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.1004718 | A Conserved Role for p48 Homologs in Protecting Dopaminergic Neurons from Oxidative Stress | Parkinson's disease (PD) is the most common neurodegenerative movement disorder characterized by the progressive loss of dopaminergic (DA) neurons. Both environmental and genetic factors are thought to contribute to the pathogenesis of PD. Although several genes linked to rare familial PD have been identified, endogenous risk factors for sporadic PD, which account for the majority of PD cases, remain largely unknown. Genome-wide association studies have identified many single nucleotide polymorphisms associated with sporadic PD in neurodevelopmental genes including the transcription factor p48/ptf1a. Here we investigate whether p48 plays a role in the survival of DA neurons in Drosophila melanogaster and Caenorhabditis elegans. We show that a Drosophila p48 homolog, 48-related-2 (Fer2), is expressed in and required for the development and survival of DA neurons in the protocerebral anterior medial (PAM) cluster. Loss of Fer2 expression in adulthood causes progressive PAM neuron degeneration in aging flies along with mitochondrial dysfunction and elevated reactive oxygen species (ROS) production, leading to the progressive locomotor deficits. The oxidative stress challenge upregulates Fer2 expression and exacerbates the PAM neuron degeneration in Fer2 loss-of-function mutants. hlh-13, the worm homolog of p48, is also expressed in DA neurons. Unlike the fly counterpart, hlh-13 loss-of-function does not impair development or survival of DA neurons under normal growth conditions. Yet, similar to Fer2, hlh-13 expression is upregulated upon an acute oxidative challenge and is required for the survival of DA neurons under oxidative stress in adult worms. Taken together, our results indicate that p48 homologs share a role in protecting DA neurons from oxidative stress and degeneration, and suggest that loss-of-function of p48 homologs in flies and worms provides novel tools to study gene-environmental interactions affecting DA neuron survival.
| Parkinson's disease is a common movement disorder with no known cure. Its characteristic motor symptoms are primarily caused by the progressive loss of midbrain dopaminergic neurons. Although studies have shown that various environmental and genetic factors both contribute to the development of the disease, the underlying mechanisms remain unknown. Here we use powerful invertebrate model organisms, fruit flies and nematode worms, and identify a new gene required for the survival of dopaminergic neurons. We show that homologs of the p48/ptf1-a gene in both flies and worms are expressed in dopaminergic neurons and mutations in p48 increase the susceptibility of dopaminergic neuron death when animals are under oxidative stress. Importantly, genetic variations in p48 in humans have been detected in the sporadic Parkinson's disease patients, indicating the possibility that similar mechanism might play a role in the death of dopaminergic neurons in humans. Oxidative stress has been regarded as a major pathogenic factor for Parkinson's disease. Our results add evidence to the link between oxidative stress and neurodegeneration, and suggest that p48 mutant flies and worms can be used to study mechanisms of neurodegeneration in Parkinson's disease.
| Dopaminergic (DA) neurons play critical roles in motor control, cognition and motivation and are affected in many neurological and psychiatric disorders [1], [2], [3], [4]. The progressive degeneration of DA neurons in the substantia nigra pars compacta (SNc) is a principal pathological feature of Parkinson's disease (PD). PD is the most prevalent neurodegenerative movement disorder, for which no preventive or restorative therapies are available [5], [6]. The discovery of the genes associated with the rare familial forms of PD has led to the development of many animal models and advanced the understanding of PD pathogenesis. However, the majority of PD cases are sporadic and likely caused by a combination of environmental factors, such as pesticide exposure, and endogenous risk factors. These endogenous risk factors remain largely unknown. A recent meta-analysis on genome-wide association studies (GWAS) for PD showed that SNPs in the genes involved in multiple aspects of neural development are highly represented in sporadic PD patients [7], suggesting that genetic variations in these pathways may contribute to PD susceptibility. Indeed, several studies in mammals have shown the critical roles of developmental genes, such as Engrailed1, foxa2 and Nurr1, in the survival of DA neurons in old age [8], [9], [10], [11]. The identification and characterization of such genes may yield a better molecular understanding of adult-onset neurodegeneration in PD.
The nervous system in invertebrate model organisms such as Drosophila and C.elegans shares many features with its mammalian counterpart and offers a powerful tool to study neural development and neurodegeneration. Drosophila DA neurons comprise multiple subclasses, some of which play roles similar to those played by the DA neurons in mammals, such as reward signaling and sleep regulation [12], [13]. The nematode C.elegans has 8 DA neurons, which are thought to have mechanosensory functions and have been shown to play a role in the modulation of locomotion [14]. Despite advances in anatomical and functional characterization, the mechanisms underlying the development and maintenance of DA neurons in flies and worms are poorly understood. Drosophila Fer2, a homolog of mammalian p48/ptf1a, belongs to the bHLH-transcription factor family, which is often involved in neurogenesis and neural subtype specification. The mammalian p48 gene is a critical regulator for neural tube development [15], in which a candidate causal SNP for PD has been detected [7,16, The database of Genotypes and Phenotypes (dbGaP; NCBI)]. Previously, we showed that Fer2 is required for the development of a subclass of circadian clock neurons, ventral Lateral Neurons (LNvs) [17]. Here we characterized additional roles of Fer2 to better understand the genetic mechanisms of neuronal subtype development and maintenance. We unexpectedly found that Fer2 is required for the development and maintenance of a subclass of DA neurons important for locomotion. Fer2 exerts its neuroprotective role in adulthood in the oxidative stress response, and loss of Fer2 expression in adulthood causes adult-onset progressive degeneration of these DA neurons. We further demonstrated that the C. elegans homolog of p48, hlh-13, is also required for the survival of DA neurons in adult worms under oxidative stress. Collectively, our results established a conserved role of p48 homologs in protecting DA neurons from oxidative stress and degeneration.
Drosophila Fer2e03248 (henceforth referred to as Fer21) mutation was induced by the insertion of PBac{RB} into the Fer2 5′UTR [18]. The Fer2MB09480 (henceforth called Fer22) allele has a Mi{ET1} transposon insertion in the 3′ end of the second exon [19] (Fig. 1A). To molecularly characterize the Fer2 mutant alleles, we determined the Fer2 mRNA levels using quantitative real-time PCR (qPCR). Consistent with our previous results, Fer2 mRNA expression of the Fer21 homozygotes was approximately 5% of the wild-type level [17]. In the Fer22 homozygous flies, Fer2 mRNA expression was reduced to about 40% of the wild-type level. Thus, Fer21 is an extreme hypomorphic allele, whereas Fer22 is a milder hypomorph. Fer2 mRNA expression of both Fer21/+ and Fer22/+ flies was only slightly reduced relative to the wild-type level, suggesting that the loss of one copy of a Fer2 gene is compensated at the mRNA level by transcriptional or post-transcriptional mechanisms (Fig. 1B). Compensation of gene dose has been observed widely in both Drosophila and mammals [20], [21], [22]. Therefore, Fer2 is a haplosufficient gene and Fer2 heterozygous mutants are expected to be phenotypically wild-type.
We noticed that Fer2 mutant flies tend to climb up the walls poorly when tapped down to the bottom of the vials. We quantified this behavior using a startle-induced climbing assay. Wild-type and heterozygous Fer2 mutants showed similar climbing abilities at least during the first 3 weeks of the adult life. In contrast, all Fer2 homozygous or hemizygous mutants displayed severely impaired climbing abilities throughout adulthood (Fig. 1C). We generated a driver fly line expressing GAL4 under the control of the Fer2 promoter (Fer2-GAL4) and a UAS line expressing FLAG-tagged Fer2 cDNA (UAS-Fer2-FLAG). The expression of Fer2-FLAG with Fer2-GAL4 partially but significantly rescued the decreased climbing ability of the Fer21 flies and restored the climbing ability of the Fer22 flies to the control level (Fig. 1D). These data indicate that Fer2 is necessary for the startle-induced climbing ability.
We have previously shown that Fer21 mutation impairs the development of LNvs, which express the neuropeptide pigment-dispersing factor (PDF) [17]. Although it has been shown that PDF is necessary for the normal negative geotaxis behavior [23], whether PDF or LNvs are necessary for startle-induced climbing has not been documented. We found that the expression of UAS-hid with Pdf-GAL4, which selectively ablates LNvs [24], does not impair the startle-induced climbing ability (Fig. 1E, F). Thus, the decrease in startle-induced climbing ability in Fer2 mutant flies is due to the deficits other than the lack of LNvs. Because loss of climbing ability is often associated with impaired CNS integrity [25] and the available transcriptome data indicate that Fer2 is almost exclusively expressed in the brain (modENCODE Tissue Expression Data, FlyAtlas [26]), we examined the integrity of major neuron types in the brains of adult Fer21 mutants. We did not find any obvious differences in the overall morphology of the cholinergic, glutamatergic and serotonergic neurons between Fer21 and controls, although we cannot exclude the possibility that there are subtle differences in the number of these neurons (Fig. S1). Interestingly, we found an evident reduction of dopaminergic (DA) neurons in Fer21 mutants. Seven DA neuron clusters were detected by anti-tyrosine hydroxylase (TH) staining in the Fer21 heterozygouse flies, which were very similar in number and morphology to those in wild-type flies [27]. In contrast, there were markedly fewer DA neurons in the PAM and PAL clusters in homozygous Fer21 flies on the first day after eclosion (day 0); there were even fewer of them in 7-day-old flies (Fig. 2A). In addition, we expressed UAS-GFP under the control of the HL9-GAL4 driver to label several clusters of DA neurons [28] and found a similar dramatic reduction of PAM and PAL neurons in the homozygous Fer21 flies (Fig. S2A). This indicates that PAM and PAL neurons were reduced in number in Fer21 mutants, rather than merely having reduced TH expression. Quantification of the HL9 > GFP-positive neurons revealed a 75% reduction in PAM neuron counts already at day 0 and a 90% reduction at day 7 in Fer21 compared to the heterozygous controls. Four out of 5 PAL neurons were undetectable at day 0 in the Fer21 flies, and most brains had no PAL neurons at day 7. The numbers of other DA neuron clusters were not different between Fer21 and controls at both ages (Fig. 2B). To further verify the loss of DA neurons in the Fer21 flies, we expressed GFP using the R58E02-GAL4 driver, which is derived from the promoter of the dopamine transporter gene and expressed almost exclusively in PAM neurons [29]. There were significantly fewer R58E02-GAL4-labeled PAM neurons in the Fer21 flies compared to the control, supporting the finding that a large fraction of PAM neurons were absent in Fer21 (Fig. 2D). The expression of Fer2-FLAG by Fer2-GAL4 restored the loss of PAM and PAL neurons in the Fer21 flies to quasi wild-type levels (Fig. 2C).
To examine if Fer2 is expressed in PAM and PAL neurons, we monitored the expression of a GFP-tagged Fer2 genomic transgene in the brain by GFP/TH double staining. FER2::GFP expression was observed in all the PAM and PAL neurons and in a few other clusters of cells. As expected, GFP/PDF double staining confirmed the expression of FER2::GFP in the LNvs, consistent with the previous RNA analysis results [17] (Fig. 3A, B). Fer2-GAL4 showed a more widespread expression pattern than FER2::GFP, as is often the case with promoter-GAL4s. Nonetheless, Fer2-GAL4 was also expressed in all PAM neurons and 4 out of 5 PAL neurons (Fig. S2B). Having validated the expression of Fer2-GAL4 in PAM and PAL neurons, we next used it to express UAS-TH in the Fer21 flies and immunostained the brains with anti-TH antibodies. Fer2 > TH slightly increased the number of neurons detected by TH-staining in the PAM and PAL clusters but not to the control level, which demonstrates again the absence of these cells (Fig. S2C). These results suggest that Fer2 is expressed in PAM and PAL neurons and further support that Fer21 mutation selectively reduces the number of these neurons.
To examine the possibility that the dopaminergic neurotransmitter identity of PAM and PAL neurons is changed in Fer21 and thus they are undetectable, we analyzed the cell lineage derived from the Fer2-GAL4-expressing cells. By combining Fer2-GAL4, UAS-FLP and UbiP63 > stop> EGFP, the Fer2-GAL4-expressing lineage was marked with GFP in the control and Fer21 flies (Fig. S2D) [30]. While most of the PAM and 4 PAL neurons were GFP-positive in the heterozygous control flies, the majority of these neurons were not present and no ectopic GFP-positive cells were observed in Fer21 (Fig. S2E, F). Therefore, the reduction of PAM and PAL neurons in Fer21 is not due to a cell-fate switch.
To learn more about the developmental impairments of PAM and PAL neurons in Fer21 mutants, we next examined DA neurons in the pupal brains with anti-TH staining, because these neurons are not present in the larval brain and some of the PAM neurons are known to be born during pupal stages [27], [31]. In the Fer21 heterozygous controls, approximately 80% of PAM and PAL neurons were clearly detectable within 5 days after puparium formation (APF). Whereas in Fer21 homozygotes, PAM and PAL neurons gradually increased in number but were significantly fewer than in the controls throughout pupal development. These observations indicate that the majority of PAM and PAL neurons were not formed or died before maturation into DA neurons in Fer21 mutants (Fig. S3A, B). Taken together with the observation that the loss of PAM/PAL neurons progresses at least up to 7 days into adulthood (Fig. 2B), these results indicate that Fer21 mutation impairs the development of the DA neurons in the PAM and PAL clusters and also causes their rapid degeneration in adulthood.
We next asked whether the integrity of DA neurons is affected in the milder hypomorphic mutant Fer22 as well. We focused our analysis on the PAM cluster DA neurons and monitored their number in the Fer22 flies by anti-TH staining. At day 0, the number of PAM neurons was slightly reduced in Fer22 compared to the control, but to a much lesser extent as in Fer21. Remarkably, the loss of PAM neurons continued progressively at least up to 28 days in Fer22 (Fig. 4A). Lineage tracing using HL9-GAL4, UAS-FLP and UbiP63 > stop > EGFP flies verified the loss of PAM neurons in Fer22 (Fig. S4A). The loss of PAM neurons was rescued by expressing Fer2-FLAG with Fer2-GAL4 (Fig. S4B).
The contrasting results between Fer21 and Fer22 mutants suggest that a moderate reduction of Fer2 expression has only a minor effect on the development of DA neurons but is sufficient to deteriorate PAM neurons in aged flies. To test this, we generated UAS-transgenic lines to express 2 independent microRNAs (miRNAs) that target Fer2 (miR Fer2-4 and -5) and one negative control miRNA (miR Fer2-N) that contains a sequence of 19 random nucleotides unrelated to any Drosophila gene (see Fig. 1A). When expressed with Fer2-GAL4, both miR Fer2-4 and -5 reduced the Fer2 mRNA levels to approximately 35% of the wild-type level, whereas miR Fer2-N had no effect on the Fer2 mRNA (Fig. S5A). As expected, constitutive expression of miR Fer2-N by Fer2-GAL4 at 25°C did not alter the number of PAM neurons. In the flies expressing miR Fer2-4 or miR Fer2-5, the number of PAM neurons remained stable until 35 days of age. However, at 49 days, many flies in either of the knockdowns had a reduced number of PAM neurons, and the reduction was significant in Fer2 > miR Fer2-5 flies (Fig. 4B). The constitutive knockdown by miR Fer2-4 or miR Fer2-5 at 25°C with R58E02-GAL4 or HL9-GAL4 resulted in a similar trend, with a significant reduction of PAM neurons in the flies aged over several weeks old (Fig. 4C, S5B). The PAL and other DA neuron clusters were not affected by any of the Fer2 knockdowns, although Fer2-GAL4 and HL9-GAL4 are expressed in most of the DA clusters including PAL neurons. While there were subtle differences in the onset of degeneration that were likely due to the differences in GAL4 expression levels, the data nevertheless illustrate that Fer2 knockdown causes adult-onset PAM neuron degeneration.
The foregoing observations indicate that a moderate reduction of Fer2 expression either by a hypomorphic mutation or by knockdown has little effect on DA neuron development but mainly affects the survival of PAM neurons in adults. This further suggests that the role of Fer2 in the survival of adult PAM neurons is independent of its role in development. To test this more directly, we knocked-down Fer2 only during adulthood using a combination of UAS-miR Fer2s, Fer2-GAL4 and temperature-sensitive GAL80 expressed under the tubulin promoter (tub-GAL80ts). These flies were reared at 18°C (a permissive temperature for GAL80ts) until eclosion, and then the temperature was shifted to 29°C (a restrictive temperature for GAL80ts) to allow for transcriptional activation by GAL4 throughout adulthood [32]. We found that the adult-specific knockdown of Fer2 induced the adult-onset progressive degeneration of PAM neurons without affecting PAL neurons (Fig. 4D). Notably, the loss of PAM neurons was more evident in these flies than in flies with constitutive knockdown (Fig. 4B), which is consistent with the greater GAL4 activity at 29°C than at 25°C [33]. These results clearly distinguish the role of Fer2 in developing and adult DA neurons and demonstrate that Fer2 expression is required for the survival of adult PAM neurons in aging flies.
We next asked whether the loss of PAM neurons, which is the most prominent cellular phenotype in Fer2 mutants, is the cause of their locomotor impairment. To test the role of PAM neurons in the startle-induced climbing ability more directly, we knocked-down Fer2 in PAM neurons by HL9-GAL4 and performed a climbing assay. Knocking-down Fer2 with either of the miRNAs resulted in significant declines in climbing ability after 49 days (Fig. 5A). This is consistent with the observation that HL9 > miR Fer2s induce the adult-onset degeneration of PAM neurons only after several weeks (Fig. S5B).
To further assess the contribution of PAM neurons in the climbing ability, we sought to rescue the loss of PAM neurons using a PAM neuron-specific driver in Fer2 mutant flies. Expression of UAS-Fer2-FLAG using HL9-GAL4 or R58E02-GAL4 did not rescue the loss of PAM neurons in Fer21. This is most likely because the majority of PAM neurons fail to form or die in Fer21 before these DA neuron-specific drivers start to be expressed, consistent with the fact that R58E02-GAL4 has little expression in the larval brain (H. Tanimoto and A. Thum, personal communication). Thus, we reasoned that PAM-neuron specific rescue might be possible in Fer22 flies, which show little impairments in DA neuron development but display progressive PAM neuron degeneration. Indeed, R58E02 > Fer2-FLAG suppressed the degeneration of PAM neurons in the Fer22 mutants in adulthood (Fig. 5B). A climbing assay revealed that R58E02 > Fer2-FLAG significantly improves the climbing impairments in the Fer22 flies (Fig. 5C). The rescue of the climbing ability was partial. This may be because some of the PAM neurons that failed to develop in Fer22 were not rescued, or because other unknown cell types affected in Fer22 contribute to the climbing ability. Nevertheless, these observations together with the results of the PAM neuron targeted-knockdown indicate that PAM neurons are necessary, although may not be sufficient, for the normal climbing ability of the flies.
The dopaminergic system is critically involved in the control of locomotion in both vertebrates and invertebrates [34], [35]. The motor symptoms of PD arise mainly from the loss of DA neurons in the SNc [5]. L-dopa, a dopamine biosynthesis precursor, remains the gold standard for treatments of PD motor symptoms. We found that the locomotor deficit of the Fer21 flies was partially but significantly rescued by feeding with L-dopa (Fig. 5D). By anti-TH staining, we observed no significant rescue of the number of DA neurons by L-dopa, which is consistent with a previous study [36]. Since L-dopa has to be converted to dopamine in the DA neuron terminals to exert its therapeutic effect, the partial rescue of the locomotion by L-dopa is also consistent with the marked loss of PAM neurons observed in Fer21 mutants.
Accumulating evidence suggests that dysfunctions in multiple aspects of mitochondrial biology are associated with the DA neurodegeneration in PD and pathogenesis of other neurodegenerative disorders [37], [38]. To examine whether mitochondrial dysfunction is involved in the loss of DA neurons caused by the loss of Fer2 expression, we visualized mitochondria in the adult PAM neurons by expressing mitochondria-targeted GFP (mitoGFP) [39] with HL9-GAL4. The majority of visible mitochondria in the cell bodies of the remaining PAM neurons in Fer21 mutants was in enlarged aggregations and did not form tubular networks as in the control flies (Fig. 6A). Similarly, DA neuron-selective Fer2 knockdown by HL9-GAL4 and Fer22 mutation lead to the accumulation of abnormally enlarged mitochondria in PAM neurons (Fig. 6B, S6A). Mitochondrial morphology in some of the TRH-GAL4-positive serotonergic neurons was indistinguishable between Fer21 homozygotes and heterozygotes, suggesting that mitochondria in PAM neurons are particularly vulnerable to loss of Fer2 expression (Fig. S6B).
Since mitochondria are the major source of ROS, mitochondrial dysfunction leads to an excessive ROS production and oxidative damages to various macromolecules. Oxidative stress causes rapid depolarization of mitochondrial inner membrane and inhibits complex I activity, exacerbating ROS production. Thus, mitochondrial defects and elevated ROS levels are interdependent and are thought to have prominent roles in PD pathogenesis [40]. We therefore asked whether ROS levels are increased in the Fer2 mutant brains and if it has a causative role on PAM neuron degeneration. We monitored intracellular ROS levels in the brains of Fer22 mutant and the heterozygous control flies using 2′,7′-dichlorofluorescein (H2DCF), which produces green fluorescence upon reacting with ROS. Interestingly, ROS levels were significantly elevated throughout the brain in 5-day-old Fer22 flies compared with the age-matched controls. Although there was no regional specificity of ROS accumulation, ROS levels within the PAM neurons were also significantly elevated in Fer22 (Fig. 6C). These suggest that loss of Fer2 expression leads to a systemic increase in oxidative stress in the brain.
The surprisingly dramatic increase of ROS levels in Fer22 mutants prompted us to further examine if Fer2 is involved in oxidative stress response. We first tested if Fer2 expression levels can be altered upon oxidative stress-challenge by feeding flies with non-lethal dose of hydrogen peroxide (H2O2, 5%) for 24 hr. We found that H2O2 treatment significantly increases Fer2 mRNA levels (Fig. 6D). We next examined whether oxidative stress-challenge aggravates the degeneration of PAM neurons in Fer2 mutants by anti-TH staining. The H2O2 treatment did not affect PAM neurons in Fer22 heterozygous flies, whereas the number of PAM neurons was significantly decreased in Fer22 homozygotes after the treatment (Fig. 6E). DA neuron counts in other clusters were unchanged by the same H2O2 treatment in Fer22 mutants (Fig, S6C), indicating that loss of Fer2 expression renders PAM neurons selectively more vulnerable to increased oxidative stress. Taken together, these results point toward a role for Fer2 in oxidative stress response and suggest that Fer2 contributes to the protection of PAM neurons against oxidative stress.
Fer2 homologs are found from nematodes to vertebrates [41]. hlh-13 is predicted to be the single homolog of Drosophila Fer2 and mammalian p48/ptf1a in C.elegans [42]. Consistent with a previous study [42], a GFP::hlh-13 genomic transgene was expressed in all DA neurons (named CEP (4 cells), ADE (2 cells) and PDE (2 cells)) and in a tail neuron in developing and adult worms. GFP::hlh-13 expression was also observed in several unidentified ventral nerve cells from L2 to L4 stages (Fig. 7A).
Since both fly Fer2 and worm hlh-13 are expressed in DA neurons, we next asked whether hlh-13 has a comparable function as Fer2 in DA neuron development or survival. We used the hlh-13 knockout mutant hlh-13(tm2279) to test the effect of hlh-13 loss-of-function on the number of DA neurons and on a dopamine-dependent behavior, the basal slowing response. The basal slowing response is a slowing of locomotion rate when worms encounter bacteria and has been shown to require dopamine signaling [43] (Text S1). The knockout mutant showed no differences in the number of DA neurons and basal slowing response compared to wild-type worms (Fig. 7D control, Fig. S7A). Therefore, unlike Fer2, hlh-13 is not required for the development, survival or function of DA neurons under normal growth conditions.
Next, to test if hlh-13 is involved in the survival of DA neurons under oxidative stress, we treated wild-type and hlh-13(tm2279) mutant adult worms with 1 mM H2O2 for 30 min and analyzed the hlh-13 mRNA levels and DA neuron integrity at subsequent time points. hlh-13 mRNA levels were upregulated by approximately 3-fold immediately after the H2O2 treatment and returned to the non-treated levels after 2 hrs (Fig. 7B). To monitor DA neurons in the wild-type or hlh-13(tm2279) background, we used the dat-1::gfp reporter driving GFP expression in DA neurons. Since it was difficult to reliably detect PDE neurons, we focused our analysis on CEP and ADE neurons located in the head. In wild-type animals, at least until 7 days after the H2O2 treatment, there were no significant differences in the number or morphology of the CEP and ADE neurons between treated and untreated groups. By contrast, H2O2- treated hlh-13(tm2279) mutants showed fragmentation of the CEP neuron projections starting from day 4 after treatment, followed by the loss of cell bodies. Similarly, the number of ADE neurons was also reduced in the H2O2-treated mutants (Fig. 7C, D). Despite the apparent change in DA neuron numbers, the basal slowing response was not different between the control and stressed worms in either genotype (Fig. S7B), which is consistent with the previous observation that basal slowing response is defective only when all 4 CEPs are ablated [43]. These results indicate that, similar to fly Fer2, hlh-13 is likely to be involved in the oxidative stress response and required for the protection of DA neurons under oxidative stress in adult worms.
To examine the extent to which hlh-13 shares the function with Fer2, we next sought to test cross-species complementation of the Fer2 loss-of-function mutation in flies by the hlh-13 gene. We generated a UAS-hlh-13 construct and expressed it with Fer2-GAL4 in the Fer21 mutant background. We found that the loss of PAM and PAL neurons in Fer21 was partially but significantly rescued by the expression of hlh-13 (Fig. 7E). Collectively, these results confirm that hlh-13 is the C.elegans ortholog of Fer2 and suggest that the protection of DA neurons against oxidative insults is a conserved role between these orthologs.
Many neurodegenerative disorders are multi-factorial, in which interactions between environmental and genetic factors play important causal roles. Oxidative stress has emerged as a major pathogenic factor for common neurodegenerative diseases, yet how such a ubiquitous phenomenon leads to the loss of selective neuronal populations remains unclear [44]. Here we presented evidence that loss-of-function in p48 homologs in Drosophila and C.elegans renders DA neurons susceptible to degeneration under oxidative stress in adult animals. Interestingly, genome-wide association studies for PD have identified candidate causal SNPs in p48/ptf1a [7], [16], suggesting the possibility that p48 loss-of-function may represent an as-yet-unknown genetic risk factor that increases susceptibility of DA neurons to environmental toxins also in mammals. Many familial PD-associated genes are widely expressed; nevertheless, mutations in these genes result in a selective loss of SNc DA neurons, suggesting that cell-type-specific factors, those similar to Fer2 and hlh-13, might contribute to the DA neuron vulnerability even in the familial PD cases. The identification of Fer2 and hlh-13 upstream and downstream pathways may thus shed light on the common mechanisms underlying the selective loss of DA neurons in diverse PD cases.
The major cellular phenotype in Fer21 mutants was the developmental defects in 2 subsets of DA neurons, in addition to the developmental loss of LNvs [17], although we cannot exclude the possibility that other neuronal types are also affected (Fig 2, S1). Judging from the results of the lineage-tracing experiments and the observation of DA neurons in the pupal brain, Fer2 is not a selector gene for dopaminergic phenotype in PAM/PAL neurons but is required for neurogenesis or survival of postmitotic neurons before phenotypic maturation (Fig. S2E, F and S3). The notion that genes required for the development of DA neurons confer important roles in adult DA neuron survival has been postulated by several studies in mammals [8], [9], [10], [11]. Although the molecular mechanisms underlying their roles in adult neurons remain elusive, these developmental genes may actively control the genetic programs required for the maintenance of cell identity in adults [45]. Our findings on the Fer2's dual roles extend this notion to invertebrate nervous systems and underscore its significance.
PAM neuron-targeted Fer2 knockdown induces PAM neuron degeneration (Fig. 4C) and mitochondrial dysfunction within PAM neurons (Fig. 6B, S6A). These results indicate that mitochondrial dysfunction and cell death can be induced by a cell-autonomous reduction of Fer2 expression within the PAM cluster. On the other hand, ROS levels are increased brain-wide in the Fer22 flies, despite the fact that Fer2 expression is restricted to several clusters of cells in the brain (Fig. 3, 6C). Thus, loss of Fer2 expression leads to both cell-autonomous and non-cell-autonomous consequences to the animal's well-being. How does the brain-wide ROS increase occur by Fer2 mutation although Fer2 is not expressed ubiquitously? An intriguing recent study in C. elegans demonstrated that mitochondrial perturbation in neuronal cells modulates mitochondrial stress response in distal tissues non-cell-autonomously [46]. Flies might exhibit similar non-cell-autonomous mitochondrial stress response that causes systemic ROS production. Systemic increase in oxidative stress is a clinical feature common to many aging-related neurological diseases including PD [47]. Studies in mammals have documented that inflammation is a major factor mediating excessive ROS production and PD pathology. Activated microglia produces ROS and mediates DA neuron death. Dying DA neurons stimulate microglia, exacerbating the ROS production and DA neurodegeneration [48]. As CNS glia in Drosophila are thought to possess immune-like function [49], similar mechanisms via inflammatory responses might mediate global elevation of ROS production in Fer2 mutants.
Are the abnormal mitochondria in PAM neurons a cause or a consequence of the ROS upregulation? Because mitochondrial defects and excessive ROS production are inter-dependent, it is not possible to clarify the causality in the current study. However, because Fer2 expression is upregulated upon H2O2 treatment and the same acute H2O2 treatment triggers PAM neuron death in the absence of Fer2 (Fig. 6D, E), we favor the hypothesis that Fer2 provides protection against oxidative stress rather than directly acting on mitochondria (Fig. 8). These phenomena are remarkably similar in C.elegans; an acute H2O2 treatment upregulates hlh-13 expression and triggers DA neuron degeneration in hlh-13 null mutants (Fig. 7B–D). These data suggest that the oxidative stress response is an ancestral role of p48 homologs. Alternatively, hlh-13's roles in neural development in worms might have been taken over by other genes. Either way, these findings suggest that loss-of-function in Fer2 and hlh-13 can be used to study pathophysiology of DA neuron degeneration under oxidative stress. Interestingly, Fer2 mRNA levels remain upregulated at least up to 12 hr after the 24-hr H2O2 treatment, whereas hlh-13 mRNA levels return to the non-treated levels 2 hr after a brief H2O2 treatment (Fig. 6D, 7B). This difference in gene expression kinetics may reflect the duration of the H2O2 treatment, RNA stability, or difference in signal transduction mechanisms. Various stress response genes show highly restricted temporal expression upon stress, as the continuous activation of these genes are often detrimental to the cell [50]. Initial upregulation of hlh-13 immediately after an acute oxidative stress might be necessary and sufficient to trigger the downstream genetic programs that continue to scavenge ROS and repair the cellular damages during the following days. Identification of the downstream genetic programs controlled by Fer2 and hlh-13 will be a key toward understanding the evolutionarily conserved mechanisms of neuroprotection.
Mild loss of Fer2 expression by Fer22 mutation or knockdown leads to a progressive loss of PAM neurons associated with mitochondrial dysfunction, increase in ROS production and progressive locomotor deficits, all of which are reminiscent of the pathological characteristics of PD. Unlike other fly PD models that are derived from genetic modifications of human PD-associated genes or their homologs, Fer2 is not an ortholog of known familial PD-associated genes. Yet, the magnitude of the DA neuron degeneration caused by the loss of Fer2 expression is markedly greater than in existing fly PD models [51]. We demonstrated that the loss of PAM neurons is at least partly responsible for the impaired climbing ability caused by Fer2 loss-of-function (Fig 5A–C). Because rescue of the PAM neuron counts in Fer21 mutants to quasi wild-type level does not restore the climbing ability to the control level (compare Fig. 1D and 2C), it is likely that some cells other than PAM and PAL neurons are somehow affected in Fer21 and contribute to the locomotor deficits. Nonetheless, our results are in agreement with the recent study by S. Birman and colleagues, which demonstrates that the progressive motor deficits in the flies expressing human α-synuclein, a transgenic model of PD, derives from the dysfunction of a subset of PAM neurons [52]. Because the selective degeneration of DA neurons within the SNc is the principal cause of the motor manifestations of PD, the Drosophila PAM neurons parallel the DA neurons of the human SNc with regard to function and vulnerability. Thus, Fer2 loss-of-function may serve as a model to better understand the mechanisms by which the loss of specific subsets of DA neurons leads to locomotor deficits in PD.
PBac {RB} Fer2e03248 (Fer21) has been previously characterized [17]. The following lines were obtained from the Bloomington Stock Center: Df(3R)Exel7328 (referred to herein as Df), Mi{ET1}Fer2MB09480 (Fer22), UAS-mitoGFP and the strain used for Fer2-GAL4 flip-out assay (w; P{UAS-RedStinger}4, P{UAS-FLP1.D}JD1, P{Ubi-p63E(FRT.STOP)Stinger}9F6/CyO) [30]. HL9-GAL4 [28] was a gift from G. Miesenböck. dvglut CNSIII-GAL4 and Cha-GAL4 [53] were gifts from A. DiAntonio. TRH-GAL4 [54] was from O. Alekseyenko. DILP2-GAL4 [55] was a gift from P. Léopold. R58E02-GAL4 [29] was from H.Tanimoto and UAS-TH [56] was from J.True. Fer2::GFP genomic transgene FlyFos022529 was derived from a fosmid clone including approximately 5 kb upstream and 20 kb downstream of the Fer2 gene. FlyFos022529 (pRedFlp-Hgr) (Fer2[16092]::2XTY1-SGFP-V5-preTEV-BLRP-3XFLAG)dFRT) was generated and generously provided by the project “A reverse genetic toolkit for systematic study of gene function and protein localization in Drosophila” (M.IF.A.MOZG8070).
To generate Fer2-GAL4, 2359 bp upstream of Fer2 ATG were amplified using the following primers: EagFer2upF, 5′- TTTCGGCCGTGGATTTGCTCTGGTTTGGATGC -3′ and XhoFer2upR, 5′- TTTCTCGAGTTTTACGCACTTCCGCTGTCC -3′. The amplified fragment was cloned into a pENTR 3c gateway vector (Invitrogen), verified by sequencing and cloned into to the transformation vector containing GAL4 (pBPGal4.2 Uw2; Addgene) using the gateway system (Invitrogen). The Fer2-GAL4 transgene was inserted into the attP16 landing site [57] on the second chromosome by PhiC31-mediated recombination by a commercial transformation service (BestGene, Inc.). UAS-miR Fer2 constructs were generated as described in [58]. Four different miRNA coding sequences were generated along with a negative control, which does not target any sequence in the Drosophila melanogaster genome. From these 4, 2 miRNAs (lines 4 and 5) were used for further experiments. The miRNA sequence of line 4 was 5′-TGAGCAAGATCGACACTCTGC-3′, the miRNA sequence of line 5 was 5′-TCAAAGCGGATAGGGCTAATT-3′ and the sequence of the negative control miRNA was 5′- TACCCGTATCGGGTTAATCGA -3′. The stem loop structure was predicted using the RNAfold Webserver of the institute of Theoretical Chemistry at the University of Vienna (http://rna.tbi.univie.ac.at/cgi-bin/RNAfold.cgi). UAS-miR Fer2 constructs were inserted into the attP40 landing site on the second chromosome (BestGene, Inc.). To generate the UAS-Fer2-FLAG construct, Fer2 cDNA with a 3xFLAG tag coding sequence at its 3′ end was amplified and cloned into pCR II Topo vector (Invitrogen) and verified by sequencing. A DNA fragment containing Fer2-FLAG coding sequence was then cloned into a UAS-containing transformation vector (pUAST-UAS-Stringer attB). The UAS-Fer2-FLAG construct was integrated into the attP40 landing site. The UAS-hlh-13 construct was generated by cloning a PCR-amplified hlh-13 full-length cDNA into the pBid-UASC-G vector [59] by Gateway cloning, and integrated into the attP40 landing site.
Throughout the text and in the figures, genotype “X>Y” indicates a combination of GAL4 (X) and UAS-effector (Y).
C.elegans were cultured using standard protocol unless otherwise indicated. The following strains were obtained from the Caenorhabditis Genetic Center: wild-type (N2), BZ555 egIs1[dat-1::gfp] and IU189 rwls1[hlh-13p::GFP::hlh-13,mec-7::RFP]. IU129 hlh-13(tm2279) was a gift from S. Lee [42].
To assay the startle-induced locomotion of the flies, we used a negative geotaxis assay with modifications [60]. Twenty flies were anesthetized with CO2 and placed in a vertical glass column (25 cm length, 1.5 cm diameter) with a conical bottom. The columns were divided into 5 equally spaced zones and graded from 1 to 5 from the bottom to the top. After a 1-hr recovery period from CO2 exposure, the flies were gently tapped to the bottom. The flies were then allowed to climb the wall for the subsequent 20 seconds. The experiments were video recorded and the videos were manually analyzed using VLC software. The numbers of flies that climbed up to each zone within 20 seconds were counted. Flies that remained at the bottom were defined to be in zone 0. A climbing index (CI) was calculated using the following formula: CI = (0×n0 + 1×n1 + 2×n2 + 3×n3 + 4×n4 + 5×n5) / ntotal, where ntotal is the total number of flies and nx is the number of flies that reached zone X. One experiment consisted of 3 trials performed at 5 min intervals. Two or three independent experiments were performed for each condition, and the mean climbing indexes of independent experiments are shown. All climbing assays were performed 2 hr after lights on (ZT2) to avoid any circadian variation in locomotor activities.
The feeding of the flies with L-dopa feeding was performed as described previously with minor modifications [61]. Flies were raised on fresh medium made from instant food (formula 4–24) containing the antioxidant ascorbic acid (25 mg/100 ml), the antifungal agent Nipagin and L-dopa (1 mM) (Sigma Aldrich, D9628). Control vials contained only ascorbic acid and Nipagin.
For the co-immunostaining of fly brains with anti-GFP and nc82 antibodies, flies were decapitated and the heads were fixed with 4% paraformaldehyde +0.3% Triton X-100 for 1 hour on ice and washed twice with PBST-0.5 (PBS, 0.5% Triton X-100). Subsequently, the head cuticle was partly removed and the heads were washed twice more and blocked in blocking solution (5% normal goat serum, PBS, 0.5% Triton X-100) for 1 hr at room temperature and incubated with the primary antibodies overnight at 4°C. After 2 washes, the heads were incubated with secondary antibodies (Alexa-conjugated) for 2 hours at room temperature. Cuticles and tracheas were removed, and the brains were mounted in Vectashield mounting medium. For staining with anti-TH antibodies with or without other antibodies, 0.3% Triton X-100 was added instead of 0.5% in PBST and in blocking solution. Primary antibodies were incubated over 2 nights at 4°C, and secondary antibodies were incubated at 4°C overnight. The primary antibodies and concentrations used in this study were as follows: rat monoclonal anti-GFP (GF090R) (Nacalai Tesque, Inc.) 1∶500, rabbit polyclonal anti-tyrosine hydroxylase (ab152) (Millipore) 1∶100, mouse monoclonal antibody nc82 (Developmental Studies Hybridoma Bank) 1∶100, mouse monoclonal anti-RFP (AKR-021) (Cell BioLabs, INC.) 1∶500, mouse monoclonal anti-tyrosine hydroxylase (22941) (Immunostar) 1∶50 and polyclonal rabbit anti-GFP (A6455) (Invitrogen) 1∶200.
Leica TCS SP5 confocal microscope was used to image fly brains. Quantification was performed using ImageJ software (NIH). To count the number of DA neurons, anti-TH-positive neurons or anti-TH, anti-GFP double-positive neurons were counted manually through confocal Z-stacks. To image the expression of hlh-13-GFP a Leica DM5500 B microscope was used. TILL Phototonics iMIC digital microscope was used to image DA neurons in worms and DA neurons were counted manually on each Z-stack using ImageJ software.
Total RNA isolation from fly heads, cDNA synthesis and quantitative-PCR (qPCR) analysis were performed as described previously [17]. mRNA levels of Fer2 were normalized to those of the housekeeping gene elongation factor 1β (Ef1β). For qPCR analysis of worm mRNAs, worms were collected from the plates with M9 buffer, placed into the Falcon tubes and left to settle for 15 min to remove bacteria in their guts. Worms were then washed twice by spinning at 3000 rpm for 1 min. Total RNAs were isolated using TRIZOL (Life Technologies), and mRNAs were reverse-transcribed and used as templates for qPCR. hlh-13 mRNA levels were normalized to the its-1 levels.
Statistical analyses were performed using StatPlus (AnalystSoft) and SPSS software (IBM). For normally distributed data sets, two-tailed Student's t-tests were used to compare the means of two groups. The data that were not normally distributed were analyzed with non-parametric statistics (Mann–Whitney U test). For all experiments, the level of significance was set at p<0.05.
The numbers of brain hemispheres examined in Figure 4A-D are as follows. (A) Control (Fer22/+): day 0, n = 6; day 1, n = 9; day 21, n = 8. Fer22: day0, n = 6; day1, n = 6; day 21, n = 9. (B) Fer2 > miR Fer2-N: day 0, n = 6; day 35, n = 9; day 49, n = 17. Fer2 > miR Fer2-4: day 0, n = 8; day 35, n = 7; day 49, n = 19. Fer2 > miR Fer2-5: day 0, n = 8; day 35, n = 7; day 49, n = 21. (C) R58E02 > miR Fer2-N: day 0, n = 10; day 21, n = 8; day 35, n = 12; day 63, n = 14. R58E02 > miR Fer2-4: day 0, n = 8; day 21, n = 12; day 35, n = 16; day 63, n = 16. R58E02 > miR Fer2-5: day 0, n = 10; day 21, n = 8; day 35, n = 14; day 63, n = 17. (D) Control without UAS-transgene (Fer2-GAL4, tub80): day 0, n = 15; day 28, n = 7; day 35, n = 18. Control miRNA (Fer2 > miR Fer2-N): day 0, n = 15; day 28, n = 13; day 35, n = 24. Fer2 > miR Fer2-4, tub80: day 0, n = 19; day 28, n = 14; day 35, n = 18. Fer2 > miR Fer2-5, tub80: day 0, n = 13; day 28, n = 20; day 35, n = 13.
Sample sizes in Figure 5B are as follows. Positive control (Fer22/+): day 0, n = 8; day 14, n = 14. Negative control (Fer22, R58E02-gal4): day 0, n = 15; day 14, n = 16. Rescue (Fer22, R58E02 > Fer2): day 0, n = 17; day 14, n = 13.
5-day-old flies were transferred into the empty vials for 6 hrs, then placed onto the food prepared from instant food (Formula 4–24 Instant Drosophila Medium, Carolina(R)) containing 5% H2O2 for 24 hrs. Control food contained only dH2O. The vials were placed in a humid box and kept at 25°C. Flies were subsequently collected for RNA analysis or placed on the normal food for 24 hrs prior to the dissection and TH staining.
Worms were synchronized by bleaching and stressed as young adults 2 days later (22.5°C). After washing worms off the plates with M9 medium, they were spun down gently at 130 g and washed once with M9 to remove bacteria in the solution. The worms were transferred in 5 ml M9 to an empty petri dish and H2O2-containing M9 (5 ml) was added to the final concentration of 1 mM. The worms were kept shaking for 30 min at room temperature. Worms were then washed 3 times with M9 to remove H2O2 and were placed back on the normal NGM plates for recovery.
In vivo detection of ROS production in fly brains was performed using 2′7′-dichlorofluorescein (H2DCF) as detailed in Owusu-Ansah et al. (Protocol Exchange, 2008, doi:10.1038/nprot.2008.23). Brains of the R58E02-GAL4, UAS-mCherry flies in the Fer22 heterozygous or homozygous background were imaged by confocal microscopy and the fluorescence intensity was measured using the FIJI software [62]. Since there was no regional specificity in the H2DCF signal, the central brain area (entire brain except for the optic lobe) was manually defined and the signal intensity in the defined region across Z-stacks was measured by performing a Z-SUM projection. To quantify the H2DCF signal within PAM neurons, PAM neurons were defined by thresholding the RFP signal and the total H2DCF signal within the defined volume was measured by a Z-SUM projection.
|
10.1371/journal.pgen.1001014 | Combinatorial Binding Leads to Diverse Regulatory Responses: Lmd Is a Tissue-Specific Modulator of Mef2 Activity | Understanding how complex patterns of temporal and spatial expression are regulated is central to deciphering genetic programs that drive development. Gene expression is initiated through the action of transcription factors and their cofactors converging on enhancer elements leading to a defined activity. Specific constellations of combinatorial occupancy are therefore often conceptualized as rigid binding codes that give rise to a common output of spatio-temporal expression. Here, we assessed this assumption using the regulatory input of two essential transcription factors within the Drosophila myogenic network. Mutations in either Myocyte enhancing factor 2 (Mef2) or the zinc-finger transcription factor lame duck (lmd) lead to very similar defects in myoblast fusion, yet the underlying molecular mechanism for this shared phenotype is not understood. Using a combination of ChIP-on-chip analysis and expression profiling of loss-of-function mutants, we obtained a global view of the regulatory input of both factors during development. The majority of Lmd-bound enhancers are co-bound by Mef2, representing a subset of Mef2's transcriptional input during these stages of development. Systematic analyses of the regulatory contribution of both factors demonstrate diverse regulatory roles, despite their co-occupancy of shared enhancer elements. These results indicate that Lmd is a tissue-specific modulator of Mef2 activity, acting as both a transcriptional activator and repressor, which has important implications for myogenesis. More generally, this study demonstrates considerable flexibility in the regulatory output of two factors, leading to additive, cooperative, and repressive modes of co-regulation.
| While genetic studies are essential to reveal the phenotypic relationships between genes, it is often very difficult to disentangle the molecular mechanism of two genes that phenocopy each other. In this study, we used global scale and single gene analysis to investigate the relationship between two transcription factors whose mutant embryos have a similar defect in myogenesis. In Drosophila, Mef2 mutant embryos display a block in myoblast fusion, which is very similar to what is observed in mutant embryos for lmd, a zinc-finger transcription factor. To understand the underlying nature of these defects we used ChIP-on-chip analysis to obtain a global view of their co-regulated enhancers, and we used expression profiling of mutant embryos to reveal their downstream transcriptional response. The results indicate that Lmd acts as a tissue specific modulator of Mef2 activity. Using in vivo and in vitro reporter assays, we show that co-binding to the same enhancer element can lead to diverse regulatory responses. The presence of Lmd has an additive, cooperative, or repressive effect on Mef2 activity, demonstrating that it acts as a molecular switch for gene expression during muscle differentiation. More broadly, our results highlight the difficulty in translating information on combinatorial binding data into a functional regulatory response.
| Development is driven by precise patterns of spatio-temporal gene expression, which are regulated through the action of transcription factors and cell signaling cascades converging on cis-regulatory modules (CRMs). CRMs are typically bound by multiple transcription factors, whose concentrations and interactions change dynamically over time. It is this combinatorial and dynamic property of CRM occupancy which makes regulatory output difficult, if not impossible, to predict based on information from a single transcription factor (TF) [1]. Understanding the regulation of complex developmental processes requires linking combinatorial binding at the molecular level to the regulation of these processes at the phenotypic level. We have assessed the contribution of two well-studied TFs, Mef2 (Myocyte Enhancing Factor 2) and Lmd (Lame duck) to the cellular process of myogenesis during Drosophila development. Although the phenotypic defects in myoblast fusion are almost identical in Mef2 or lmd loss-of-function mutant embryos [2]–[4], the molecular relationship between these TFs activity is poorly understood.
Members of the Mef2 family of MADS-box proteins were first characterized in vertebrates as important regulators downstream of the MyoD family of transcription factors, and have since been identified as part of an evolutionarily ancient regulatory network driving myogenesis from flies to man [5]. In vertebrates, Mef2 transcription factors act as central regulators of cell proliferation, survival, apoptosis and differentiation in a range of cell types, including skeletal, cardiac and smooth muscle, brain, neural crest, lymphocytes and bone (reviewed in [5]). This diversity in Mef2 function is achieved through regulation by extracellular signals and cooperative activity with specific co-regulators. In skeletal muscle, for example, Mef2 acts together with bHLH transcription factors to regulate the expression program that drives myogenic differentiation [6], [7]. In neural crest cells, Mef2c acts cooperatively with the DLX5 and DLX6 homeodomain TFs to regulate craniofacial development [8], [9], while in smooth muscle cells Mef2 acts together with myocardin [10]. Thus, Mef2 TFs have little inherent instructive potential by themselves but rather act together with tissue-specific TFs to drive specific gene expression programs. Given the diverse roles of the Mef2 gene family during development, many more co-regulators are likely to be required to generate the spectrum of transcriptional responses elicited by these factors.
In Drosophila embryos, the single Mef2 ortholog is expressed exclusively in mesoderm and its muscle derivatives. Even in this relatively simple model system, Mef2 regulates distinct batteries of target genes in precise spatial patterns (e.g. in the dorsal vessel, the somatic mesoderm and visceral mesoderm [11], [12]), and in a specific temporal order [11], [13]. Global in vivo occupancy experiments revealed dynamic Mef2 enhancer binding; although Mef2 is expressed continuously, it binds to one group of enhancers only early in development and to another group only at late developmental stages [11]. The temporal shift in the expression onset of Mef2 target genes [11], [13] as well as their spatial diversity, indicates a requirement for co-regulators, similar to the mechanism of Mef2 action in vertebrates [5]. holes-in-muscle (him) was recently identified as a potential repressor of Mef2-dependent transcriptional activation via the recruitment of the general co-repressor Groucho [14]. Regulation by Him therefore provides one mechanism to alter the temporal output of Mef2 activity once it is bound to an enhancer. However, other co-regulators are clearly required to modulate Mef2's temporal enhancer occupancy and to restrict its spatial activity.
The Drosophila body wall muscle or somatic muscle is formed from two heterogeneous populations of cells- the founder cells (FCs), which represent 30 distinct cells in each hemisegment of the embryo, and the fusion competent myoblasts (FCMs) [15]. Once specified, a single FC will fuse with a defined number of FCMs to give rise to a syncytial myotube of distinct identity, defined by its size, shape and attachments. Mef2 is required to initiate a program that regulates myoblast fusion and drives the differentiation program of the resulting myotube into a contractile myofiber [2]. The zinc-finger transcription factor Lame duck maintains Mef2 expression in FCMs, and like Mef2, is also essential to regulate a program of muscle differentiation, the first step of which is myoblast fusion [3], [4], [16]. lmd mutant embryos have defects in the specification or maintenance of FCMs [4], which results in an expansion of Zfh1-expressing pericardial cells and adult muscle precursor-like cells [17]. In contrast to Mef2, the molecular function of Lmd is more poorly understood; its only known direct target gene being Mef2 itself [3].
Given the extensive co-expression of lmd and Mef2 and the similarity in the myoblast fusion phenotype observed in their loss-of-function mutants, we hypothesized that Lmd may act as an FCM-specific modulator of Mef2 activity. To assess this, we have systematically compared the in vivo enhancer occupancy of Lmd and Mef2 and identified a large number of combinatorially bound enhancers during myogenesis. Expression profiling of loss-of-function lmd and Mef2 mutants revealed that, although these TFs co-occupy the same enhancer region, they have different regulatory effects on the expression of the target genes. We used a combination of in vivo and in vitro approaches to demonstrate differential integration of regulatory input from Lmd and Mef2 at individual CRMs. Taken together, these data emphasize the diversity of transcriptional responses that can be generated by two transcription factors and identify Lmd as a new context-specific modulator of Mef2 activity.
As a first step towards understanding the phenotype of lmd mutant embryos and its potential combinatorial regulation with Mef2, we set out to identify Lmd-bound enhancer regions and directly regulated target genes. To identify Lmd-bound enhancers within the developing embryo, we used chromatin immunoprecipitation followed by microarray analysis (ChIP-on-chip) during defined stages of muscle development. Lmd-associated DNA was precipitated from tightly staged embryos at two consecutive developmental time points, spanning most of the developmental stages when lmd is expressed (stages 10–13). To obtain data with high sensitivity and specificity we performed a total of eight independent chromatin immunoprecipitations per time-point using two different anti-Lmd antibodies for each time point (Materials and Methods). The enriched DNA sequences were analyzed on microarrays containing overlapping 3 kb fragments tiling across ∼50% of the Drosophila genome [11]. Genomic regions were considered bound by Lmd if they were significantly enriched with both antibodies, thereby reducing potential false targets caused by non-specific antibody effects.
Lmd binding was detected at 154 unique genomic regions at one or both developmental time points (Table S1), including the only known Lmd-binding site upstream of the Mef2 locus ([3]; Figure 1A). In addition, Lmd binds to a previously characterized enhancer of sns ([18]; Figure 1B), a transmembrane protein that requires lmd for its expression in FCMs [3] indicating that Lmd directly regulates sns expression. The expression of the bHLH transcription factor twist persists longer in lmd loss-of-function mutants than in wild-type embryos [4]. As the DNA-binding domain of Lmd is similar to the Gli-family of transcription factors, which can act both as transcriptional activators and repressors, it was proposed that Lmd may directly repress Twist [4]. However, no significant Lmd-binding was detected in the twist locus (data not shown). Although we cannot exclude low-level Lmd occupancy below the detection limit of our assay, this result suggests an indirect regulatory connection between Lmd and twist.
The recovery of enhancers of both Mef2 and sns, genes known to be genetically downstream of lmd, underscores the accuracy of the ChIP-on-chip results. Moreover, a number of Lmd-bound regions overlap previously characterized muscle enhancers, including betaTub60D [19], Act57B [20], CG14687 and CG9416 [11] (Figure S1 and Table S2) and are dependent on Lmd for their activity (see below). In addition, we have characterized the activity of four previously unknown Lmd-bound enhancers that are responsive to Lmd both in vivo (Figure 2) and in vitro (see below).
The activities of Lmd and Mef2 are required for the initiation of myoblast fusion, presumably due to the regulation of a battery of genes essential for this process or for the identity of the FCMs themselves. To investigate potential co-regulation of target genes, we compared the enhancer binding data from Lmd (presented here) to our previously reported ChIP-on-chip data for Mef2 performed at the same developmental time points [11]. 106 out of 154 (68.8%) Lmd-bound regions are co-bound by Mef2 during muscle development (Figure 1C), suggesting that the majority of these regions are co-regulated by both TFs. This is likely to be a conservative estimate as regions bound by one or both transcription factors just below our thresholds are not considered. Nevertheless, the extensive level of enhancer co-occupancy (Figure 1C) indicates that combinatorial regulation by these two TFs is an important feature within the myogenic program. In many cases the temporal profile of Lmd and Mef2 binding to shared enhancers is identical, again indicating that these two TFs act together to co-regulate enhancer output. For example, both TFs only bind to the ttk (tramtrack) enhancer at stages 10–11, but not later, while the blow (blown fuse) and CG5080 enhancers are co-bound at stages 12–13, and not earlier (Figure 3).
The combinatorial binding of Lmd and Mef2 to shared CRMs raises several interesting questions. How much regulatory input does each factor contribute to the activation of an enhancer? Are Lmd and Mef2 acting in a co-operative or additive manner to regulate target gene expression? Are both transcription factors required for enhancer activation, or do they act redundantly? We have used several approaches, both in vivo and in vitro, to address these questions. First, we used global expression profiling to determine which genes require Lmd and Mef2 activity for their correct expression in vivo (Figure 1E).
We performed a developmental time-course of gene expression, comparing the transcriptional state of wild-type embryos to that of lmd-mutant embryos at six consecutive one-hour windows of development, providing a high-resolution map of lmd-dependent changes in gene expression. These experiments identified 640 genes that are genetically downstream of lmd during the stages of myoblast fusion and the initiation of terminal muscle differentiation (Table S3). By integrating this differential gene expression data with information on Lmd enhancer occupancy (ChIP-chip data) and muscle-specific gene expression patterns (from BDGP in situ hybridizations, [21]) we defined a high-confidence set of 74 target genes [11] that are likely to be directly regulated by Lmd (Table S1 and Table S4). Among these are a number of genes known to be involved in myoblast fusion, including Mef2, sns and blow, as well as genes with characterized roles in other aspects of muscle development, suggesting that Lmd may have a broader role in myogenesis than previously anticipated.
In a previous study, we used ChIP-on-chip experiments and expression profiling to identify a stringent set of Mef2 direct target genes at multiple stages of development [11]. Comparing these data to that of Lmd revealed that a large (79.7%) and highly significant (p<2.2×10−16, Fisher's exact test) proportion of Lmd direct target genes are also directly regulated by Mef2 (Figure 1D). Thus the majority of the regulatory input provided by Lmd is mediated in conjunction with Mef2, which is not the case in the other direction. Mef2 regulates many target genes independently of Lmd, reflecting its broader expression and role in muscle development (Table S4).
To assess the regulation of Lmd direct target genes we first examined their transcriptional response to loss of regulatory input from either Lmd or Mef2. 57 of the 74 Lmd direct target genes are differentially expressed in lmd and/or Mef2 mutant embryos compared to stage-matched wild-type controls (visualized by K-means clustering in Figure 1E). This analysis revealed an unanticipated diversity in transcriptional responses, despite the fact that the majority of genes have an enhancer bound by both transcription factors. One group of genes (cluster I, Figure 1E) is downregulated in both lmd and Mef2 mutants compared to the stage matched wild-type embryos. This group contains many genes coding for structural muscle proteins, including Act57B, Act87E and betaTub60D (Figure 1E). As Mef2 expression is also strongly reduced in lmd mutant embryos [3], these target genes either depend on input from Mef2 alone or on a combination of Mef2- and Lmd-mediated activation. Other genes are affected differently in the two mutants. Several are upregulated in lmd mutants (e.g. CG9416 and CG30035), but are either unchanged or have slightly decreased or increased levels in Mef2 mutants (cluster II, Figure 1E). In contrast, a third cluster of genes, including blow, goliath (gol) and tramtrack (ttk), have decreased expression at the late time points in lmd mutants and increased expression in Mef2 mutants, suggesting activation by Lmd and repression by Mef2 (Figure 1E). We note that, although we have used several methods to assess the role of Mef2 in regulating the expression of these genes (see below), we have not found any evidence that Mef2 may act as a transcriptional repressor. Therefore the apparent de-repression of these genes is most likely due to a secondary effect within the Mef2 mutant embryos. Despite this, the vast majority of genes known to be genetically downstream of Mef2 had significantly reduced expression, indicating that the expression profiling data accurately recapitulates what is expected from genetic studies [11].
As a complementary approach to assess the regulatory inputs of Lmd and Mef2, we asked if these transcription factors are sufficient, either alone or in combination, to induce target gene expression in vivo. The transcription factors were ectopically expressed in parasegmental stripes under the control of the engrailed-Gal4 driver [22] (Figure 2). Lmd has been reported to activate Mef2 expression in the CNS but not in the remainder of the ectoderm under these conditions [16], allowing us to assess the contribution of the two transcription factors independently. As the transcription factors are acting outside of their normal cellular context, this is a stringent assay to investigate regulatory connections.
The transcriptional response of shared target genes to ectopic TF expression was examined using colorimetric in situ hybridization (ISH) (Figure 2), and confirmed by double fluorescent ISH (Figure 2, Figure S2). This analysis revealed a range of regulatory responses. We examined three genes that showed reduced expression in both lmd and Mef2 mutant embryos (Figure 1E, cluster I). betaTub60D and Act57B are ectopically induced by Mef2 alone, but not by Lmd alone (Figure 2A–2B″). As expected, co-expression of both transcription factors also led to ectopic expression (Figure 2A′″, 2B′″). A third gene, CG5080, was neither ectopically activated by Lmd nor Mef2 alone (Figure 2C′–2C″). However, when both transcription factors were co-expressed, their combined activity was sufficient to drive ectopic expression, revealing a synergistic regulation of this target gene (Figure 2C′″). Ubiquitous over-expression of Mef2 using a daughterless-Gal4 driver was previously reported to ectopically activate CG5080 in the head mesoderm [13]. The fact that Mef2 is sufficient to regulate CG5080 expression in this context, but not in ectodermal stripes, strongly suggests that Mef2 requires additional tissue-specific co-activators also in other tissues of the embryo.
The expression levels of blow and sug were also strongly reduced in lmd mutants, and weakly reduced in Mef2 mutants (Figure 1E, cluster III). Similar to CG5080, neither expression of lmd nor Mef2 alone was sufficient to activate expression of blow, sug or sns, yet ectopic activation was detected upon co-expression of both transcription factors (Figure 2D–2F′″). Although Mef2 is not required for sns expression [23], our data demonstrates that Mef2, in combination with Lmd, is sufficient to activate the expression of sns in ectodermal cells. CG14687 showed the opposite response to bTub60D and act57B, in that it could be activated by Lmd alone, but not by Mef2 (Figure 2G′–2G″). These data correlate with the expression profiling data, showing a strong requirement of lmd activity for CG14687 expression (Figure 1E, cluster IV). Although in situ hybridization is not quantitative, the fluorescent ISH suggests a higher level of expression when both Lmd and Mef2 are co-expressed (Figure 2G, fluorescent panels, Figure S2).
The gene CG9416 revealed yet another mode of regulatory integration: Mef2 activated ectopic expression of CG9416 in the absence of Lmd, but this effect appears to be attenuated when both transcription factors were co-expressed (Figure 2H–2H′″, fluorescent panel, Figure S2), indicating opposing regulatory inputs from Mef2 (activation) and Lmd (inhibition). The repressive effect of Lmd is consistent with the dramatic increase in CG9416 expression in lmd mutant embryos (Figure 1E). Finally, the gol gene represents the only example tested where even the combination of both Mef2 and Lmd was not sufficient to induce ectopic expression (Figure 2I′″).
In summary, although all genes investigated are directly co-regulated by Lmd and Mef2, ectopically supplying one or both factors revealed considerable flexibility in how information is integrated at each individual locus. In higher eukaryotes, many genes have multiple regulatory elements, which collectively contribute to the complete expression pattern of a gene. To investigate whether the different transcriptional responses to Lmd and Mef2 activity are reflected by the integration of inputs at single enhancers or by the combined activity of multiple cis-regulatory elements, we next studied regulatory integration at the CRM level.
Individual enhancer regions in Drosophila commonly range from 0.5 to 1 kb in size. The Lmd-bound DNA fragments immunoprecipitated in our ChIP experiments were in a similar size range, however the genomic tiling arrays used in this study limited our resolution to overlapping 3 kb sequences. To achieve higher resolution, we used quantitative real-time PCR to assay the enrichment of shorter sequences within individual 3 kb-bound regions using both the Lmd and Mef2 chromatin immunoprecipitates. In all eight cases examined, the highest enrichment of Mef2 and Lmd binding coincided within a common 0.1 to 1 kb region (data not shown), suggesting that the transcription factors co-occupy a single enhancer element. In addition, each refined sequence was found to contain at least one Mef2 consensus binding site conserved in Drosophila pseudoobscura (data not shown).
We tested the ability of the refined Lmd-Mef2-bound regions to regulate expression in vivo by generating transgenic reporter lines. All tested enhancer regions specifically activated GFP-reporter expression in the developing muscle (Figure 3 and Figure S1). At stage 11, when both lmd and Mef2 are co-expressed in fusion-competent myoblasts, the enhancers of blow (Figure 3A′) and gol (Figure 3C′) activated GFP-expression broadly in the visceral and somatic mesoderm. At the same stages, the CG5080 (Figure 3B′) and tramtrack enhancers (ttk, Figure 3D′) induced GFP-expression in a subset of myoblasts. At stage 13, when myoblast fusion is in progress, all four enhancers showed almost identical expression patterns throughout the somatic muscle (Figure 3A″–3D″). We also re-examined the spatio-temporal activity of the previously characterized Act57B [20], betaTub60D [19] and CG14687 [11] enhancers (Table S2 and Figure S1) and included them in the set of combinatorially-bound enhancers investigated in the remainder of this study.
We used the in vivo enhancer-reporter lines to study the integration of Lmd and Mef2 regulatory input by comparing CRM activity in wild-type and mutant embryos. Six transgenic reporter lines (Figure 3, Table S2) were placed in the genetic background of lmd1 and Mef222.21, two characterized loss-of-function alleles for these transcription factors [2], [3] (Figure 4, Figure S3).
The expression of the betaTub60D gene is controlled by several independent cis-regulatory modules [19], [24], [25]. An upstream enhancer, 5′ to the betaTub60D gene, requires Mef2 activity for its full activation [24]. In contrast, the intronic betaTub60D enhancer under study here, although co-occupied by Mef2 and Lmd, appeared unaffected in Mef2 mutant embryos while having strongly reduced expression in lmd mutants (Figure 4A′ and 4A″). The strong reduction in the expression of the betaTub60D gene in Mef2 and lmd mutant embryos detected by expression profiling (Figure 1E) therefore reflects the combined activity of at least two enhancers: one strongly responsive to Mef2 levels and a second one depending on Lmd (but not Mef2) for activation.
The Act57B enhancer drives GFP-expression in somatic and visceral muscles in wild-type embryos at stage 13 (Figure 4B). This expression was completely lost in lmd mutant embryos (Figure 4B′), while Mef2 mutant embryos showed reduced, but detectable reporter expression, as observed previously [20] (Figure 4B″). In contrast, expression driven by the CG5080 enhancer was reduced in lmd (Figure 4C and 4C′) and to a lesser extent in Mef2 mutant embryos (Figure 4C and 4C″). Similarly, reporter expression in the somatic muscle driven by the blow enhancer was lost in both lmd and Mef2 mutant embryos (Figure 4D–4D″). Enhancer expression in the hindgut visceral muscle persisted in Mef2 mutant embryos (Figure 4D″), indicating additional tissue-specific input at this enhancer.
The CG14687 enhancer is activated in somatic and visceral muscle in wild-type embryos (Figure 4E). Expression in somatic muscle required lmd expression (Figure 4E′), but is unaffected in Mef2 mutant embryos (Figure 4E″). Interestingly, expression in the visceral muscle was independent of both lmd and Mef2 expression (Figure 4E′ and 4E″), implicating additional tissue-specific factors in the activation of this enhancer. Both the homeodomain transcription factor bagpipe (bap) and the fork head domain transcription factor biniou are recruited to this enhancer in vivo [26] and most likely activate gene expression in this tissue. Finally, the gol enhancer required lmd activity (Figure 4F and 4F′), but robustly activated gene expression in the absence of Mef2 (Figure 4F″).
In summary, all six muscle enhancers examined showed reduced activity in one or both mutant conditions, demonstrating that the in vivo occupancy of these modules by Mef2 and Lmd has regulatory function. lmd mutants generally displayed a stronger reduction in enhancer activity compared to Mef2 mutant embryos. As Lmd is required to maintain Mef2 expression, lmd mutant embryos are effectively double mutants for both transcription factors. This is reflected by the stronger reduction in enhancer activity in this genetic background and underscores the combinatorial regulation of these enhancers by both transcription factors.
We next assessed if the combined regulatory inputs of Lmd and Mef2 on these enhancers are integrated in an additive, cooperative or repressive manner. Drosophila S2 cells, which express neither endogenous lmd nor Mef2 [27], were used to study the regulatory logic of the different enhancers in vitro. Eight regulatory regions that are co-bound by Lmd and Mef2 in vivo were placed upstream of a minimal Hsp70 promoter driving a firefly luciferase reporter and co-expressed with increasing amounts of Lmd and/or Mef2 expression vectors.
Co-transfection of either Lmd or Mef2 alone was sufficient to activate the CG14687, CG5080 and gol enhancers (Figure 5A–5C), while co-expression of both regulators led to an approximately additive level of reporter activity. For example, transfection of 10 ng of the Lmd expression vector led to a 3.6 fold increase in the luciferase activity driven by the CG14687 enhancer, while 1 ng of the Mef2 expression vector led to a 2.4 fold increase in expression. Co-expression of both factors resulted in a 5 fold increase in enhancer activity (Figure 5A).
In contrast, Lmd and Mef2 acted cooperatively to regulate the ttk, blow and betaTub60D enhancers (Figure 5D–5F). For example, expression of either Lmd or Mef2 alone yielded only low levels of reporter gene activity via the blow enhancer (Figure 5E). However, co-expression of both transcription factors resulted in much higher levels of activity, indicating a cooperative interaction between Lmd and Mef2 in the context of this enhancer.
Conversely, the CG9416 enhancer is readily activated by Mef2, but cannot be induced by Lmd (Figure 5G). Instead, co-expression of both transcription factors revealed that Lmd counteracts the positive input of Mef2 to this module, essentially blocking activation by Mef2 in a dose dependent manner. This repressive activity, in combination with the in vivo occupancy of Lmd on this enhancer (Figure S1), the increase in CG9416 gene expression in lmd mutant embryos (Figure 1E) and the ability of Lmd to attenuate the ectopic activation of CG9416 by Mef2 (Figure 2H, fluorescent panels), provides strong evidence that Lmd can provide direct inhibitory input to enhancer activity. Similar to CG9416, the expression of endogenous CG30035 was de-repressed in lmd mutant embryos (Figure 1E). The Lmd-Mef2 bound enhancer region close to the CG30035 locus displayed a similar dose-dependent inhibitory effect of Lmd on Mef2-mediated transcriptional activation (Figure 5H).
Collectively, our results demonstrate that Lmd and Mef2 can induce different regulatory responses depending on the context of the enhancer. This may reflect differences in the relative positioning of Mef2 and Lmd binding to each other or the recruitment of additional unknown factors. As there is no consensus binding site known for Lmd, we used de novo motif discovery [28] to identify possible Lmd binding motifs. Since we observed that Lmd and Mef2 are commonly bound within close proximity to each other, we reduced the search space to a 400 bp window around each predicted Mef2 site within the group of 57 co-bound regions. This analysis did not reveal any candidate motifs matching the only known site occupied by Lmd [3], precluding further analysis of individual Lmd binding sites.
Metazoan cells must activate and inactivate the expression of large cohorts of genes in a precise spatio-temporal manner to progress through development. To achieve a molecular understanding of the regulatory networks controlling cellular decision-making, it is essential to understand how inputs from different regulators are being integrated to give rise to defined patterns of gene expression. In this study, we approached this challenge from a genomic perspective by examining the combinatorial input of two key myogenic regulators, Mef2 and Lmd. ChIP-on-chip experiments and expression profiling of loss-of-function mutants were used to systematically identify the direct target genes of the zinc-finger protein Lmd, an important regulator of myogenesis, for which only a single target gene had previously been identified. Integrating these data with data previously obtained for Mef2 revealed that Lmd regulates the majority of its targets in a combinatorial manner together with Mef2. In a few cases these two transcription factors target the same locus through different regulatory regions (e.g. ladybird-early, PAK-kinase or short stop), however in the majority of cases Lmd- and Mef2-binding could be mapped to the same genomic location (Table S1, Table S4). Examining the contribution of both Lmd and Mef2 to regulatory activity, at both the enhancer and gene level, revealed a number of important insights into the contribution of both transcription factors to the myogenic developmental program.
Genes that are co-regulated by the same two (or more) transcription factors are generally expected to have very similar spatio-temporal expression profiles. In fact, this assumption has been used by many studies to computationally predict the location of enhancer elements by searching for common TF binding motifs in the vicinity of clusters of co-expressed genes (or synexpression groups) [29]–[32]. It was therefore surprising when our comparison of experimentally-identified enhancer regions bound by the same two transcription factors uncovered a diverse range of regulatory responses. The 59 genes with enhancer elements co-bound by Lmd and Mef2 at the same stages of development are regulated either in a cooperative, additive or repressive manner depending on the individual enhancers. These data suggest that enhancer regions integrate regulatory inputs more flexibly than previously anticipated. By focusing on individual enhancer elements, we evaluated how Lmd and Mef2 influence regulatory activity in different contexts both in vivo and in vitro. Combining a number of complementary approaches allowed us to identify three different modes of TF integration at developmental enhancers leading to additive, cooperative or repressive regulation.
Mef2 and Lmd provide an additive positive input to the regulation of the Act57B locus. Ectopic Mef2 expression in the ectoderm is sufficient to induce Act57B expression, while providing Lmd alone is not (Figure 2B–2B″). Conversely, enhancer-reporter gene expression is completely lost in lmd mutant embryos and only slightly reduced in Mef2 loss-of-function mutant embryos (Figure 4B–4B″). Together, these data reveal a role for both transcription factors at this enhancer. Previous studies demonstrated that the initiation of Act57B expression at stage 11 requires Mef2 for its activation. Yet, artificially increasing Mef2 levels at this stage does not lead to premature activation of this locus [13]. Our findings offer an explanation for this result: at this stage of development, combined input from Lmd and Mef2 is required to drive gene expression, while the presence of Mef2 alone is not sufficient to activate transcription. At later stages, when lmd expression is lost, Mef2 concentration has increased sufficiently to maintain Act57B expression. Conversely, the CG14687 locus can be activated by ectopic Lmd in the ectoderm, but not by Mef2 alone (Figure 2G′–2G′″) and requires lmd, but not Mef2, for its expression in the somatic muscle (Figure 4E′–4E″). Combined ectopic expression of the two TFs, on the other hand, leads to a marked increase of reporter signal, again indicating combinatorial positive regulation by both TFs (Figure 2G, fluorescent panels). These findings are supported by the ability of both Lmd and Mef2 to separately activate reporter gene expression in vitro and to yield additive reporter activity in combination (Figure 5A).
The blow enhancer shows a different mode of regulation and is synergistically activated by both factors. While neither Mef2 nor Lmd alone are sufficient to activate ectopic gene expression in vivo, supplying both factors simultaneously leads to robust target gene expression (Figure 2D–2D′″). Assaying for reporter gene activation in the two mutant backgrounds yields a complementary result; Mef2 and Lmd activity is required to activate transcription in the somatic mesoderm via the blow enhancer (Figure 4D–4D″). Moreover, the in vitro reporter assay reveals a positive interaction between the two proteins (Figure 5E), indicating that the blow enhancer functions as a cooperative switch.
Analysis of the CG9416 enhancer revealed an antagonistic interaction between Lmd and Mef2. While ectopic expression of Mef2 leads to enhancer activation (Figure 2H–2H″), simultaneous expression of Lmd markedly attenuates the transcriptional output from this locus (Figure 2H, fluorescent panels, Figure S2H). This effect can be reproduced in vitro: while providing Mef2 alone leads to robust activation of the CG9416 enhancer, Lmd is not able to activate gene expression (Figure 5G). Instead, Lmd antagonizes the activating input of Mef2 in a concentration-dependent manner. To our knowledge, this is the first example of direct negative regulation by Lmd. To identify additional examples of a repressive role for Lmd, we re-examined the expression profiles of lmd and Mef2 mutant embryos (Figure 1E). CG9416 is markedly upregulated in lmd mutants, but shows reduced expression in embryos lacking Mef2 (Figure 1E). We selected another direct target gene with similar expression changes in these genetic backgrounds, CG30035 (Figure 1E) and after determining the limits of the ChIP-enriched region we assayed its ability to drive reporter gene expression in vitro. Similar to the CG9416 enhancer, the CG30035 enhancer is robustly activated by Mef2, and this activation is inhibited by Lmd in a dose-dependent manner (Figure 5H). This provides a second, independent example for Lmd-mediated repression of gene expression.
In summary, starting from a genomic perspective, we have identified a large cohort of genes co-regulated by a pair of tissue-specific transcription factors. Lmd modulates the activity of Mef2 at different enhancers in a context-dependent fashion, allowing for additive, cooperative or antagonistic interactions in the same cells. In this way, the timing and expression levels of Mef2 target genes can be further refined, as exemplified by the Act57B locus, which may owe its early activation during embryonic development to the combined activity of both proteins. Lmd shows homology with the Gli superfamily of transcription factors [3], which can act both as transcriptional activators and repressors, depending on proteolytic cleavage regulated by the hedgehog signaling pathway. To date, there is no evidence for proteolytic cleavage of Lmd and an irreversible conversion of Lmd from a transcriptional activator to an inhibitor is difficult to reconcile with our observation that Lmd can perform both roles at different loci at the same time, in the same tissue. For the same reason, we also consider it unlikely that Lmd interferes with transcriptional activation simply by binding to Mef2 and sequestering the protein in the cytoplasm. Instead, we propose that Lmd exerts a dominant inhibitory influence over a transcriptional activator, either by locally quenching Mef2's activity or through direct repression of the locus, similar to transcriptional repressors described in other developmental networks [33], [34]. Our results provide a molecular understanding for the genetic observation that restoring Mef2 activity in lmd mutant embryos is not sufficient to rescue muscle differentiation [4]. Both transcription factors are required to provide different regulatory inputs to a large number of co-regulated target genes during myogenesis. Their associated enhancers have revealed considerable flexibility in integrating regulatory inputs from these two TFs at individual cis-regulatory regions.
Embryo collections and chromatin immunoprecipitations were performed as described previously [11], [35]. Two antisera were raised against the amino terminus of Lmd and purified from E. coli by poly-His tag affinity purification. Four independent staged wild-type embryo populations were collected at 6–8 and 8–10 hrs after egg-laying and fixed with formaldehyde. For each time point, chromatin from all four populations was precipitated with both antisera as well as the respective preimmunesera, leading to a total of 16 reactions (8 mock, 8 anti-Lmd) per time point. DNA amplification, labeling and hybridizations were performed as described previously [11], [35] and dye swaps were included to account for possible dye biases.
The assayed lmd1 [3] line was outcrossed to wild-type flies (Canton S) twice to remove any spurious mutants. Six one-hour embryo collections were assayed in an expression profiling timecourse (between 5 and 11 hours after egg-laying). At each time point, 4 independent populations of lmd mutant and stage-matched Canton S embryos were collected and aged. Homozygous mutants were selected with an automated embryo sorter [16], [36]. The staging of all collections was verified by formaldehyde fixation of a small sample to ensure that wild-type and mutant embryos were tightly stage matched. Total RNA was extracted using Trizol (Invitrogen, Carlsbad, US), amplified, reverse-transcribed and labeled as described previously [11].
For expression profiling analysis, mutant and stage-matched control cDNA was hybridized directly against each other. Raw data was normalized using print-tip LOESS. Differentially expressed genes were identified using Significance analysis of microarrays (SAM) [37]. Genes with a q<1% and a fold change >1.6 (log2>0.7 or <−0.7) were considered to be differentially regulated (Table S5). Immunoprecipitated DNA from Lmd-specific or mock precipitations was hybridized against a total genomic reference DNA sample. Sequences significantly enriched by the anti-Lmd-antibodies were identified by comparing rank products [38] and the false-discovery rate was estimated. Only fragments with an FDR <2% and a fold enrichment >1.5 (log2 >0.58 or <−0.58) were considered to be significantly enriched (Table S1).
Automatic assignment of ChIP-enriched fragments to target genes was performed as described previously [11]. The majority of regions co-occupied by Mef2 and Lmd was independently assigned to the same target genes using either Mef2-mutant or lmd-mutant expression profiling data. For a small number of regions, data from this study indicated a more likely target gene than had been assigned previously with Mef2 data alone [11]; in these cases, we chose the updated target prediction for further analysis. A complete list of ChIP-enriched regions, expression profiling results and target assignments are available in Tables S1, S3, S4, S5. All raw microarray data is available from ArrayExpress (Lmd ChIP (E-TABM-895) and lmd expression profiling (E-TABM-894). Lmd- and/or Mef2-bound regions and mutant expression data can be visualized at http://furlonglab.embl.de/data/.
Fragments within the following coordinates (based on BDGP genome release 5) were cloned into the pH-stinger (AF242365) vector for germline transformation [39]: chr2R:16831306-16831372 (actin57B), chr2R:20197035-20197429 (betaTub60D), chr3R:6619371-6620063 (CG14687), chr3R:27529661-27530409 (ttk), chr2R:3472616-3473387 (blow), chr3R:27538572-27539618 (ttk), chr2R:8813219-8814579 (sug), chr2R:20966587-20969610 (gol). For all constructs at least two independent transgenic lines were obtained and assayed. The UAS-Mef2 line used in this study has been described previously [11]. The UAS-lmd line was previously referred to as UAS-gfl [16].
Double fluorescent in situ hybridizations were done as described previously [16]. To minimize experimental differences, the embryo fixations and the in situ hybridizations were done in parallel and the confocal imaging was performed with identical laser and gain settings for each gene in the four genetic backgrounds. The following ESTs were used to generate digoxigenin or fluorescein-labeled probes: RE53159 (betaTub60D), LD04994 (act57B), LD34147 (CG5080), LP02193 (blow), LD36528 (sug), RE74890 (CG14687), RE28322 (CG9416), GH20973 (gol), AT15089 (twi) and RE02607 (wg). The full-length sns cDNA was a kind gift from S. Abmayr. GFP expression in transgenic animals was detected by immunohistochemistry with rabbit α-GFP antibody (Torrey Pines Biolabs) at a concentration of 1:500. Biotinylated secondary antibodies were used in combination with the Vector Elite ABC kit (Vector Laboratories).
Drosophila S2 cells were transiently transfected using Cellfectin (Invitrogen). Lmd and Mef2 were expressed from full-length ESTs (LD47926 and GH24154, respectively) in pAc5.1 vector (Invitrogen). The enhancers (coordinates given above) were assayed in a pGL3 luciferase reporter vector (Promega) with an Hsp70 minimal promoter and the luciferase activity was normalized to Renilla standard (Promega). The total amount of transfected DNA was kept constant by supplementing empty pAc5.1 vector. The measurements were performed according to the supplier's recommendations (Dual-Luciferase Reporter Assay, Promega) with a PerkinElmer 1420 Luminescence Counter.
|
10.1371/journal.ppat.1002568 | Short ORF-Dependent Ribosome Shunting Operates in an RNA Picorna-Like Virus and a DNA Pararetrovirus that Cause Rice Tungro Disease | Rice tungro disease is caused by synergistic interaction of an RNA picorna-like virus Rice tungro spherical virus (RTSV) and a DNA pararetrovirus Rice tungro bacilliform virus (RTBV). It is spread by insects owing to an RTSV-encoded transmission factor. RTBV has evolved a ribosome shunt mechanism to initiate translation of its pregenomic RNA having a long and highly structured leader. We found that a long leader of RTSV genomic RNA remarkably resembles the RTBV leader: both contain several short ORFs (sORFs) and potentially fold into a large stem-loop structure with the first sORF terminating in front of the stem basal helix. Using translation assays in rice protoplasts and wheat germ extracts, we show that, like in RTBV, both initiation and proper termination of the first sORF translation in front of the stem are required for shunt-mediated translation of a reporter ORF placed downstream of the RTSV leader. The base pairing that forms the basal helix is required for shunting, but its sequence can be varied. Shunt efficiency in RTSV is lower than in RTBV. But in addition to shunting the RTSV leader sequence allows relatively efficient linear ribosome migration, which also contributes to translation initiation downstream of the leader. We conclude that RTSV and RTBV have developed a similar, sORF-dependent shunt mechanism possibly to adapt to the host translation system and/or coordinate their life cycles. Given that sORF-dependent shunting also operates in a pararetrovirus Cauliflower mosaic virus and likely in other pararetroviruses that possess a conserved shunt configuration in their leaders it is tempting to propose that RTSV may have acquired shunt cis-elements from RTBV during their co-existence.
| Ribosome shunting, first discovered in plant pararetroviruses, is a translation initiation mechanism that combines 5′ end-dependent scanning and internal initiation and allows a bypass of highly-structured leaders of certain viral and cellular mRNAs. Here we demonstrate that a similar shunt mechanism has been developed by the RNA picorna-like virus RTSV and the DNA pararetrovirus RTBV that form a disease complex in rice. Leader sequences of the RTSV genomic RNA and the RTBV pregenomic RNA possess a conserved shunt configuration with a 5′-proximal short ORF (sORF1) terminating in front of a large stem-loop structure. Like in RTBV and a related pararetrovirus Cauliflower mosaic virus, shunt-mediated translation downstream of the RTSV leader depends on initiation and proper termination of sORF1 translation and on formation of the basal helix of the downstream secondary structure. Given that RTBV-like shunt elements with identical sequence motifs are present in all RTSV isolates but absent in related picorna-like viruses, it is likely that RTSV could have acquired these elements after its encounter with RTBV. Alternatively, the RTSV shunt elements could have evolved independently to adapt to the rice translation machinery. Our study highlights on-going genetic exchange and co-adaptation to the host in emerging viral disease complexes.
| Rice tungro disease is a significant constraint for rice cultivation in South and Southeast Asia. It is caused by a synergistic interaction of two viruses, Rice tungro bacilliform virus (RTBV) and Rice tungro spherical virus (RTSV). Individually these viruses exhibit rather mild symptoms: RTSV causes mild or indistinct symptoms, whereas RTBV infection causes yellowing and reddening of the leaves and results in stunted growth. The RTBV symptoms are accentuated in plants co-infected with RTBV and RTSV. Moreover, RTBV on its own cannot be transmitted from plant to plant, but it can do so with the help of RTSV that encodes an insect transmission factor [1]. This suggests that the two viruses have co-evolved into a unique disease complex, in which partners may have developed not only specialized but also shared mechanisms enabling the complex to establish systemic infection and to accumulate in the same plant tissues in order to be co-transmitted. Indeed, both RTBV and RTSV are phloem-restricted. It can be further suggested that during converging evolution the two viruses may have exchanged or independently developed certain cis-acting elements and sequence motifs to adapt to the host cell machinery and to synchronize their life cycles. Our study provides initial evidence for this hypothesis.
RTSV belongs to genus Waikavirus in the family Secoviridae of picorna-like viruses [2]. Its single-stranded, polyadenylated genomic RNA of 12.4 kb contains one large ORF encoding a viral polyprotein [3]. The polyprotein ORF is preceded with an unusually long leader sequence (514-nt in the type species NC_001632) which has several short ORFs (sORFs) and a high propensity to form stable secondary structure (see below): both features are known to inhibit 5′ end-dependent, scanning-mediated translation initiation on eukaryotic ribosomes [4]. Thus, translation of RTSV genomic RNA may involve either internal ribosome entry or 5′ end-dependent ribosome shunting. An internal initiation mechanism operates in animal picornaviruses that possess long and highly-structured leaders [5] and it is therefore an attractive possibility that plant picorna-like viruses have also evolved an internal ribosome entry site (IRES) to initiate translation. However, so far there is little evidence that viruses of the family Secoviridae use internal initiation of translation and the IRES elements identified in short leaders of two distinct viruses from the family Potyviridae do not resemble each other and those of animal picornaviruses [6]. Instead, compelling evidence indicates that plant pararetroviruses have evolved a ribosome shunt mechanism, which combines features of 5′ end-dependent scanning and internal initiation, to translate their pregenomic RNAs that all possess long and highly structured leaders [7]–[10].
RTBV is the only member of genus Tungrovirus in the family Caulimoviridae of pararetroviruses [11]. Its circular double-stranded DNA genome of 8 kbp is transcribed by Pol II into a pregenomic RNA (pgRNA) of more-than-genome length as a poly(A) signal located 195 bp downstream of the transcription start site is recognized efficiently only at its second encounter. The pgRNA is a polycistronic mRNA for three consecutive overlapping ORFs (I, II and III) that are translated by a leaky scanning mechanism [12]. This mechanism operates efficiently owing to the lack of additional AUGs within about 1 kb region between the start codons of ORFs I and III, the feature also conserved in a closely-related badnaviruses (genus Badnavirus of the Caulimoviridae) which have similar organization of ORFs I–III [13]. Unlike badnaviruses, RTBV has an additional ORF, ORF IV, located downstream of ORF III. This ORF is translated from a spliced version of pgRNA, in which the first sORF of the pgRNA leader is fused to ORF IV [14].
Translation of RTBV pgRNA is initiated by ribosome shunting that overcomes the obstacles of a 700-nt leader sequence with multiple sORFs and a stable stem-loop structure [8]. This mechanism operates efficiently in rice protoplasts and involves (i) 5′ end-dependent ribosome scanning until the first sORF is encountered, (ii) translation of this sORF and its termination just in front of the stem basal helix, the formation of which is crucial for efficient shunting, (iii) ribosome shunting over the structured region, and (iv) resumption of scanning at the shunt landing site, where a fraction of the shunting ribosomes (about 10%) also initiates translation at the AUU start codon of ORF I [8], [15] (Figure 1A). The RTBV shunt strikingly resembles the shunt mechanism evolved by Cauliflower mosaic virus (CaMV) from genus Caulimovirus of plant pararetroviruses [7], [15]. Notably, in both cases, initiation and proper termination of the first sORF translation (but not an encoded peptide) are essential for shunting. Furthermore, the RTBV shunt elements including the sORF, the stem base section and the shunt landing sequence could functionally replace the corresponding elements in the CaMV genome in driving efficient polycistronic translation of CaMV pgRNA and in supporting infection of the chimeric virus in CaMV-host plants [16].
The shunt configuration comprising an sORF terminating in front of the stable secondary structure has been identified in the pgRNA leader of most plant pararetroviruses [9], suggesting its evolutionary conservation within this family. Whether or not a shunt mechanism was also evolved in other families of plant viruses remained unknown so far. It is worth mentioning that an sORF-dependent shunt mechanism of the CaMV/RTBV-type has also evolved in a human spumavirus [17] and a human gene [18]. Here we provide evidence that sORF-dependent ribosome shunting operates in RTSV.
Our computer-aided comparison of the 697-nt RTBV and the 514-nt RTSV leader sequences revealed remarkable similarities, suggesting that RTSV has co-evolved ribosome shunting (Figure 1):
Thus, all the cis-acting elements known to drive ribosome shunting in RTBV are also present in RTSV, strongly supporting the idea that RTSV could have co-evolved an sORF-dependent shunt mechanism. Moreover, the identity of certain sequence motifs within these elements raises a possibility of their horizontal transfer from one virus to another during co-evolution. Alternatively, these motifs could have co-evolved independently through adaptation to the rice translational machinery.
To test the hypothesis that translation of RTSV genomic RNA is initiated by an sORF1-dependent ribosome shunting, we used well-established translation assays based on rice protoplasts and wheat germ extracts, in which translation of a reporter ORF encoding chloramphenicol acetyl transferase (CAT) placed downstream of the RTSV leader sequence or its mutant versions was monitored. We followed the same experimental settings and protocols as those used previously in a comparative study of molecular mechanisms of the RTBV and CaMV shunting [15].
In rice protoplasts, both RTSV and RTBV leaders drove relatively efficient translation of the reporter ORF, although the RTBV leader allowed a 1.6-fold higher initiation rate. Confirming our previous results, knock out (KO) mutations of the start (AUG to UAG) or stop (UAG to UAC) codon of RTBV sORF1 drastically reduced translation (Figure 2). The same KO mutations of the RTSV sORF1 start or stop codons resulted in a significant decrease in downstream translation, albeit less dramatic than in the case of RTBV. This indicates that translation initiation downstream of the RTSV leader is sORF1-dependent, which is not consistent with internal ribosome entry at the 3′ end of the leader. Interestingly, the stop codon KO had a more pronounced effect by reducing the translation rate to 28%, whereas the start codon KO reduced translation only to 56% of the wild type level. This suggests that the RTSV leader lacking the first sORF AUG allows a relatively efficient linear ribosome migration towards the 3′ end, i.e. by leaky scanning through the remaining five AUGs and/or translation at some of the remaining five sORFs followed by reinitiation event(s). In CaMV, such a linear ribosome migration along the leader sequence has been investigated by mutating nine AUGs individually and in combinations and found to be 5 times less efficient than ribosome shunting in plant protoplasts [20] and wheat germ extracts [24]. In the case of RTBV, linear ribosome migration is even less efficient, likely because of a larger number of the intervening AUGs (twelve) and sORFs (eleven) (Figure 1). The KO of stop codon should not affect the initiation step of sORF1 translation but should result in termination of this translation event downstream of the shunt take-off site, which would diminish shunting but would not affect linear ribosome migration following sORF1 translation.
To further verify that sORF1-dependent translation downstream of the RTSV leader is initiated by ribosome shunting and evaluate a contribution of linear ribosome migration, we used a 40-nt Kozak-stem (KS) sequence which forms a perfect, compact stem-loop structure and blocks linear migration of scanning ribosomes [25]. In the case of RTBV and CaMV, insertion of KS in the leader region upstream of the first sORF abolished downstream translation, whereas its insertion within the leader region which is bypassed by shunting ribosomes had no dramatic effect on downstream translation [7], [8], [15], [20]. Likewise, insertion of KS at the 5′-end of the wild-type RTSV leader or its mutant versions with the sORF1 start or stop codon KO mutation nearly abolished downstream translation (Figure 3). This indicates that translation initiation in RTSV is 5′ end-dependent, thus ruling out internal initiation. Insertion of KS in the middle of the wild type RTSV leader did not abolish downstream translation, although the initiation rate was reduced to 42%. With KS inserted in the middle of the RTSV leader, KO mutation of either start or stop codon of sORF1 abolished downstream translation (Figure 3).
Taken together, we conclude that almost half of the ribosomes entering at the 5′ end of the RTSV leader and initiating translation of sORF1 are able to shunt over the structure and re-initiate translation at the 3′-end of the leader. Notably, like in CaMV and RTBV, this mechanism depends on proper termination of sORF1 translation in front of the structured region. Extension of RTSV sORF1 by the stop codon KO mutation should lead to termination at the in-frame stop codon located 10 triplets downstream, i. e. within the ascending arm of the structure. This would melt the stem basal helix and bring the terminating ribosome away from the take-off and landing sites.
To test if the stem basal helix structure is required for RTSV shunting, twelve point mutations were introduced either in its 5′-proximal or 5′-distal arms, which would disrupt secondary structure, and the compensatory mutations in both arms, which would restore stable secondary structure (Figure 4). The basal helix mutants with and without the KS sequence in the middle of the RTSV leader were constructed. Transient expression of the resulting constructs in rice protoplasts showed that disruption of the basal helix drastically reduced translation downstream of the RTSV leader, whereas restoration of the helix structure by the compensatory mutations almost fully restored downstream translation (Figure 4). We conclude that integrity of stable secondary structure but not primary sequences involved in formation of the stem basal helix is essential for ribosome shunting in RTSV.
Interestingly, in the absence of KS, the mutations in the 5′-proximal arm nearly abolished translation (5% of the wild type level), whereas the mutations in the 5′-distal arm reduced translation to 36% of the wild type level. The latter mutations in the presence of KS nearly abolished downstream translation (5% of the wild type level) (Figure 4). This suggests that, besides shunting, linear ribosome migration following translation of sORF1 is also abolished by the mutations in the primary sequence just downstream of sORF1. By contrast the mutations of the 5′-distal arm sequence located far away of sORF1 do not appear to affect linear ribosome migration, which would account for relatively high translation efficiency in this case, comparable to the translation efficiency of the RTSV constructs lacking sORF1. Notably, the negative effect of the 5′-proximal arm mutations is also evident when the RTSV basal helix is restored by compensatory mutations in the 5′-distal arm.
We have established previously that, in wheat germ extracts supporting efficient ribosome shunting driven by the CaMV shunt elements [23], [24], the RTBV shunting is about 7 times less efficient [15]. This is entirely due to incompatibility of the RTBV landing sequence, because the wheat germ translation machinery prefers A-rich rather than U-rich sequences and perhaps other unknown cis-elements present in the CaMV landing site but absent in the RTBV one [15]. Similar to RTBV, translation downstream of the RTSV leader was also relatively inefficient in wheat germ, although the RTSV leader allowed a 1.6-fold higher initiation rate (Figure 2). The KO mutation of sORF1 start codon increased downstream translation 1.9-fold. This is unlike RTBV, in which the sORF1 start codon KO reduced downstream translation about 4-fold (Figure 2). As discussed above, the RTSV leader allows much more efficient linear ribosome migration downstream of sORF1 than the RTBV leader, which explains a positive effect of the RTSV sORF1 start codon removal in the wheat system where shunt efficiency is diminished. KO mutation of the RTSV sORF1 stop codon abolished downstream translation in the wheat system (Figure 2). This shows that most of translation downstream of the RTSV leader depends on proper termination of sORF1 translation.
Taken together, we demonstrate here that translation initiation of RTSV genomic RNA is controlled by its long leader and mediated largely by sORF1- and stem basal helix-dependent ribosome shunting. Further research is needed to characterize this mechanism in more detail. But given the striking similarity of all the shunt elements in RTSV and RTBV and especially the identity of certain sequence motifs in the shunt take-off and landing sites, it is very likely that both RTSV and RTBV use a similar shunt mechanism.
Our comparison of five isolates of RTSV (NC_001632 and AM234048, AM234049, U71440, and AB064963) showed that the leader sequence is remarkably conserved with only 35 polymorphic positions including 33 single nucleotide substitutions and 1-nt and 2-nt insertions/deletions (not shown). Only four substitutions occur in the shunt elements – one in the shunt landing sequence between the pyrimidine stretch and the non-AUG codon and three in the stem basal helix primary sequence (but not in the secondary structure) (Figure 1B). Notably, in some regions downstream of the leader, RTSV sequences have a much higher polymorphism than the leader itself (not shown).
We conclude that the shunt elements are well preserved in all RTSV isolates, indicating their biological importance for the virus. When an infectious clone of RTSV becomes available it will be important to test the role of sORF1 and other cis-elements identified in this study for viral infectivity. Previously, it has been shown that sORF-dependent ribosome shunting is essential for infectivity of CaMV [21] and that the RTBV shunt elements can functionally substitute for the corresponding CaMV elements in systemic infection with a chimeric virus [16].
Maize chlorotic dwarf virus (MCDV), the second recognized member of genus Waikavirus [2], also possesses a long leader (434-nt in the type species NC_003626) with several sORFs (a total of 6 AUGs in NC_003626) and stable secondary structure (−154 kcal/mole in NC_003626; Figure S2). However, this leader sequence is highly polymorphic in three known isolates (less than 40% nucleotide identity) (data not shown). This suggests that a translation initiation mechanism may not be conserved. Interestingly, in all three isolates the first sORF in the MCDV leader is preserved in length (5 codons) but not nucleotide content. However, it terminates 145 nts upstream of the main structure in NC_003626 (Figure S2), which is not compatible with ribosome shunting. Furthermore, owing to high polymorphism, the shape, stability and position of the main structure are not preserved in MCDV isolates and the number and configuration of sORFs is also variable (data not shown). This again argues against shunting as the initiation mechanism. Nevertheless, the preservation of the first sORF suggests its importance in controlling translation initiation on MCDV genomic RNA which may occur via linear ribosome migration following translation of the first sORF. In support of this hypothesis, our above results for the RTSV leader indicate that in addition to ribosome shunting, sORF1-dependent linear ribosome migration also contributes to translation initiation downstream of the leader. It can therefore be proposed that in waikaviruses a linear ribosome migration-dependent mechanism has evolved earlier than shunting and that the ribosome shunt is a so-far unique acquisition by RTSV following its encounter with RTBV in a disease complex. However we cannot exclude an independent evolution of ribosome shunting in RTSV in the process of adaptation of the virus to the host plant translational machinery.
Among other viruses of the family Secoviridae, Parsnip yellow fleck virus (PYFV), the only recognized member of genus Sequivirus, is most closely related to RTSV and MCDV [2]. Unlike RTSV and MCDV, this virus has a shorter leader sequence (278 nts) that does not contain sORFs and cannot fold into stable secondary structure as predicted by MFOLD (data not shown). This suggests a linear scanning-dependent mechanism of translation initiation in PYFV. We cannot rule out, however, that PYFV (and MCDV) may use an internal initiation mechanism similar to that of potyviruses [5].
It is thought that the ribosome shunt mechanism in plant pararetroviruses has evolved in order to protect the viral coat protein-binding, secondary structure element located within the leader [26] – an RNA packaging signal – from being melted by linearly-migrating scanning ribosome [16]. A mechanism of packaging in RTSV is unknown: but conservation of the shunt mechanism between RTBV and RTBV raises a possibility that a packaging element may reside within the structured region of the RTSV leader.
The RTBV leader constructs “Wild type” and “KO start” have been described earlier [15]. The RTSV leader construct “Wild type” (Figure 1) is a derivative of the corresponding RTBV construct, in which the RTSV genomic RNA sequence from position +1 till position +535 (i.e. the leader sequence followed with a 21-nt segment of the polyprotein ORF) was inserted between the CaMV 35S promoter and the CAT reporter ORF in place of the RTBV leader (as a PCR-amplified RTSV fragment flanked with Cla I and Xho I and cloned into the corresponding sites of the vector). Note that this construct contains the natural polyprotein ORF start codon in a strong initiation context followed by 6 codons of this ORF and the CAT ORF fused to these 7 codons lacks its own ATG. In the RTBV constructs the CAT ORF begins with its own ATG in a strong context, which is in frame with the upstream AUU initiation codon located in the RTBV shunt landing site [15]. Point mutations of the RTSV sORF 1 start (ATG to taG) or stop (TAG to TAc) codons were introduced by PCR-based mutagenesis, yielding constructs ‘KO start’ and ‘KO stop’, respectively. The Kozak-stem (KS) sequence was introduced at the 5′ end of RTSV leader by cloning of a pre-annealed, self-complementary oligonucleotide CGGGGGCGCGTGGTGGCGGCTGCAGCCGCCACCACGCGCCCC (the self-complementary KS sequence is underlined) into the Cla I site of the RTSV plasmids “Wild type”, “KO start” and “KO stop”. The KS sequence shown above was also introduced in the middle of the RTSV leader sequence (in place of a guanosine at position 255) by a PCR ligation method, similar to that which we described previously [15]. Note that this insertion does not disrupt the RTSV secondary structure except that one of its branches is extended by KS (Figure S1C).
The RTSV leader constructs “Stem Disrupt L”, “Stem Disrupt R” and “Stem Restore L+R” were obtained using 353 bp and 200 bp synthetic DNA fragments of the RTSV wild type construct that contain sequences from Cla I to EcoRV and from EcoRV to Xho I, respectively, each with 12 point mutations shown in Figure 4. These fragments were introduced into the RTSV wild type construct individually or in combination by a two fragment ligation method. Same mutations were also introduced in the above-described construct carrying the KS sequence in the middle part of the RTSV leader: in this case 262 bp and 330 bp synthetic DNA fragments of the RTSV+KS construct were used, which contain sequences from Cla I to Pst I (located in the KS sequence) and from Pst I to Xho I, respectively, each with the 12 point mutations.
For the in vitro translation experiments, the T7 promoter was introduced just upstream of the RTSV full-length leader and their variants with the sORF1 mutations by subclonning the Cla I-Sph I fragment from the RTSV plasmids “Wild type”, “KO start” and “KO stop” in place of the corresponding fragment of the T7 promoter-RTBV leader-CAT ORF plasmid described previously [15].
Protoplasts from suspension culture of O. sativa were prepared and transfected with plasmid DNA by a polyethylene glycol method as described previously [8], [15]. Briefly, 0.6×106 protoplasts were transfected with 10 µg CAT-expressing plasmid and 2 µg β-glucuronidase (GUS)-expressing plasmid or 5 µg green fluorescent protein (GFP)-expressing plasmid. The GUS or GFP plasmid served as an internal control of transfection efficiency. Following incubation for 19–24 hrs at 27°C in the dark, protoplasts were harvested, protein extracts prepared and assayed for CAT and GUS (or GFP) accumulation, as described previously [20]. Relative GUS activities were taken for normalization of the CAT expression levels given in Figure 2 and Figure 3, while relative GFP activities were taken for normalization of the CAT expression levels given in Figure 4. For each construct, the values given are the means of at least three experiments in independent batches of protoplasts. Deviations from the mean values generally did not exceed 20%. The levels of CAT mRNA accumulation were measured by quantitative RT-PCR with CAT ORF-specific primers using previously-described protocols for total RNA preparation, cDNA synthesis and real time PCR [27] and found to be comparable for all the RTSV constructs (data not shown).
The in vitro experiments were performed as described in detail earlier [15]. Briefly, the T7-promoter plasmids were linearized by Sph I and transcribed in the presence of the cap analog 7mGpppG (in 6-fold molar excess over GTP) by incubation with T7 RNA polymerase (Biofinex). The integrity of the synthesized transcripts was evaluated on a 6% denaturing polyacrylamide gel. Equimolar amounts of capped transcripts (0.5 pmol) were translated for 1 hour at 27°C in a wheat germ extract. Accumulation of CAT protein in translation mixture was measured in duplicate by CAT ELISA (Roche) as recommended by the manufacturer. For each construct, in vitro translation was performed at least three times with freshly prepared capped RNA, yielding similar results.
Secondary structures at 25°C were predicted using the MFOLD program (Wisconsin Package, version 6.0; Genetics Computer Group, Madison, WI, USA). The most optimal and suboptimal secondary structures of the 515 nt RTBV leader sequence are shown in Figure S1. Folding of the RTSV leader sequence extended by either the natural RTSV coding sequence or the CAT reporter ORF sequence (present in the RTSV constructs tested here in the translational assays) did not affect the formation of the base section present in both optimal and suboptimal conformations (data not shown). Notably a free energy of the most optimal leader structure in RTSV (deltaG = −194.1 kcal/mol) is much more negative than that of fully randomized sequences of the same length (deltaG = ca. −100 kcal/mol; [28]).
MFOLD prediction of RNA secondary structure has proven to be reliable. For example, an MFOLD-predicted, large stem-loop structure of the 612-nt CaMV leader has been largely confirmed in vitro using chemical and enzymatic methods, though alternative conformations were also revealed in that study [29].
|
10.1371/journal.pbio.1002563 | Spermidine Suppresses Age-Associated Memory Impairment by Preventing Adverse Increase of Presynaptic Active Zone Size and Release | Memories are assumed to be formed by sets of synapses changing their structural or functional performance. The efficacy of forming new memories declines with advancing age, but the synaptic changes underlying age-induced memory impairment remain poorly understood. Recently, we found spermidine feeding to specifically suppress age-dependent impairments in forming olfactory memories, providing a mean to search for synaptic changes involved in age-dependent memory impairment. Here, we show that a specific synaptic compartment, the presynaptic active zone (AZ), increases the size of its ultrastructural elaboration and releases significantly more synaptic vesicles with advancing age. These age-induced AZ changes, however, were fully suppressed by spermidine feeding. A genetically enforced enlargement of AZ scaffolds (four gene-copies of BRP) impaired memory formation in young animals. Thus, in the Drosophila nervous system, aging AZs seem to steer towards the upper limit of their operational range, limiting synaptic plasticity and contributing to impairment of memory formation. Spermidine feeding suppresses age-dependent memory impairment by counteracting these age-dependent changes directly at the synapse.
| Neurons communicate by sending impulses, in the form of secretion of neurotransmitters, across small spaces called synapses. It is these synapses that undergo structural and functional changes during formation and retrieval of memories. Though alterations in synaptic performance are believed to accompany aging, the causal relationship between age-dependent memory impairment and synaptic changes remains largely unknown. Using the fly Drosophila melanogaster as a model, we found that feeding them spermidine—a polyamine compound—suppresses age-induced decline in olfactory memory, providing us with a tool to further decipher mechanisms associated with age-dependent memory impairment. In this study, we investigated the relationship between synaptic changes and age-dependent memory impairment by studying the olfactory circuitry. We observed an age-related increase in the levels of the synaptic proteins Bruchpilot and Rim-binding protein, which caused an enlargement of the presynaptic active zone—the complex of proteins that mediate neurotransmitter release—and enhanced synaptic transmission. Interestingly, feeding of spermidine was sufficient to abolish these age-associated presynaptic changes, further emphasizing the relationship between presynaptic performance and age-dependent memory impairment. Furthermore, flies engineered to express an excess of the core active zone protein Bruchpilot showed a premature impairment in memory formation in young flies. Based on our data, aging plausibly steers the synapses towards the upper limit of their operational range, limiting synaptic plasticity and contributing to impairment of memory formation.
| Age-dependent memory impairment (AMI), which is associated with both psychiatric and neurodegenerative disorders, starts in midlife and worsens with advancing age, suggesting that the greatest driving factor is age itself. The lack of effective treatments that prevent, halt, or reverse the condition is contributing to a diminishing quality of life for many senior citizens. Therefore, animal models that allow one to monitor physiological changes across their lifespan and to test for a causal character of age-induced changes might be helpful in exploring the mechanistic basis of AMI. D. melanogaster, with its short lifespan of around 60 d and advanced molecular genetic tools, provides an efficient experimental model to unravel mechanisms underlying AMI. Additionally, the olfactory nervous systems of insects and mammals exhibit many similarities, suggesting that the mechanisms for olfactory learning may be shared [1]. Moreover, aversive short-, intermediate-, and long-term olfactory memories have been found to be subject to age-induced decline in Drosophila, with an onset at about 10 d of age and plateaus at about 30 d of age [2–6]. Notably, we recently found a simple dietary supplementation of spermidine, a polyamine that specifically protects from AMI in Drosophila.
External stimuli are believed to be represented in the brain as spatiotemporal patterns of neural activity within a set of neuronal connections. Changes in synaptic communication (“plasticity”) within certain neuron populations are meant to ultimately encode behavioral adaptations such as learning and memory. Thus, dysfunctioning of synaptic plasticity might well be relevant to age-dependent deterioration of learning and memory [7,8]. One of the fundamental problems of studying AMI, however, is the inability to differentiate causative changes from adaptive or protective changes. Moreover, the brain undergoes changes at multiple levels with advancing age, including alterations in circuits, individual neurons, and single synapses, further complicating the situation [8]. Nonetheless, recent work has linked AMI to subtle synaptic alterations in the hippocampus and other cortical brain areas, rather than to the loss of neurons [7,9]. At the same time, the age-associated modulation of molecular entities underlying learning and memory that define and change synapse function remain poorly understood. Therefore, we set out to determine the role of age-induced changes in the organization and function of synapses in AMI, using dietary supplementation with spermidine as a tool to identify synaptic changes that can potentially contribute to AMI.
To accomplish this, we analyzed age-induced changes in the ultrastructural, molecular, and functional organization of synapses within the olfactory system of flies by comparing aged flies fed with normal food to aged flies fed with spermidine-supplemented food. We found that aging is associated with an increase in the average size of active zone (AZ) scaffolds, structures recently shown to scale with synaptic vesicle (SV) release. Consistent with this, optophysiological analysis showed that more SVs are released in response to natural odor stimuli in aged flies. Interestingly, these age-associated changes were suppressed by spermidine feeding, indicating that these changes might be causally relevant to AMI. In fact, genetic manipulation provoking an increase of T-bar size in young animals was sufficient to induce a premature decline in memory performance. We suggest that a cumulative increase in the size and function of presynaptic AZ scaffolds might reduce the operational range of synaptic plasticity processes, and thus, hamper the formation of new memories with age. Additionally, levels of postsynaptic neurotransmitter receptors and postsynaptic Ca2+ signals remained largely unaffected with age, suggesting that homeostatic adaptations might be involved in increasing the threshold for memory formation with advancing age.
It is known that the ability to acquire new memories declines with advancing age. Based on previous study [10], one plausible explanation for this observation might be the increase in the threshold required for memory formation with age. In fact, when we analyzed olfactory conditioning in aged flies (30-d-old flies or 30d) we found that greater number of exposures to the unconditioned stimulus in order to attain saturated levels of memory scores, which, however, never reached the same maximal learning scores found in young flies (3-d-old flies or 3d), indicating that the dynamic range of memory formation is altered with advancing age (S1 Fig). Multiple lines of evidence suggest that presynaptic plasticity processes are responsible for forming olfactory associative memory in Drosophila [11–13]. Therefore, we set out to determine the role of age-induced changes in the organization and function of synapses in AMI, using dietary supplementation with spermidine as a tool to identify synaptic changes that can potentially contribute to AMI.
In order to identify synaptic mechanisms plausibly contributing to AMI, we used opto-physiological assays to characterize overall neuronal responses in synaptic terminals of live intact flies. For these experiments, we focused on projection neuron (PN) to kenyon cell (KC) synapses within the mushroom body calyx of the olfactory system for two reasons: first, aversive olfactory learning involves coincidence detection of a conditioned stimulus (odor) with an unconditioned stimulus (electric shock), causing changes in the odor-specific synaptic activity of second order PNs and third order mushroom body KCs [1,14]; second, the superficial position of the calyx within the fly brain enabled us to perform efficient optical analysis [15], since sensor signals could be retrieved from discrete synaptic bouton areas.
We started by expressing cytosolic GCamp3.0 in the PNs (using GH146-Gal4) and found the basal expression of GCamp3.0 to remain largely unchanged with age (S2A Fig). Next, we monitored the PN boutons for intracellular Ca2+ responses to two odors typically used for olfactory conditioning, 3-Octanol (3-Oct) and 4-methylcyclehexanol (MCH), through two-photon microscopy. Similar to our previous observations [3], we found no significant difference in the amplitude or time course of cytosolic GCamp3.0 signals of young (3d) and aged (30d) animals (S3 Fig). Thus, in the context of odor information processing, odor-evoked action potential frequency or presynaptic Ca2+ influx remained rather unaffected by the age of the animal.
Next, we asked whether the release of SVs was altered with advancing age and analyzed the odor-driven SVs release. To this end, we used SynaptopHluorin (SynpH), a pH-sensitive green fluorescent protein (GFP) fused to the luminal side of the SV membrane protein Synaptobrevin (Syb) [16]. SynpH is nonfluorescent at the acidic pH inside SVs; however, when SVs are released, SynpH is exposed to the neutral extracellular space, and the presynaptic terminal becomes brightly fluorescent. Following endocytosis, SVs become reacidified, and the cycle can start again [17]. SynpH was expressed within PNs, and the release of SVs in response to two odors was monitored, again, at PN-to-KC synapses (Fig 1A–1H). We found a profound increase in the amplitude of SynpH signals in aged (30d) animals when compared to young (3d) flies (Fig 1A–1H). In contrast, spermidine administration to 30d flies prevented this age-dependent increase of odor-induced SynpH signals (30dSpd; Fig 1A–1H). Alterations in the endocytotic clearance of newly released SVs might, per se, explain the increase in SynpH signals observed; however, the decay constants of the poststimulus SynpH signal remained essentially unchanged with aging (S4 Fig), indicating that the endocytic clearance cannot be responsible for the difference in odor-driven SynpH signals observed in aged animals. In addition, neither the basal expression of SynpH before odor stimulation nor the maximal SynpH signal determined by high-molar KCl treatment showed systematic differences between young and aged cohorts (S2B and S5 Figs). These experiments, thus, indicate that the exocytosis of SVs underlies the increase in SynpH response with advancing age.
In addition to measuring the SV release at the PN presynaptic terminals within calyx, we also measured odor-evoked changes within the axonal projections of KCs within the mushroom body horizontal lobes by expressing SynpH using mb247-Gal4. Though relative signals were smaller (when signal was normalized to the whole mushroom body horizontal lobe), likely reflecting the well-documented sparse odor coding of KCs [18], we still observed a substantially higher amplitude of SynpH signals in aged (30d) than in young (3d) flies, and, again, spermidine administration (30dSpd) protected from this age-dependent increase (S6 Fig). Thus, two major neuron populations of the olfactory system—PNs and KCs, showed an increase in odor-evoked fluorescence changes in response to odor stimuli, indicating higher release of SVs in aged animals.
Since Ca2+ influx into presynaptic terminals was apparently not responsible for the profound age-induced increase in SV release, presynaptic mechanisms downstream of Ca2+ signaling might be involved. In order to address the molecular and cellular basis of this age-associated increase in SV release, we started by analyzing proteins directly associated with SVs: Synapsin, Syb, and Synaptotagmin-1. Synapsin is a SV-associated phosphoprotein important for controlling the number of SVs available for release [19], and Syb is a core component of SNARE complex that drives the exocytosis of SVs [20,21]. We found the levels of Synapsin as well as Syb to remain unchanged with advancing age (comparing aged flies: 30-days old or 30d with young flies: 3d), regardless of spermidine feeding (30dSpd; Fig 2A–2D and S7 Fig). Synaptotagmin-1 is a vesicular protein with a central role as a Ca2+ sensor for SNARE-dependent SV fusion [22]. Synaptotagmin-1 decreased slightly with age, and feeding with spermidine had no discernable influence on this age-dependent change (Fig 2E–2H), indicating that these moderate changes are seemingly not associated with AMI.
The release of neurotransmitters is a sophisticated process that requires SVs to be in close vicinity to voltage-gated Ca2+ channels, and this precise spacing is orchestrated by interplay among several proteins that form the AZ scaffold [23,24]. In flies, the ELKS-family protein Bruchpilot (BRP) is an essential building block of the AZ scaffold and is needed to effectively cluster Ca2+ channels as well as regulate the release of SVs [25–28]. When whole-mount brains were stained for BRP using two different antibodies (BRPNc82 and BRPN-term), we observed a substantial increase in the levels of BRP with advancing age (Fig 2I, 2J, 2L and 2M). Similarly, Rim-binding protein (RBP) [29], another structurally and functionally important component of the AZ scaffold, was found to be significantly increased in brains of 30d flies compared to 3d animals (Fig 2I, 2J and 2N). Furthermore, flies analyzed at shorter intervals throughout their lifetime exhibited a progressive increase in the levels of both BRP and RBP (S8 Fig). Notably, the age-dependent increase in BRP and RBP signals was suppressed in aged flies fed with spermidine (30dSpd; Fig 2I–2N).
The staining efficacy could potentially be influenced by the sheer age of the tissue, e.g., due to differences in antibody penetration. To rule this out, flies expressing a GFP-tagged genomic BRP construct (rescuing the lethal brp null mutant [28]) were aged on normal food or food supplemented with spermidine. We found the endogenous GFP signals to be significantly increased in 30d flies in comparison to 3d flies, while feeding with spermidine again prevented this age-related increase (S9 Fig).
Since the AZ scaffold has previously been reported to effectively cluster Ca2+ channels [26–28], we asked whether the age-associated increase in levels of core AZ-proteins might influence synaptic levels of Ca2+ channels. To address this, we expressed a GFP-labeled genomic construct of α1 subunit Cacophony (Cac), which is the only representative of the mammalian Cav2.1/2.2 family present in Drosophila [28], and stained the flies for GFP and BRP. We found the levels of Cac (quantified using an antibody against GFP) to remain unchanged with aging (S10 Fig). Besides its role in Ca2+ channel clustering, the AZ scaffold has been suggested to create a stereotypic arrangement that defines SV release slots by clustering SV release machinery [28]. In fact, the levels of Unc13, a protein essential for priming SVs by rendering them fusion-competent [24], were also increased in brains of 30d flies compared to 3d flies (S11 Fig). Again, spermidine administration suppressed this age-dependent increase (30dSpd; S11 Fig). Taken together, our data suggest that synaptic levels of core AZ scaffold proteins, BRP and RBP, as well as the levels of critical release factor Unc-13 increased with advancing age.
Next, we asked whether the increase of both BRP and RBP labeling in aged brains reflects an increase in the number of AZs or just the increase in local amounts of these proteins at individual AZs. To resolve this, we performed ultrastructural analysis on PN-to-KC synapses within the mushroom body calyx. In contrast to presynaptic terminals of KCs, presynaptic PN terminals within the calyx exhibit a well-defined morphology [30,31], by which synapse types can be reliably identified in EM micrographs. Moreover, the superficiality of the calyx enabled us to perform stimulation emission depletion microscopy (STED) analysis (see below).
In order to allow for an unbiased quantification, we applied automated data collection to acquire more than a thousand transmission electron microscopic images covering nearly a whole calyx cross-section, which were then “stitched” together into a single high-magnification image (see Materials and Methods). As described previously [30], PN boutons could be easily identified, and light-colored boutons containing clear-core SVs were used for analysis. We recognized that plasma membranes between cellular elements were less aligned, with an increase in extracellular spacing between cellular elements, in aged (30d) flies when compared to young (3d) flies (S12 Fig). Spermidine feeding appeared to substantially alleviate this age-related change (S12 Fig). Driven by the finding that SV release is increased with age, we decided to analyze the AZs within PN boutons. We found aged animals (30d) to display reduced numbers of AZs per unit bouton-area in comparison to 3d flies, with no apparent influence of spermidine feeding on this age-dependent decline (Fig 3A–3E). The density of SVs in proximity to the AZ scaffold appeared unchanged in aged flies (30d as well as 30dSpd), when compared to young flies (3d; Fig 3F). Additionally, the number of SVs docked at the AZ plasma membrane appeared essentially unaltered with advancing age (Fig 3G).
The AZ scaffold exhibits an electron-dense structure in electron microscopy (EM), and due to its T-shaped structure in Drosophila, this scaffold is often referred to as a T-bar [24,26,27]. We found the average size of the T-bars to be significantly increased in 30d animals in comparison to 3d flies (Fig 4A–4D). Feeding flies with spermidine suppressed this age-induced increase in T-bar size (30dSpd; Fig 4A–4D). We have previously introduced STED in the analysis of AZ suborganization [26–28]. At peripheral neuromuscular synapses of Drosophila larvae, STED allowed us to unmask the “nano-architecture” of AZs where BRP and RBP organize a scaffold that provides slots for SV release and concentrates Ca2+ channels in the AZ center [28,29]. When planar AZs were imaged using the BRP C-terminal epitopes at neuromuscular synapses, they display a ring-shaped structure whose diameter correlated with the EM-derived physical size of individual T-bar/AZ scaffold [32]. We applied STED to PN-to-KC synapses of the calyx and found ring-like BRP structures at planar-oriented AZs (S13 Fig). Subsequently, the analysis of these STED images revealed an increase in the ring diameter of BRP spots with advancing age, while spermidine treatment was able to suppress this age-associated increase (Fig 4E–4H). Finally, we performed coimmuno-EM labeling against BRP and RBP on calycal slices. The number of gold particles positive for BRP as well as RBP was found to increase in aged flies (30d) in comparison to both young (3d) flies and aged flies fed with spermidine (30dSpd; Fig 4I–4N). Taken together, the morphological EM, immuno-EM, and STED analysis consistently show that aged animals display larger AZ scaffolds, plausibly due to an increase in local amounts of the critical scaffold components: BRP and RBP.
Recent in vivo analysis of larval Drosophila neuromuscular junctions has shown that the local amounts of BRP at a given AZ scale directly with the probability of evoked SV release [33–37]. Consistent with these studies, we found SV release to increase and AZ scaffolds to enlarge with age, while importantly both these age-related changes were suppressed by dietary supplementation with spermidine. Therefore, we next wanted to determine the influence of these synaptic changes on olfactory memory formation.
Presynaptic plasticity processes have been reported to be critical for forming olfactory associative memory in Drosophila [11–13]. Based on our findings, we suggest that the scale-up in the size and function of AZ scaffolds is likely to change the “operational range” of synaptic plasticity processes and thus change the threshold for memory formation. Thus, we wanted to test whether genetically provoking an artificial enlargement of AZ scaffolds, independent of the aging process, might affect memory formation. Since BRP is a major essential building block of the AZ scaffold in Drosophila [26–28,32], we decided to increase the gene copy number of BRP from two to four copies by combining two additional genomic copies of brp [28] with two endogenous copies. As a result, BRP signals increased substantially in 3d flies expressing four-copy BRP (4xBRP) when compared to 3d flies expressing two-copy BRP (2xBRP; Fig 5A–5E). Additionally, RBP levels also increased concomitantly with BRP (Fig 5A–5D and 5F), consistent with the suggested role of BRP to operate as a “master molecule” in shaping the size (and functional performance) of the AZ scaffold [28,29,36]. In order to confirm that the increase in BRP levels resulted in an increase of the average size of AZ scaffolds, we took advantage of STED imaging. Again, a considerable increase in the ring diameter of BRP spots was observed in 2xBRP flies with advancing age (Fig 5G–5K). Meanwhile, we found young flies (3d) expressing 4xBRP to have increased BRP ring diameters when compared to age-matched control flies (2xBRP), and the ring diameter of BRP spots in 4xBRP flies remained rather unchanged with age (Fig 5G–5K).
Having created a genetic state wherein levels of AZ core scaffold proteins increased prematurely in young animals, we decided to investigate the influence of this manipulation on memory formation. Before doing so, however, we wanted to ascertain whether the innate behavior was affected in 4xBRP flies. Thus, we measured naïve odor response and shock reactivity and found 4xBRP flies to show odor avoidance and shock reactivity scores that were indistinguishable from 2xBRP age-matched control flies (2xBRP; S1 Table). Subsequently, we started by measuring short-term memory (STM), and found 4xBRP flies to exhibit lower memory scores “already” at a young age (3d), and their memory scores declined only negligibly with age (Fig 5L). In contrast, control flies (2xBRP) exhibited normal AMI (Fig 5L).
As mentioned earlier, intermediate-term memory (ITM) has also been reported to decline with age [2–4]. Consistently, we found that 30d 2xBRP flies show substantially reduced ITM scores (measured 3-h post-training) when compared to 3-d 2xBRP flies (Fig 5M). By contrast, the 4xBRP flies showed lower ITM scores at a young age (3 d) and, again, the ITM scores did not decrease further in 30-d 4xBRP flies (Fig 5M). In fact, the learning performance of 3-d 4xBRP flies was comparable to that of 30-d 2xBRP flies. Based on distinct genetic mutants and specific pharmacological sensitivities [2,4,38,39], the ITM can be dissected into anesthesia-sensitive memory (ASM) and anesthesia-resistant memory (ARM) components. The ASM, unlike the ARM, has been shown to be strongly impaired with aging [3,4]. The ASM can be calculated by subtracting ARM scores, measured after amnestic cooling, from ITM. Consistent with previous studies [2–4,40], we found ARM in 2xBRP and 4xBRP flies to remain relatively unaffected with age (Fig 5M). In contrast, ASM was nearly absent in 30-d 2xBRP flies when compared to 3-d 2xBRP flies. Reaffirming our idea, we found the young (3-d) 4xBRP flies to show lower ASM scores in comparison to age-matched control (2xBRP) flies, while their ASM scores declined negligibly with age (Fig 5M). These experiments indicate that a genetically provoked “up-scaling” of the average AZ scaffold size is sufficient to induce an “early” decline in memory, similar to AMI, which physiologically occurs over a time course of 20–30 d.
A reduction in BRP levels, per se, might be expected to slow down the onset of AMI. To address this possibility, we removed a single gene copy of brp, and found BRP heterozygotes (brp69/+ or 1xBRP) to exhibit a considerable reduction in the levels of both BRP and RBP (S14A–S14F Fig), indicating that our antibody stainings can detect subtle changes and reaffirming that BRP levels can directly steer the local amounts of other AZ components in the Drosophila brain. We found that 3d flies expressing only one BRP copy (brp69/+) displayed memory scores comparable to 3d control flies (2xBRP); however, these brp69/+ flies still exhibited a normally-occurring AMI (30d; S14G Fig). AZ scaffold-dependent control of neuronal plasticity is undoubtedly a complex process [24,41], and other mechanisms, operating in parallel to modulations in the amounts of scaffold proteins, might well contribute to the pace and extent of AMI. Lysine-acetylation of BRP was recently identified as a major node to control the SV release at larval AZs [42,43]. In particular, the loss of histone deacetylase-6 (HDAC6) was found to cause hyperacetylation of BRP and provoke a reduction in evoked SV release at AZs [43]. Interestingly, using immunoprecipitation followed by mass spectroscopic analysis, we found at least 13 lysine sites within BRP to be target for (de)acetylation, (S15 Fig). Next, we asked whether loss of HDAC6 might affect memory. While the odor avoidance and shock reactivity were mainly unaffected by knockdown of hdac6 (S1 Table), memory scores of both young and aged flies with pan-neuronal knockdown of hdac6 were higher than those of age-matched driver controls (Fig 5N). These findings are consistent with the idea that driving down the AZs towards the lower limit of their operational range might facilitate memory formation in aged animals. Though any implications of acetylation of BRP or potentially other AZ scaffold proteins with respect to aging process still require extensive analysis, this result shows that BRP-directed modifications, reported to reduce SV release, can in fact increase the efficacy of memory formation in aged animals.
Finally, we asked how the postsynaptic compartment might respond to these age-associated presynaptic structural and functional changes. To address this question, we used GCaMP3.0 fused to the postsynaptic protein Homer [15] and found the basal expression of Homer-GCamp3.0 to be largely unaffected with age (S2C Fig). Moreover, the sensor was found to be effectively targeted to the postsynaptic density of the PN::KC synapses, as manifested by its specific enrichment within the postsynaptic specializations formed by claw-like dendritic endings of multiple KCs surrounding a single PN bouton (Fig 6A). However, postsynaptic Ca2+ signals did not increase with age. Rather, a slight tendency towards a decrease of postsynaptic Ca2+ signals was observed in normally aged animals when compared to young controls (Fig 6A–6H). At the same time, aged flies treated with spermidine (30dSpd) produced signals more similar to untreated 3d-Homer-GCaMP3.0 flies than to untreated aged animals (Fig 6A–6H). In order to be certain that Homer-GCamp3.0 signals were not saturated, we used high-molar KCl treatment to determine the maximal postsynaptic Ca2+ response. Unlike the odor-evoked maximum change in Homer-GCamp3.0 fluorescence of about 55%, KCl stimulation resulted in a substantially higher ΔF/F0 value of more than 300% (S16 Fig), suggesting that sensor sensitivity was not a limiting factor for the postsynaptic Ca2+ signals. Meanwhile, when the cumulative postsynaptic Ca2+ activity was critically analyzed during the odor stimulation, we found that the Ca2+ responses reduced significantly in aged (30d) flies relative to young flies, while the Ca2+ signals were comparable between young flies and spermidine-fed aged animals (30dSpd; S17 Fig).
PNs provide cholinergic input to the KCs within the calyx [44]. We used a fusion of mushroom body-specific enhancer mb247 to the Dα7 subunit of the acetylcholine receptor (mb247::Dα7GFP) to explicitly visualize postsynaptic acetylcholine receptors. We showed previously that expression of Dα7-GFP from KCs localized specifically to the KC postsynaptic densities, where it closely matched the AZs of the PNs [45]. While we observed an age-related increase in BRP in 30d mb247::Dα7GFP flies in comparison to 3d mb247::Dα7GFP flies, the levels of Dα7 subunit (quantified using an antibody against GFP fused to the α7 subunit of acetylcholine receptors) did not change with age, and spermidine feeding had no effect on the level of the α7 subunit of acetylcholine receptors (Fig 7A–7E). Similarly, when we stained for endogenous Drep2, a postsynaptic scaffold protein that is known to express strongly within the postsynaptic densities of PN::KC synapses [46], we also found Drep2 to remain unchanged with age (S18 Fig).
At first glance, the increase in release of SVs might be expected to translate into increased postsynaptic responses; however, ample evidence from various studies in different model organisms, including Drosophila, support the existence of homeostatic controls, allowing neurons to remain within a certain range of excitation [47,48] in order to avoid epileptic states and Ca2+-induced degeneration. In an attempt to directly examine the existence of such homeostatic controls, we wanted to determine whether an increase in the amount of depolarization required to trigger an action potential might influence the architecture of the apposed AZ scaffold. To achieve this, we used dORK1ΔC, a constitutively open K+ selective pore that causes hyperpolarization of neurons and subsequent inactivation of neuronal function [45,49]. dORK1ΔC was specifically expressed in the KCs, and presynaptic terminals of PNs within the calyx were analyzed for BRP levels (Fig 7F–7I). Indeed, we found a substantial increase in the levels of BRP in the calyces of both 3d as well as 10d mb247>dORK1ΔC flies, when compared to age-matched controls (Fig 7J). Thus, a drop in postsynaptic excitability can drive a homeostatic increase in presynaptic AZ scaffolds, leading to a potential increase in SV release at olfactory synapses—a finding similar to the one we found at aging synapses. Though the exact mechanisms allowing for homeostatic compensation of the elevated presynaptic release remain to be further worked out, it is tempting to speculate that homeostatic mechanisms coupling postsynaptic excitability to presynaptic release function might drive aging synapses towards the upper limit of their operational range and be critically involved in AMI (see model in Fig 7K)
The aging process, causing progressive deterioration of an organism, is subject to a complex interplay of regulatory mechanisms. One of the primary aims of aging research is to use the understanding of this process to delay or prevent age-related pathologies, including AMI. We previously showed that restoration of polyamine levels by dietary supplementation with spermidine suppressed AMI in fruit flies [3], providing us with a protective paradigm to identify candidate processes that might be functionally associated with AMI. As an insight towards the synaptic basis of AMI, we describe an age-induced increase in the levels of core AZ proteins, BRP, and RBP and of the functionally critical release factor Unc13, together with a shift towards an enlargement of AZ scaffolds within the olfactory system. In addition, based on SynpH experiments, we observed a substantial increase in the release of SVs at aged synapses (PN-to-KC and KC-to-mushroom body output neuron [MBON] synapses) in response to odors used for learning experiments. Importantly, spermidine feeding was able to “protect” from both the functional and structural changes at aged AZs, arguing in favor of specific synaptic changes to be causally relevant for AMI. Indeed, installing 4xBRP not only increased the size of BRP rings in young flies, similar to those found in aged animals, but also provoked memory impairment in young flies.
Notably, a reduction of BRP levels has previously been reported to affect ARM but not ASM [50]. Here, we report that an increase in BRP levels (by changing the gene copy number of BRP from two to four copies) severely affected ASM. These findings suggest that the two complementary forms of memory (ARM and ASM) might rely on the recruitment of distinct presynaptic “functional modules.” The loss of brp has been shown to severely reduce release function in response to single low frequency, but not in response to high-frequency stimulation [27], indicating that SV release at low-frequency stimulation might be particularly relevant for forming ARM, a memory component that develops gradually after training. On the other hand, mobilization of the SVs during high frequency stimulation has been suggested to be critical for formation of ASM [50], a memory component that predominates early memory and decays with age. Thus, the increase in the size of the AZ scaffolds might potentially contribute directly to AMI by interfering with mechanisms facilitating SV availability in the course of forming ASM.
Though the exact mechanisms underlying age-induced synaptic changes remain to be fully worked out, a reduction in autophagy-mediated protein degradation might well be involved [51–53]. Autophagy is a cellular digestion pathway that involves the sequestration of cytoplasmic components within a double-membrane vesicle called autophagosome, which fuses with lysosomes (autolysosomes) to degrade autophagic cargo by acidic hydrolases [52]. Interestingly, spermidine was shown to induce autophagy in several model systems, including rodent tissues and cultured human cells [51,54,55]. Moreover, amelioration of a-synuclein neurotoxicity due to spermidine administration was accompanied by autophagy induction [56]. Of note, we also found that spermidine feeding prevented accumulation of poly-ubiquitinated proteins by plausibly halting normally occurring age-induced decline of autophagic clearance [3,57]. The gene atg7 encodes an E1-like enzyme required for activation of both Atg8 and Atg12, a step critical for the completion of the autophagic pathway [53]. We found that atg7-mutant flies (atg7-/-) exhibit reduced memory scores at a young age (3d), which declined further with age (20-d of age or 20d) [3]. Concurrently, spermidine-mediated protection from memory impairment was eliminated in atg7−/− flies (for both 3d- and 20d-flies) [3,57]. Therefore, we wondered whether the decrease in the autophagic pathway might, per se, provoke increase of AZ scaffold components. When staining for BRP in atg7-mutant brains (atg7-/-), we found a brain-wide increase in levels of BRP (for both BRPNc82 and BRPN-term antibodies), and spermidine feeding was unable to prevent this age-related increase (S19 Fig). The finding that spermidine feeding in atg7−/− flies neither blocked the increase in BRP levels (S19 Fig) nor suppressed memory impairment [3] suggests that the integrity of the autophagic system is crucial for the spermidine-mediated protection from age-associated increase in AZ scaffold components. Spermidine effects were recently shown to involve widespread changes of both nuclear and cytosolic protein acetylation [58,59].
In primary neurons, autophagosomes have previously been observed to form at the distal end of the axon, indicating compartmentalization and spatial regulation of autophagosome biogenesis [60,61]. More recently, autophagosomes were demonstrated to form directly near synapses and were found to be required for presynaptic assembly at developing synaptic terminals of Caenorhabditis elegans [62]. Moreover, the crucial release factor Unc13 was found to accumulate under conditions of defective endosomal microautophagy (a specialized form of autophagy) at developing neuromuscular synapses of Drosophila, suggesting Unc13 to be a substrate of this form of autophagy [63]. Interestingly, we have shown recently that the synaptic levels of Unc13-A isoform scale tightly with the levels of the BRP/RBP scaffold [64]. Thus, it is conceivable that some of the AZ proteins, whose levels increase with age (BRP/RBP/Unc13), might be direct substrates of “pre-synaptic autophagy,” and that spermidine feeding might augment effective autophagic degradation of these proteins at aging synapses.
We also observed a moderate decrease in synapse numbers in aging Drosophila brains, a phenotype that was unaffected by spermidine feeding. Our data compare favorably with studies in mammals. For example, loss of synapses in aged rodents has been reported in the dentate gyrus as well as the CA1 area of the brain [8,65,66]. Additionally, the “unitary” intracellular-evoked amplitude elicited by minimal stimulation protocols has been found to be greater in old than in young rodents [67], suggesting that the “surviving synapses” are stronger [68]. It is of note that the induction threshold for long-term potentiation, considered to be a synaptic correlate of learning, has been reported to increase in aged rodents [10]. Similarly, an age-related increase in the amplitude of endplate potentials evoked has been reported at mouse neuromuscular synapses [69,70]. By contrast, a study at neuromuscular junctions of C. elegans revealed that aged motor neurons undergo a progressive reduction in synaptic transmission [71]. In flies, however, an age-related increase in the amplitude of the excitatory postsynaptic potential at adult neuromuscular junctions has been reported recently; this increase was suggested to tune the response of the homeostatic signaling system and establish a new homeostatic set point [72]. Collectively, these findings suggest that the dynamic range of synaptic plasticity may change with advancing age and, thus, contribute to AMI.
Why would an increase in the odor-evoked SV release and ultrastructural size of AZ scaffolds impair the efficacy of forming new memories? Synapses appear to display a “finite ceiling and floor” that define a synaptic operating range [73]. In rodents, the formation of new memories seems to drive synaptic strength to the upper limit of a fixed operating range, thereby creating an imbalance [73]. As a result, if the synapses are not returned to the midpoint of the synaptic modification range, then additional strengthening required for new memory formation might be blocked, and the system is driven to employ homeostatic compensatory mechanisms to balance the change [74]. In our experiments, we found dendritic Ca2+ signals and postsynaptic receptor levels to remain largely unchanged with age, suggesting the existence of homeostatic mechanisms that might allow the up-scaling of presynaptic release to be compensated by lowering the postsynaptic response to a given amount of neurotransmitter released. On the other hand, this upscaling of presynaptic structure and function might also be a homeostatic response to a reduction in postsynaptic excitability or Ca2+ homeostasis, steering retrograde enlargement of AZ scaffold and higher release of SVs (Fig 7K). In fact, the influx of postsynaptic Ca2+ through glutamate receptors at the peripheral glutamatergic synapses of Drosophila has been reported to control presynaptic assembly by retrograde signalling [47,48,75]. While the exact nature of homeostatic controls connecting pre- with postsynaptic neurons in the olfactory system remains to be resolved, changes in plasma membrane excitability, a change in postsynaptic neurotransmitter sensitivity, or an increase in inhibitory GABAergic drive are obvious candidate processes. Taken together, we propose these synaptic changes steer the presynaptic AZs to function towards the upper limit of their operational range, making these synapses unable to react adequately to conditioning stimuli and provoke potentiation or depression of synapses in order to encode memory formation [11,12,76].
Sleep is widely believed to be critical for formation and consolidation of memories [77]. In sleep-deprived animals, neuronal circuits would exceed available space and/or saturate, thereby affecting an individual’s ability to learn [77]. Importantly, sleep deprivation has also been associated with widespread increases of BRP levels in the Drosophila brain [78]. Notably, we also observed a brain-wide increase in BRP levels in aged brains. It is tempting to speculate that both sleep deprivation and aging change the operational range over several synaptic relays and thereby affect memory formation—a topic that deserves further investigation in future.
Taken together, our data show that upscaling of presynaptic structure and function contribute to an AMI in Drosophila. Furthermore, and restoration of polyamine levels prevents these age-associated alterations as well as AMI. Thus, spermidine feeding provides a unique opportunity to further the molecular and functional dissection of the mechanisms underlying AMI with the ultimate goal of restoring memory function in older humans.
All fly strains were reared under standard laboratory conditions [79] at 25°C and ≈70% humidity, with constant 12:12 h light/dark cycle. Flies from an isogenized w1118 strain were used as the wild-type control for all experiments. Flies carrying P(acman) cacGFP, P(acman) brp83GFP and P(acman) brp83 [28] and mb247::Dα7GFP [45] were described previously. The generation of UAS-homer-GCaMP3.0 flies are described elsewhere [15]. Briefly, cDNA of dhomer was amplified from w 1118 flies and inserted with a C-terminal linked GCaMP336 into pUAST. Both UAS-GCamp3.0 (on the 3rd chromosome) [80] and UAS-SynpH [81] were kindly provided by Gero Miesenböck. Atg7d14 and Atg7d77 flies were kind gifts from Thomas Neufeld [53]. In addition, mb247-Gal4 [82] and gh146-Gal4 [83] were used.
As previously described [3], the fly food was prepared according to Bloomington media recipe (www.flystocks.bio.indiana.edu/Fly_Work/media-recipes/media-recipes.htm) with minor modification, which was called Spd−or normal food. Spermidine (Sigma Aldrich) was prepared as a 2 M stock solution in sterile distilled water, aliquoted in single-use portions and stored at −20°C. After food had cooled down to 40°C, Spermidine was added to normal food to a final concentration of 1 mM or 5 mM Spd, and called Spd1mM+ or Spd5mM+, respectively. Parental flies mated on either Spd−or Spd5mM+ food for all experiments, and their progeny were allowed to develop on the respective food. Flies used in all experiments were F1 progeny. The flies were collected once a day for aging, as a results-specific age indicated is day ± 24 h.
Behavioral experiments were performed in dim red light at 25°C and 80% relative humidity with 3-Oct (1:150 dilution in mineral oil presented in a 14 mm cup) and MCH (1:100 dilution in mineral oil presented in a 14 mm cup) serving as olfactory cues, and 120V AC current serving as a behavioral reinforcer. Standard single-cycle olfactory associative memory was performed as previously described [3,4,46,84,85], with minor modifications. Briefly, about 60–80 flies received one training session, during which they were exposed sequentially to one odor (conditioned stimulus, CS+; 3-Oct or MCH) paired with electric shock (unconditioned stimulus, US) and then to a second odor (CS−; MCH or 3-Oct) without US for 60 s with 30 s rest interval between each odor presentation. During testing, the flies were exposed simultaneously to the CS+ and CS− in a T-maze for 30 s.
The conditioned odor avoidance was tested immediately after training for STM (memory tested immediately after odor conditioning). Subsequently, flies were trapped in either T-maze arm, anesthetized, and counted. From this distribution, a performance index was calculated as the number of flies avoiding the shocked odor minus the number avoiding the nonshocked odor divided by the total number of flies and, finally, timed by 100. A 50:50 distribution (no learning) yielded a PI of zero, and a 0:100 distribution away from the CS+ yielded a PI of 100. A final performance index was calculated by the average of both reciprocal indices for the two odors.
For ITM, flies were trained as described above, but tested 3 h after training. As a component of ITM, ARM was separated from ASM by cold-amnestic treatment, during which the trained flies were anesthetized 90 s on ice at 30 min before testing. In the end, ASM was calculated by subtracting the performance index of ARM from that of ITM for each training session on the same day, respectively.
Brains were dissected in HL3 solution and fixed for 20 min at room temperature with 4% paraformaldehyde and 0.2% Glutaraldehyde in a buffer containing 50 mM Sodium Cacodylate and 50 mM NaCl at pH 7.5. Afterwards, brains were washed twice in the buffer and dehydrated through a series of increasing alcohol concentrations. Samples were embedded in London-Resin (LR)-Gold resin by incubating them in Ethanol/LR-Gold 1:1 solution overnight at 4°C, followed by Ethanol/LR-Gold 1:5 solution for 4 h at room temperature. Thereafter, the samples were washed first with LR-Gold/0.2% Benzil overnight, a second time for 4 h, and then again overnight. Finally, the brains were placed in BEEM capsules covered with LR-Gold/0.2% Benzil resin and placed under a UV lamp at 4°C for 5 d to allow for polymerization of the resin.
Following embedding, sections 70–80 nm, each, were cut using a Leica Ultracut E ultramicrotome equipped with a 2 mm diamond knife. Sections were collected on 100 mesh nickel grids (Plano GmbH, Germany) coated with 0.1% Pioloform resin and transferred to a buffer solution (20 mM Tris-HCl, 0.9% NaCl, pH 8.0). Prior to staining, sections were blocked for 10 min with 0.04% BSA in buffer. Sections were incubated with the primary antibody (guinea pig-anti RBPSH3II+III and rabbit-anti BRPlast200, 1:500 dilution) in blocking solution overnight at 4°C. After washing four times in buffer, the sections were incubated in buffer containing the secondary antibody (goat anti-guinea pig 10 nm colloidal gold, goat anti-rabbit 5 nm colloidal gold British Biocell, 1:100) for 2–3 h at room temperature. Finally, the sections were washed four times in buffer and three times in distilled water. Contrast was enhanced by placing the grids in 2% uranyl acetate for 30 min, followed by washing three times with water and, afterwards, incubation in lead citrate for 2 min. Afterwards, the grids were washed three times with water and dried. Images were acquired on a FEI Tecnai Spirit, 120 kV transmission electron microscope equipped with a FEI 2K Eagle CCD camera.
The following primary antibodies were used: MαBRPNc82 (ref. 9, 10; 1:100), GPαBRPN-term (1:800) [25,27], RbαRBPC-term (1:800) [29], MαSynapsin (1:20) [89], RatαSyb (1:100) [90], RbαSynaptotagmin-1C-term (1:500) [91], RbαGFP (Molecular Probes; 1:500), RbαDrep2C-term (1:500) [46], RbαUnc13C-term (1:500) [63], RbαBRPlast200 (1:500), and GPαRBPSH3II+III (1:500).
The following secondary antibodies were used: GαM Alexa 488 (Molecular Probes; 1:400), GαR Alexa 488 (Molecular Probes; 1:500), GαGP Alexa 555 (Invitrogen; 1:800), GαM Cy3 (Dianova; 1:500), and GαR Cy5 (Invitrogen; 1:400).
For Immunoprecipitation, BRPlast200 and IgG were used at final amount of 50 ug per 500 ul. For western blots, secondary antibody was used at a dilution 1:1,000.
Female 3d or 30d flies were briefly anesthetized on ice and immobilized in a small chamber under thin sticky tape. A small window was cut through the sticky tape and the cuticle of the head capsule using a splint of a razor blade. Trachea were carefully removed and the brain was covered with Ringer’s solution (5 mM HEPES, 130 mM NaCl, 5 mM KCl, 2 mM MgCl2, 2 mM CaCl2, pH = 7.3). Imaging was performed using an LSM 7 MP two-photon microscope (Carl Zeiss) equipped with a mode-locked Ti-sapphire Chameleon Vision II laser (Coherent), a 500–550 nm bandpass filter, and a Plan-Apochromat 20×1.0 NA water-immersion objective (Carl Zeiss). A custom-built device to supply odorous air with a constant flow rate of 1 ml/s directly to the fly’s antennae was attached to the microscope. Odor stimulation (MCH or 3-Oct, diluted 1:100 or 1:150, respectively, in mineral oil or pure mineral oil) was controlled using a custom-written LABVIEW program (National instruments). GCamp3.0, homer-GCaMP, and SynpH were excited at 920 nm and fluorescence monitored at an image acquisition rate of 5 Hz. The odorants were presented with a 20 s break between stimulation, and each fly was exposed to five to six repetitive experiments.
The images were aligned to reduce small shifts in the X–Y direction using a custom written ImageJ plugin. The mean intensity within the region of interest of five images before stimulus onset was used as baseline fluorescence (F0). The difference in intensity (ΔF) was calculated by subtracting F0 from the fluorescence intensity value within the ROI of each image (Fi) and, subsequently, divided by the baseline fluorescence. ΔF/F0 values of three or more repetitions were averaged for each fly.
Odor-induced fluorescence changes of SynpH were considered in calycal PN boutons showing ΔF/F0 values more than twice the standard deviation of the baseline fluorescence. The boutons with the five highest odor-induced ΔF/F0 amplitudes were considered for further analysis. We found SynpH to exhibit rapid photo-bleaching, therefore, bleaching correction was performed on its ΔF/F0 values. For this, first, ΔF/F0 values from the onset of the stimulus until the decay of the signal were removed and then the best least square fit was obtained using the remaining ΔF/F0 values (second order polynomial decay function). Subsequently, this decay function was subtracted from the entire original ΔF/F0 curve, and the new modified data are the bleaching corrected data.
Fluorescence emission of cytosolic GCamp was determined within specific boutons in the calyx that respond to the odor stimulus, and only the boutons showing ΔF/F0 values of more than 100% in four to five stimulations were averaged for each fly and considered for final analysis. Fluorescence changes of mb247-Gal4; UAS-homer-GCamp flies were averaged over the five most responsive microglomerular structures, as anatomically defined by basal fluorescence.
False color-coded images were obtained by subtracting the image just before stimulus onset from the image at the maximum of the intensity difference (i.e., at 2 s after odor onset) and divided by the baseline fluorescence.
The KCl experiments were performed using a fluorescence microscope (Zeiss) equipped with a xenon lamp (Lambda DG-4, Sutter Instrument), a 14-bit CCD camera (Coolsnap HQ, Photometrics) and a 20 × NA = 1 water-immersion objective. Images were acquired at 5 Hz using Metafluor (Visitron Systems). After recording some initial frames, KCl was added to the Ringer’s solution covering the fly brain (final concentration 0.05 M). Fluorescence changes were determined in a circular region covering the calyx (d = 20 μm), and background fluorescence determined outside the calyx was subtracted.
For the identification of (de)acetylated residues of BRP, we did “conventional” protein extractions from Drosophila heads combined with BRP immunoprecipitations. The protocol could be divided into four main sections. 1) Precoupling of antibodies to matrix (50 ug antibody per reaction): 3 LoBind cups (2 ml; Eppendorf) containing Affiprep Protein A matrix were prepared: 1 X 30ul for specific antibody, 1 X 30ul for IgG control, 1 X 60 ul for head extract preclearing. The cups were washed 3 X with 500 ul H-buffer (25 mM HEPES pH 8.3 (NaOH), 150 mM NaCl, 1 mM MgCl2, 1 mM EGTA, 10% Glycerol) by inverting several times, followed by centrifugation 1,000 gmax (3,000 rpm) for 1 min. 500 ul H-buffer (+ BRPlast200 or IgG) per coupling was prepared. 500 ul antibody solution (= 50 ug IgG) was added per 30 ul washed Protein A-beads. Beads were incubated with antibody solution for 2 h on the wheel at 4°C. The Affiprep beads-antibody were collected by centrifugation for 3 min at 1,000 gmax. Affiprep beads-Antibody were washed 3 X by inverting tubes and 3 X for 10 min on wheel with IP buffer. 2) Homogenizing fly heads from stored fly heads [–80°C]. Fly heads were transferred with a clean spatula into 1 ml glass homogenizer. For 300 ul frozen fly heads, 300 ul Homogenization buffer (without detergent) was added, and heads were sheared at 900 rpm using an electronic overhead stirrer. Samples were collected in LoBind cups (2 ml; Eppendorf). 2 X 300 ul was added to rinse pestle and homogenizer (Total volume in cups ~1,100–1,200 ul). Sodium-deoxycholic Acid (DOC) was added to a final concentration of 0.4% (28 ul of 10% stock spiked into homogenate (1:25 v/v)). Triton X-100 was added to a final concentration of 1% (35 ul of 20% stock spiked into homogenate (1:20 v/v)). The samples (Homogenate) were incubated for 60 min at 4°C at level 8 (slow) on wheel. 20 ul of homogenate was stored for SDS-PAGE analysis for monitoring antigen during extraction/pull-down procedure. Homogenate (H) was centrifuged for 15 min at 17,000 gmax. Supernatant (yellow in color) was transferred to a fresh LoBind cup. Centrifugation of S1 was repeated 4X to get rid of fat and remaining head debris. After final centrifugation step, remaining supernatant was diluted 1:1 with H-buffer (without detergent). Total volume of Input was ~1,400 ul and of following composition: 25 mM Hepes pH 8.05 (NaOH), 150 mM NaCl, 0.5 mM MgCl2, 0.5 mM EGTA, 5% Glycerol, 0.2% DOC, 0,55% Triton X-100. 3) Preclearing of fly head extract on Protein A-IgG beads: Diluted fly head extract was applied to preclearing beads and incubated for 60 min at 4°C while rotating on wheel. Precleared extract was separated by centrifugation for 3 min at 1,000 gmax. Supernatant (IP input) was recovered. 4) Precipitation: Precleared extract (IP input) was applied to antibody-bead matrix (600 ul to specific Antibody-beads, 600 ul to control IgGs) and antibody–antigen binding was performed overnight at 4°C. Immunoprecipitates were collected by centrifugation at 1,000 gmax for 4 min at 4°C. Affiprep Beads-Antibody-Antigen were washed 3 X with a quick rinse followed by 2 X 20 min with 1 mL IP Buffer (H-buffer + 0.5% Triton-X 100 + 0.2% Na-DOC). Affiprep Beads-Antibody-Antigen were resuspended in 1,000 ul IP buffer and transferred to a clean LoBind cup (2 ml; Eppendorf). Affiprep Beads-Antibody-Antigen were centrifuged, and most of the supernatant was removed (without removing beads). 4.) Elution: For elution, 100 ul of 2X Laemmeli Buffer was added to Affiprep Beads-Antibody-Antigen and heated for 10 min at 95°C, 600 rpm, followed by centrifugation for 5 min at 1,000 gmax. Supernatant (IP eluate) was transferred into a fresh LoBind Cup (2 ml; Eppendorf). Immunoprecipitation was verified with western blot.
For identification of (de)acetylated lysine residues in BRP, IP eluate was heated in SDS-PAGE loading buffer, reduced with 1 mM DTT (Sigma‐Aldrich) for 5 min at 95°C and alkylated using 5.5 mM iodoacetamide (Sigma‐Aldrich) for 30 min at 20°C. The protein mixtures were separated on 4%–12% gradient SDS‐PAGE (NuPAGE, Invitrogen). The gel lanes were cut into ten equal slices, the proteins were in-gel digested with trypsin (Promega) [92], and the resulting peptide mixtures were processed on STAGE tips [93] and analyzed by LC-MS/MS.
Mass spectrometric (MS) measurements were performed on an LTQ Orbitrap XL mass spectrometer (Thermo Fisher Scientific) coupled to an Agilent 1200 nanoflow–HPLC (Agilent Technologies GmbH, Waldbronn, Germany) [94]. HPLC–column tips (fused silica) with 75 μm inner diameter (New Objective, Woburn, MA, USA) were self-packed with Reprosil–Pur 120 ODS–3 (Dr. Maisch, Ammerbuch, Germany) to a length of 20 cm. Samples were applied directly onto the column without a precolumn. A gradient of A (0.5% acetic acid (high purity, LGC Promochem, Wesel, Germany) in water and B (0.5% acetic acid in 80% acetonitrile (LC–MS grade, Wako, Germany) in water) with increasing organic proportion was used for peptide separation (loading of sample with 2% B; separation ramp: from 10%–30% B within 80 min). The flow rate was 250 nl/min and for sample application 500 nl/min. The mass spectrometer was operated in the data-dependent mode and switched automatically between MS (maximum of 1 x 106 ions) and MS/MS. Each MS scan was followed by a maximum of five MS/MS scans in the linear ion trap using normalized collision energy of 35% and a target value of 5,000. Parent ions with a charge state from z = 1 and unassigned charge states were excluded for fragmentation. The mass range for MS was m/z = 370–2,000. The resolution was set to 60,000. MS parameters were as follows: spray voltage 2.3 kV; no sheath and auxiliary gas flow; ion transfer tube temperature 125°C.
Data were analyzed in R v3.1.2 using the additional CRAN package dunn.test v1.2.2. Asterisks are used in the figures to denote significance: * p < 0.05, ** p < 0.01, *** p < 0.001, ns = not significant. Nonparametric methods were used because of the small sample sizes and because of failure of tests for normality for parts of the data (Shapiro-Wilk test). Unless indicated otherwise, the different groups in each figure were first compared using the Kruskal-Wallis test, followed by Dunn’s test for posthoc multiple comparisons. Nonparametric tests were used in order to avoid being biased by outliers, which are represented by solid circles. All p-values that are reported have been subject to Bonferroni correction for the number of comparisons. Additional relevant information is indicated in the figure legends. The data for the behavioral studies were collected with the investigator blind to the genotypes, treatment, and age of genotypes. There was no blinding in the other experiments. The data were collected and processed side by side in randomized order for all experiments. In order to analyze the difference in Homer-GCaMP3.0 responses (Fig 6 and S17 Fig), two-sided Kolmogorov-Smirnov tests were conducted in R, and the GCaMP3 responses only during odor stimulation and were compared.
|
10.1371/journal.pntd.0002052 | An Integrated Linkage, Chromosome, and Genome Map for the Yellow Fever Mosquito Aedes aegypti | Aedes aegypti, the yellow fever mosquito, is an efficient vector of arboviruses and a convenient model system for laboratory research. Extensive linkage mapping of morphological and molecular markers localized a number of quantitative trait loci (QTLs) related to the mosquito's ability to transmit various pathogens. However, linking the QTLs to Ae. aegypti chromosomes and genomic sequences has been challenging because of the poor quality of polytene chromosomes and the highly fragmented genome assembly for this species.
Based on the approach developed in our previous study, we constructed idiograms for mitotic chromosomes of Ae. aegypti based on their banding patterns at early metaphase. These idiograms represent the first cytogenetic map developed for mitotic chromosomes of Ae. aegypti. One hundred bacterial artificial chromosome clones carrying major genetic markers were hybridized to the chromosomes using fluorescent in situ hybridization. As a result, QTLs related to the transmission of the filarioid nematode Brugia malayi, the avian malaria parasite Plasmodium gallinaceum, and the dengue virus, as well as sex determination locus and 183 Mbp of genomic sequences were anchored to the exact positions on Ae. aegypti chromosomes. A linear regression analysis demonstrated a good correlation between positions of the markers on the physical and linkage maps. As a result of the recombination rate variation along the chromosomes, 12 QTLs on the linkage map were combined into five major clusters of QTLs on the chromosome map.
This study developed an integrated linkage, chromosome, and genome map—iMap—for the yellow fever mosquito. Our discovery of the localization of multiple QTLs in a few major chromosome clusters suggests a possibility that the transmission of various pathogens is controlled by the same genomic loci. Thus, the iMap will facilitate the identification of genomic determinants of traits responsible for susceptibility or refractoriness of the mosquito to diverse pathogens.
| About half of the human population is under risk of dengue infection. Because of the absence of a vaccine or drug treatment, the prevention of this disease largely relies on controlling its major vector mosquito Aedes aegypti. Availability of the complete genome sequence for this mosquito offers the potential to help in the identification of novel disease control strategies. An efficient vector of arboviruses, Ae. aegypti is also a convenient model for laboratory studies. A number of genetic loci related to the remarkable ability of this mosquito to transmit various pathogens were genetically mapped to the three linkage groups corresponding to the three individual chromosomes of the mosquito. However, the exact physical positions of the genetic loci and genomic sequences on the chromosomes were unknown. In this study, we developed maps for mitotic chromosomes of Ae. aegypti and localized 100 clones carrying major genetic markers, which were previously used for mapping genetic loci associated with the pathogens' transmission. Finally, linkage, chromosome, and genome maps of Ae. aegypti were integrated. Anchoring of the genomic sequences associated with genetic markers to the chromosomes of Ae. aegypti will help to identify candidate genes that might be utilized for developing advanced genome-based strategies for vector control.
| Mosquitoes are vectors of numerous human pathogens such as malaria parasites transmitted by the subfamily Anophelinae; lymphatic filarial worms transmitted by both Anophelinae and Culicinae subfamilies; and arboviruses whose transmission is largely associated with the subfamily Culicinae. Aedes aegypti is recognized as a principal vector of dengue and yellow fever viruses. These two diseases have a significant world-wide impact on human health. Dengue fever is currently considered the most important vector-borne arboviral disease of the 21st century [1]. The disease is a threat to 3.6 billion people with an annual incidence of 230 million cases of infection resulting in 21,000 deaths per year. Since the 1950s, the incidence of dengue fever has expanded globally. The World Health Organization (WHO) estimated a 30-fold increase in the incidence of dengue infections over the past 50 years [2]. The disease became endemic in 100 countries in Africa, West Asia, and America [3] and is a growing threat to the United States [4]. In addition to dengue, yellow fever, a devastating disease of the 19th century in North America and Europe, still affects up to 600 million lives and remains responsible for about 30,000 deaths annually [5]. The disease is currently endemic in 32 countries in Africa and 13 in South America. The spread of the pathogens is associated with the extremely tight connection of their major vector Ae. aegypti to humans. Despite all control campaigns, Ae. aegypti currently occupies most subtropical and tropical regions in the world.
Ae. aegypti represents both an efficient vector of arboviruses and a convenient model system for experimental laboratory research. This species can be easily colonized and is highly tolerant to inbreeding [6]. Unlike Anopheles eggs, Ae. aegypti eggs are resistant to desiccation and can be stored in a dry place for several months. As a result of these advantages, genetic (linkage) mapping conducted on Ae. aegypti was very successful. The genetic mapping was originally inspired from the study of the inheritance of dichlorodiphenyltrichloroethane (DDT) resistance as a single dominant trait [7]. A similar mechanism of inheritance, as a single gene or a single block of chromosome material, was demonstrated for sex determination [8]. The sex determination alleles were described as Mm in males and mm in females and linked to homomorphic chromosome 1 [9]. In addition, 28 of 87 morphological mutations described for Ae. aegypti were mapped to the three linkage groups corresponding to the three chromosomes of this mosquito [10]. The linkage map was extended by additional mapping of physiological and enzyme loci [11]. The classical linkage map included about 70 loci of morphological mutants, insecticide resistance, and isozyme markers [12].
A possibility of using DNA molecular markers opened a new era in genetic mapping of mosquito genomes. The first molecular-marker-based linkage map for Ae. aegypti was constructed using restriction-fragment-length polymorphism (RFLP) of complementary DNA (cDNA) clones [13]. This map included 50 DNA markers and covered 134 centimorgan (cM) across the three linkage groups. Thereafter, the polymerase chain reaction (PCR) was used to generate a map based on random-amplified polymorphic DNA (RAPD) loci which consisted of 96 RAPD loci covering 168 cM [14]. Linkage maps based on single-strand conformation polymorphism (SSCP) and single nucleotide polymorphism (SNP) markers were also constructed. A composite map for RFLP, SSCP, and SNP markers incorporated 146 loci and covered 205 cM [6]. Later, an additional map using amplified fragment-length polymorphism (AFLP) was also developed for 148 loci and covered about 180 cM of the genome [15]. Finally, the genetic map of Ae. aegypti was extended by incorporating microsatellite loci [16]. The linkage map was used as a tool to localize several quantitative trait loci (QTLs) related to pathogen transmission: the filarioid nematode Brugia malayi [17], the avian malaria parasite Plasmodium gallinaceum [15], [18], and dengue virus [19], [20]. Among all mosquitoes, the map developed for Ae. aegypti is the most densely populated.
Cytogenetic mapping on Ae. aegypti and other culicine species is difficult due to the absence of high-quality, easily spreadable polytene chromosomes [21], [22] The majority of cytogenetic studies for Ae. aegypti have been conducted on mitotic chromosomes from brain ganglia or meiotic chromosomes from male testis [23]–[26]. These studies led to the conclusion that Ae. aegypti has a karyotype consisting of three pairs of metacentric chromosomes [27]. The chromosomes first numbered as chromosomes I, II, and III in order of increasing size [23] were later renumbered as chromosomes 1, 2, and 3 in correspondence to the linkage map developed for Ae. aegypti, resulting in the longest chromosome III becoming chromosome 2 [9]. Chromosomes from brain ganglia of Ae. aegypti were first utilized for the successful nonfluorescent in situ hybridization of two ribosomal genes [28]. Fluorescent in situ hybridization (FISH) technique has been developed using mitotic chromosomes from the ATC-10 cell line of Ae. aegypti, resulting in direct positioning of 37 cosmid clones onto chromosomes [29]. In addition, 21 cDNA genetic markers and 8 cosmid clones containing the RFLP markers have been mapped to the chromosomes from this line [30]. This map was the first attempt to integrate linkage and physical maps for Ae. aegypti.
The genome of Ae. aegypti was among the first three mosquito genomes sequenced in the last decade [31]. As compared with Anopheles gambiae [32] and Culex quinquefasciatus [33] genomes, the genome of Ae. aegypti is the largest and consists of 1,376 Mb. The availability of the Ae. aegypti genome provides an opportunity to integrate linkage, chromosome, and genome maps for this mosquito. A total of 106 bacterial artificial chromosome (BAC) clones carrying major genetic markers have been identified by screening an Ae. aegypti BAC library prepared from the Liverpool strain [34]. In addition, a new cytogenetic approach based on mitotic chromosomes from imaginal discs (IDs) of 4th instar larvae has been recently developed [35]. Instead of using cell lines, which usually accumulate chromosomal rearrangements [36], this method utilizes live larvae for cytogenetic analysis. A preparation slide of one ID contains ∼175 chromosome spreads. This number is 6-fold greater than that of two brain ganglia. Clearly visible banding patterns of mitotic chromosomes from IDs allowed the construction of preliminary idiograms without numbered divisions for the chromosomes at mid-metaphase [35]. A FISH technique was optimized for using BAC clones as probes [37] resulting in the assignment of 10 BAC clones and ribosomal 18S DNA to bands on the idiograms [35].
In the current study, we constructed new idiograms for the longer early-metaphase chromosomes with numbered divisions and subdivisions. These idiograms facilitated assignment of 100 BAC clones carrying major genetic markers to chromosomal bands. BAC clones within each band on the chromosomes were additionally ordered based on multicolor FISH on higher resolution prophase or polytene chromosomes. Finally, because each BAC clone also represents a supercontig in the Ae. aegypti genome assembly, the total of 183 Mb or 13.3% of the genomic sequences was also incorporated into the map. We define our map as integrated linkage, chromosome, and genome map or an iMap of Ae. aegypti.
This study was performed on the Liverpool IB12 strain of Ae. aegypti, which was previously used for the genome sequencing project [31]. This strain originated from the Liverpool strain (LVP) following several rounds of inbreeding. LVP was a major strain for conducting genetic and QTL mapping in the past [13], [17], [18]. Originally, mosquitoes for this strain were collected in West Africa and then kept by the Liverpool School of Medicine [38].
Chromosome preparations were prepared from imaginal discs (IDs) or salivary glands of the 4th instar larvae of Ae. aegypti (Timoshevskiy et al., 2012). For ID dissection larvae were placed on ice for several minutes for immobilization. Individual larva was decapitated in a drop of cold hypotonic solution (0.5% sodium citrate). Then the thoracic part of the larva was dissected, and cuticle from the ventral side of the larval thorax was cut by dissecting scissors (Fine Science Tools, USA) and opened. The gut and fat body particles were removed, and a new drop of hypotonic solution was applied. After 10 min., the hypotonic solution was removed using filter paper, and a drop of fixative solution (ethanol/acetic acid in 3∶1 ratio) was applied. IDs were isolated using dissecting needles (Fine Science Tools, USA), placed in a drop of 50% propionic acid for maceration, and squashed under a cover slip (22×22 mm). Salivary glands were dissected from the larvae prefixed in fixative solution (ethanol/acetic acid in 3∶1 ratio) for at least 24 hrs and then squashed in a drop of 50% propionic acid. Presence of chromosomes on the slide was determined by using a phase-contrast microscope Olympus CX41 (Olympus America, Inc., USA) at 200× magnification. Slides suitable for further applications were placed in liquid nitrogen and cover slips were removed. Finally, slides were dehydrated in an ethanol series (70, 80, 100%) and stored at −20°C.
BAC clone DNA was prepared by the Clemson University Genomics Institute in 96-well plates. For the probe preparation, BAC clone DNA was labeled by nick-translation. The reaction mix with final volume of 25 µl contained 0.5 µg of DNA, 0.05 mM each of unlabeled dATP, dCTP, and dGTP, and 0.015 of mM dTTP, 0.5 µl of Cy3-, or Cy5-dUTP (GE Healthcare UK Ltd, Buckingham-shire, UK), or 1 µl fluorescein-12-dUTP (Fermentas, Inc., USA), 0.05 mg/ml BSA, 2.5 µl of 10× nick translation buffer, 10 u of DNA polymerase I, and 0.0006 u of DNAse I (Fermentas, Inc., USA). DNA polymerase/DNAse ratio was selected empirically to obtain the probe with a size range from 300 to 500 base pair. For performing FISH with additional colors (besides Cy3, Cy5, and fluorescein), a pair combination of equal volumes of differently labeled probes was used.
Optimized methods for isolation of the repetitive DNA fraction for Ae. aegypti was described by Timoshevskiy et al., 2012. Genomic DNA was isolated from adult mosquitoes using Qiagen Blood & Cell Culture DNA Maxi Kit (Qiagen Science, USA). For individual extractions, approximately 500 mg of adult mosquitoes were taken. For further manipulation, DNA was dissolved in 1.2× SSC buffer to final concentration of 1 µg/µl. For shearing, a denatured DNA solution was heated at 120°C for 2 min. Reassociation of the DNA was performed at 60°C for 10 min or 15 min. After reannealing, samples were placed in ice, and 10× S1-nuclease buffers and S1 nuclease (100 U per 1 mg DNA) were added. Nuclease treatment was performed at 42°C for 1 hr. Isolated repetitive DNA fractions were precipitated by isopropanol and dissolved in TE-buffer. According to our estimation, the repetitive fractions isolated using this approach correspond to Cot2 or Cot3 DNA fraction and contained all highly repetitive and part of medium-repetitive DNA fragments (Trifonov et al., 2009). Final outcome of Cot DNA fraction accounts for ∼30% of the Aedes genomic DNA. Repetitive DNA fractions of genomic DNA were utilized to suppress repetitive sequences in hybridizations to the chromosomes.
In situ hybridization was performed using a modified standard human protocol (Timoshevskiy et al., 2012). Slides were placed in 2× SSC for 30 min at 37°C, pretreated with 0.1 mg/ml of pepsin for 5 min at 37°C, denaturated in 70% formamide in 2× SSC at 72°C for 2 min, and dehydrated in a series of cold (−20°C) ethanol (70, 80, 100%) for 3–5 min each. Hybridization mix contained: 50% formamide, 10% dextran sulfate, 100 ng of each probe per slide, and 3 µg of unlabeled repetitive DNA fractions per probe. DNA/probe mix was precipitated by adding 1/10 volume of sodium acetate and 2 volumes of 100% ethanol. The DNA pellet was dissolved in “master mix” (10 µl per slide) that contained 50% formamide, 10% dextransulfate, and 1.2× SSC. After that, DNA was denatured at 96°C for 7 min. Denatured DNA was placed on ice for 1 min and incubated at 37°C for 30 min for pre-hybridization with unlabeled repetitive DNA fractions. Ten µl of hybridization mix was placed on a slide, which had been preheated to 37°C, under a 22×22 mm cover slip, and glued by rubber cement. Slides were hybridized at 37°C in a dark humid chamber overnight. After hybridization, slides were dipped for washing in a Coplin jar with 0.4× SSC, 0.3% Nanodept-40 at 72°C for 2 min, and then in 2× SSC, 0.1% Nanodept-40 at RT for 5 min. Thereafter, slides were counterstained using Prolong with DAPI (Invotrogen Corporation, USA) or incubated with 1 µM YOYO-1 solution in 1× PBS for 10 min in the dark, rinsed in 1× PBS, and then enclosed in antifade Prolong Gold (Invitrogen Corporation, USA) under a cover slip. Slides were analyzed using a Zeiss LSM 510 Laser Scanning Microscope (Carl Zeiss Microimaging, Inc., USA) at 1000× magnification.
To develop idiograms, the best images of the chromosomes from IDs stained with YOYO-1 were selected. The colored images were converted into black and white images and contrasted in Adobe Photoshop as described previously [39]. The chromosomal images were straightened using ImageJ program [40] and were aligned for comparison. In total, ∼90 chromosomes at early metaphase were analyzed. FISH images were also filtered using ImageJ program [40]. For ordering genetic markers, chromosomes at various levels of condensation were used. Prometaphase and early metaphase chromosomes were utilized for assigning genetic markers to the particular chromosome bands. Prophase and polytene chromosomes were used for ordering markers within the same chromosome band. From 10 to 20 images were analyzed to obtain reproducible ordering patterns.
Chromosomes were measured as described previously [34] using Zen 2009 Light Edition software [41]. The relationship between the physical locations of markers and their linkage positions was assessed by assigning genes of known physical position an integer score. These scores were 1–26 (1p3.4–1q4.4) on chromosome 1, 1–36 (2p4.4–2q4.4) on chromosome 2, and 1–32 (3p4.4–3q4.4) on chromosome 3. This integer score was then regressed upon the cM position of the gene as determined in a number of previous independent linkage mapping studies and F1 intercross families. Linear regression analysis was performed using R (2.14.1) [42]. These regression models were then used to predict the physical position of the markers for which we have linkage positions in cM.
Our previous study developed preliminary idiograms – diagrammatic representations of the chromosome banding patterns – for the mid-metaphase chromosomes of Ae. aegypti [35]. This stage of mitosis is the most representative in any chromosome preparation. At mid metaphase, chromosomes and chromosome arms can be easily distinguished from each other based on their length and presence of specific landmarks. Chromosomes on preliminary idiograms were not divided into divisions and subdivisions, and these idiograms served only for chromosome and chromosome arm identification. In the current study, we developed idiograms for the chromosomes of Ae. aegypti at an earlier stage of mitosis – early metaphase. The average chromosome lengths at this stage are 11.86, 16.19, and 13.64 µm for chromosomes 1, 2, and 3, respectively, or ∼5.5 µm longer than at mid metaphase. At stages of mitosis previous to metaphase, such as prophase and prometaphase, homologous chromosomes of Ae. aegypty are usually tightly paired [35]. Although chromosomes at these stages are longer than at early metaphase, the banding patterns of the chromosomes are variable. At early metaphase, chromosomes finally segregate from each other, resulting in a visible number of chromosomes becoming equal to 6 and chromosome patterns becoming more reproducible (Figure 1A). We consider this stage of mitosis as the most reliable for the development of the chromosome map, which can be used for the detailed physical mapping.
Similarly to our previous study [35], we used chromosome images stained with YOYO-1 iodide. This dye provides much clearer banding patterns as compared with the most commonly used DAPI (4′,6-diamidino-2-phenylindole fluorescent stain). Images of the chromosomes were converted into black and white images (Figure 1A) and straightened. Similarly to the idiograms of human chromosome [43], we identified chromosome bands of 4 different intensities – intense, medium intensity, low intensity, and negative (Figure 1B). Chromosomes were finally subdivided into 23 numbered divisions and 94 bands. The following regions can be considered as landmarks for the chromosome arm recognition: intense band in division 1q21; intense double band in divisions 2q21–23; and 2 low-intensity bands in region 3q32, 3q34. These regions are shown by asterisks on Figure 1. Large negative bands indicate the boundaries between all divisions on the chromosomes.
Chromosome idiograms constructed in this study represent the first cytogenetic map developed for mitotic chromosomes of Ae. aegypti. In studies conducted before on chromosomes from cell lines of Ae. aegypti, the positions of the markers on the chromosomes were measured by FLpter: a fractional length from the short-arm telomeric end p-terminus [30]. As a result, this mapping provides only approximate coordinates for the markers. The idiograms recently developed for mid-metaphase chromosome [35] were designed mostly for individual chromosome and chromosome arm recognition. The map for early metaphase chromosome presented here is designed for the more detailed band-based mapping. It finally permits assignment of the location of the specific DNA signals to the particular numbered subdivision on the chromosomes.
Previous efforts identified 106 BAC clones that carry genetically mapped marker sequences [34]. We used a “two-step” physical mapping approach for 1) assigning BAC clones to the chromosomal bands and 2) ordering them within the band. For the first step, we utilized FISH on chromosomes stained with the green dye YOYO-1 iodide. BAC clone DNA was labeled with Cy3 (red) and Cy5 (infrared) fluorescent dyes. Examples of FISH results on early metaphase chromosomes are shown in Figure 2. Each FISH allowed us to place two BAC clones to a specific band on idiograms. Eight BAC clones produced more than one hybridization signal. In these cases, the most intense signal was considered as a major position of the BAC clone on the chromosomes. In total, 100 out of 106 BAC clones were successfully assigned to specific bands on Ae. aegypti chromosomes (Table 1).
For ordering BAC clones within one band, FISH was performed on prophase chromosomes from imaginal discs or, for higher resolution, on polytene chromosomes from salivary glands of 4th instar larvae of Ae. aegypti. Probes were labeled with three different dyes: fluorescein (green), Cy3 (red), and Cy5 (infrared), or with a combination of these dyes (Figure 3). Chromosomes after FISH were stained with DAPI (ultraviolet). This approach allowed mapping of up to 6 BAC clones simultaneously. Our FISH results showed that two probes have to be separated by a distance of ∼0.5 Mb on prophase chromosomes in order to be distinguished from each other (Figure 3A). The resolution of mapping using polytene chromosomes was even higher, ∼300 kb (Figure 3D). As a result of this additional mapping, all 100 BAC clones were placed in correct order on the chromosomes (Figure 4).
The physical map constructed in this study is the most populated physical map developed for Ae. aegypti thus far. Our current mapping effort placed 100 BAC clones and an 18S rDNA probe to their particular regions on the chromosomes. The “two-step” mapping approach significantly improved the resolution of the mapping. Using long prophase chromosomes and low-polytenized chromosomes from salivary glands, in addition to early metaphase chromosomes, provided the resolution similar to that obtained on polytene chromosomes from ovaries of An. gambiae, which is equal to ∼100 kb [44]. The current study developed a simple and robust technique for high-resolution physical mapping that can be further applied for more detailed physical mapping of the Ae. aegypti genome and other mosquito genomes. Similar to studies conducted on Drosophila [45]–[48], the physical mapping approach based on the banding patterns of mitotic chromosomes can also be used for the additional mapping of An. gambiae heterochromatin, which is under-replicated in normal polytene chromosomes.
Mapping of the BAC clones that carry particular genetic markers allowed us to clarify the order of the genetic markers (Figure 4). Genetic markers physically mapped in this study span the entire chromosome complement. The longest chromosome 2 was the most densely populated with 45 genetic markers. The highest number of markers was found in areas close to the telomeres, especially on the p arm of chromosome 2. In contrast to the previous study [30], some markers were located around the centromeres. Surprisingly, two areas in the middle of the short arms on chromosome 1 and 2 (regions 1p2 and 2p2) had extremely low density of markers. A linear regression analysis demonstrated a good overall correlation between positions of the markers on the physical and linkage maps: R2 equaled to 0.69, 0.73, and 0.86 for chromosomes 1, 2, and, 3 respectively (Figure 5). On average, 1 cM on the linkage map corresponds to the half of a cytogenetic band or to 6.88 Mbs on the physical genome map. However, we found large discrepancies between the two maps with respect to the distances among markers located in areas near the centromeres and telomeres. These discrepancies probably caused by the high rate of recombination near telomeres and the low rate of recombination near the centromeres.
Based on previous studies, we were able to physycally map the positions of the QTLs related to the ability of Ae. aegypti to transmit different pathogens. The QTLs associated with the transmission of dengue virus 2 (DEN2) [19], [20]; filarioid nematode Brugia malayi [17] and the avian malaria parasite Plasmodium gallinaceum [15], [18] are indicated by different colors on Figure 4. As a result of physical mapping, 12 QTLs on the linkage map “collapsed” into five clusters of QTLs on the chromosome map in regions: 1p31–34; 2p33–42; 2q31–32 and 3p31–32 (Figure 6). Interestingly, four QTLs on 2p arm related to the transmission of various pathogens, such as filarioid nematode, avian malaria parasite, and dengue virus 2, were placed by different studies within a large region on the linkage map (31.8 cM or about 50% of chromosome 2). However, they were physically mapped within four chromosomal bands, which encompass only ∼11% of chromosome 2. These results suggest that the susceptibility of Ae. aegypti to diverse pathogens is controlled by fewer genomic loci than it was previously considered.
In addition to QTLs, the location of marker LF284, an anchoring marker for the sex determination locus [49], was also determined on chromosome 1 (Figure 4). This marker was localized in the intensively stained band in region 1q21. This region is located next to the ribosomal locus in negative band 1q22 that usually forms secondary constriction and can be easily identified on the chromosomes. Two BAC clones with markers AEGI10 and LF231 were found in conflict with previous mapping positions on different chromosomes. This result is not unexpected, as the genetic linkage map is a composite based on results of multiple independent genetic crosses [44].
Finally, the availability of the Ae. aegypti genome allowed us to map 100 genomic supercontigs to the chromosomes (Figure 4). Four supercontigs—1.123, 1.219, 1.14, and 1.1, which contained two or more genetic markers,—were oriented on the chromosomes. The orientation of these supercontigs is indicated by arrows on Figure 4. Physical mapping also helped us to identify potentially misassembled supercontigs if two or more BAC clones located in the same genomic supercontigs were mapped to different chromosomes. Our data suggests that three genomic supercontigs 1.148, 1.1, and 1.209 were misassembled in the previous study [31]. Potentially misassembled supercontigs are indicated in bold on Figure 4. In total, our mapping effort placed 183 Mb of genomic supercontigs, which is equal to 13.3% of the genome, to the chromosomes. The chromosome-based genome map for Ae. aegypti developed in this study is the second after the An. gambiae genome map developed for mosquitoes [32], [44], [50].
The genomes of the three most dangerous for the human health species of mosquitoes—Aedes aegypti, Anopheles gambiae, and Culex quinquefasciatus—were sequenced in the last decade. The genome of Ae. aegypti is the largest among the three species and consists of 1,376 Mb. Our physical mapping effort incorporated 94 cytogenetic bands, 100 molecular genetic markers, and 183 Mb of the genome into one iMap of Ae. aegypti. The locations of anchor markers for QTLs related to dengue virus, filarial nematode, and malaria parasite transmission were determined on the chromosomes, as well as for the sex determination locus. Our discovery of the localization of multiple “unrelated” QTLs in a few major chromosome clusters suggests a possibility that the transmission of different pathogens is controlled by the same genomic loci. The study also demonstrated that physical mapping can orient genomic supercontigs and identify potential mistakes in genome assembly. Thus, the iMap developed here will facilitate the identification of genomic determinants of traits responsible for susceptibility or refractoriness of the mosquito to diverse pathogens and will also guide future efforts to improve the assembly of Ae. aegypti genome.
|
10.1371/journal.pgen.1004318 | R-loops Associated with Triplet Repeat Expansions Promote Gene Silencing in Friedreich Ataxia and Fragile X Syndrome | Friedreich ataxia (FRDA) and Fragile X syndrome (FXS) are among 40 diseases associated with expansion of repeated sequences (TREDs). Although their molecular pathology is not well understood, formation of repressive chromatin and unusual DNA structures over repeat regions were proposed to play a role. Our study now shows that RNA/DNA hybrids (R-loops) form in patient cells on expanded repeats of endogenous FXN and FMR1 genes, associated with FRDA and FXS. These transcription-dependent R-loops are stable, co-localise with repressive H3K9me2 chromatin mark and impede RNA Polymerase II transcription in patient cells. We investigated the interplay between repressive chromatin marks and R-loops on the FXN gene. We show that decrease in repressive H3K9me2 chromatin mark has no effect on R-loop levels. Importantly, increasing R-loop levels by treatment with DNA topoisomerase inhibitor camptothecin leads to up-regulation of repressive chromatin marks, resulting in FXN transcriptional silencing. This provides a direct molecular link between R-loops and the pathology of TREDs, suggesting that R-loops act as an initial trigger to promote FXN and FMR1 silencing. Thus R-loops represent a common feature of nucleotide expansion disorders and provide a new target for therapeutic interventions.
| Friedreich ataxia and Fragile X syndrome are among 40 human diseases associated with expansion of repeated sequences. In both disorders repeat expansion leads to gene silencing, the molecular mechanism of which is not well understood, impeding the development of specific therapies to treat these disorders. It is proposed that formation of unusual DNA structures (RNA/DNA hybrids, or R-loops) over repeat regions may play a role, but their molecular function has not been investigated in vivo. We show that R-loops form on expanded repeats of FXN and FMR1 genes in cells from FRDA and FXS patients. These R-loops are stable, correlate with repressive chromatin marks and hinder FXN transcription in patient cells. We studied the relationship between repressive chromatin and R-loops. Decrease in the amount of repressive chromatin has no effect on R-loop levels. In contrast, increase in the R-loops leads to transcriptional silencing of FXN gene and formation of repressive chromatin, providing a direct molecular link between R-loops and pathology of expansion diseases. This discovery is important for understanding the basic molecular mechanism underlying the pathology of expansion diseases. The ability of R-loops to trigger transcriptional silencing makes them an attractive target for future therapeutic approaches to treat these diseases.
| Around forty human diseases are associated with expanded repeat sequences [1]. Friedreich ataxia (FRDA) is the most frequent autosomal recessive ataxia (2–4 cases/100,000), caused by a GAA expansion in the first intron of the frataxin (FXN) gene, which encodes a mitochondrial protein involved in iron-sulfur cluster biogenesis [2], [3]. The GAA expansion leads to reduced levels of FXN mRNA and protein [4]–[6]. Several mechanisms mediating FXN transcriptional silencing have been proposed, including the formation of unusual DNA structures (triplex DNA and RNA/DNA hybrids) and repressive heterochromatin over expanded repeats [5]–[10].
RNA/DNA hybrids (R-loops) are formed during transcription, when nascent RNA hybridizes to the DNA template behind the elongating RNA polymerase (Pol II). R-loops are detected in organisms from bacteria to humans and implicated in many processes [11]. In mammalian cells, R-loops were originally discovered in the immuno-globulin class switch regions, essential for generating the antibody diversity in mouse activated B cells [12], [13]. R-loops also accumulate in cells depleted of the key splicing factor SRSF1, resulting in genome instability and appearance of double-strand breaks [14]. Recent studies demonstrated that R-loops are enriched over CpG promoters and may be involved in protection of these regions from DNA methylation and maintaining the hypomethylated state of CpG promoters [15]. We recently showed that R-loops formed over the G-rich pause sites downstream of the polyA signal in human genes are essential for the process of transcriptional termination of RNA Pol II [16]. Furthermore RNA/DNA hybrids are induced at GAA repeats following in vitro transcription and in bacteria [17], [18]. Also R-loops formed on plasmids containing CTG/CAG repeats in E.coli and mini-gene constructs in human cells promoted repeat instability, pointing towards their role in disease pathology [19], [20]. However, the direct involvement of R-loops on endogenous expanded alleles in the pathology of FRDA has not yet been investigated in vivo.
Our study shows that RNA/DNA hybrids (R-loops) form on expanded repeats of endogenous FXN and FMR1 genes, associated with Friedreich ataxia and Fragile X (FXS) disorders, in patient cells. These transcription-dependent R-loops are resistant to cellular degradation and co-localise with repressive H3K9me2 chromatin marks, characteristic of these diseases. Using nascent nuclear run-on analysis we show that R-loops over expanded repeats impede RNA Polymerase II transcription of the FXN gene in patient cells. We investigated the interplay between repressive chromatin marks and R-loops on the FXN gene. We show that a decrease in repressive H3K9me2 chromatin mark has no effect on R-loop levels and FXN transcription. In contrast, increasing R-loop levels leads to transcriptional repression of FXN gene, providing a direct molecular link between R-loops and pathology of FRDA. These data suggest that R-loops formed over expanded repeats act as an initial trigger to promote FXN and FMR1 silencing, and represent a common feature of nucleotide expansion diseases, contributing to their pathology in vivo.
We examined transcriptional regulation of the FXN gene in immortalized lymphoblastoid cells derived from FRDA patients, where FXN mRNA expression is reduced by ∼80% (Figure 1A–C). Pol II chromatin immuno-precipitation (ChIP) analysis in these cells showed that Pol II is enriched over the exon 1, positioned at the major transcriptional start site (TSS2) in lymphoblasts, correlating with the promoter-specific histone H3 depleted region [4] (Figure 1D, S1). Pol II levels over exon 1 were significantly reduced in FRDA cells. Similarly, using quantitative RT-PCR (RT-qPCR) in three independent control and three FRDA cell lines, a dramatic reduction in nascent RNA was detected over exon 1 in FRDA cells, further confirming a defect in transcription initiation (Figure 1E, S2A and S2B left panels). We also observed ∼10-fold reduction in the nascent RNA downstream of the expansion in regions D–G in FRDA cells. Overall Pol II ChIP and RT-qPCR results suggest transcriptional initiation and elongation defects triggered by expanded repeats, in line with previous reports [4], [6], [21].
Recently we established the DNA immuno-precipitation (DIP) method, which allows detection of R-loops on endogenous human genes in vivo using S9.6 antibody which recognizes RNA/DNA hybrids [16], [22]. Here we employed DIP to investigate R-loop distribution on the FXN gene. As a positive control we used the intron 1 region of the γ-actin gene, where high levels of R-loops are detected [15]. Significantly, we observed ∼3-fold enrichment of R-loops over regions B, C and D in the FXN intron 1 in FRDA cells, compared to control cells (Figure 1F). R-loops were concentrated over the expanded repeat region and were low in the downstream regions E–G. γ-actin R-loop levels were similar in control and FRDA cells (Figure 1F). Similar R-loop enrichment over expanded GAA repeats was detected in two additional independent FRDA cell lines (Figure S2A–B, right panels). Interestingly, when we compared the DIP data from all control and FRDA cell lines we observed that the level of R-loops correlates with expansion length (Figure S2C). To confirm the specificity of the DIP signal, we treated the samples with RNase H, which specifically degrades the RNA in RNA/DNA hybrids, prior to immuno-precipitation. Following RNase H digestion, the signal was strongly reduced for control γ-actin and FXN regions, suggesting that genuine R-loops are formed over the expanded GAA repeats (Figure 1G). High level of R-loops detected in FRDA cells may also suggest that these structures are particularly stable over FXN expanded repeats. Therefore, R-loops could act in cis to affect FXN gene expression in FRDA cells.
To understand the function of R-loops in FRDA pathology, we further characterized these structures over the expanded FXN allele. In particular, we treated cells with the transcriptional inhibitor actinomycin D. Following this treatment for 21 hours we observed an ∼80% reduction in γ-actin nascent RNA and R-loop signal, suggesting that γ-actin R-loops are quickly turned over in the cell (Figure 2A, B). In contrast, although nascent FXN RNA decreased following actinomycin D treatment, no change in R-loop levels was detected. However, we did finally observe a significant decrease in the level of R-loops over expanded repeats following prolonged treatment with actinomycin D for 48 hours (Figure S3). Overall these results suggest that R-loops associated with FXN expanded repeats are resistant to degradation. This may relate to the expanded GAA repeat property of transcriptional repression and repeat instability. We also observed that enrichment of R-loops correlated with highest peaks of the repressive histone modification H3K9me2 over the FXN regions B–D in FRDA cells (Figure 2C and 1F). This suggests that R-loops over expanded GAA repeats may functionally associate with repressive heterochromatin and be involved in mediating transcriptional repression.
Previously we showed that R-loops formed at the 3′ends of human genes promote transcriptional termination of RNA Pol II [16]. We therefore investigated if R-loops over FXN expanded repeats affect Pol II elongation. Here we employed nuclear run-on (NRO) analysis with Br-UTP labelled nucleotide [16], which measures actively transcribing Pol II (Figure 2D), in contrast to ChIP, which detects the total Pol II level on the gene. Using NRO we observed a substantial decrease in active transcription upstream of the GAA expansion (regions in1 and B) in FRDA cells, confirming our Pol II ChIP results (Figure 2E and 1D). In addition, we also detected ∼3-fold reduction in active transcription in FRDA cells over FXN regions D and E, positioned 210 nt and 4.5 kb downstream of expansion, respectively (Figure 2E). This elongation defect is not due to the increased distance caused by GAA expansion (∼3 kb), since we observed no decrease in active transcription between regions D to E, separated by ∼4.3 kb, in both cell lines. This suggests that expanded repeats directly interfere with active Pol II transcription in FRDA cells.
To test if expanded GAA repeats trigger the formation of R-loops in a different genomic location, we employed HEK293 cells, containing a copy of the FXN gene with either six (FXN-Luc) or ∼310 (FXN-GAA-Luc) GAA repeats, fused to the luciferase gene, integrated on chromosome 1, while the endogenous FXN gene lies on chromosome 9 (Figure 3A) [23]. We confirmed the presence of the GAA expansion using PCR on genomic DNA extracted from these cells. As demonstrated in Figure 3B, the FXN-GAA-Luc cell line indeed contains ∼310 expanded repeats. The presence of GAA repeats caused a reduction of ∼37% in FXN-luciferase nascent RNA levels as determined by RT-qPCR (Figure 3C) and an increase in the repressive histone modification H3K9me2 (Figure S4A), recapitulating repression of gene expression seen in FRDA cells. This smaller reduction in the RNA levels in FXN-GAA-Luc cells compared to FRDA lymphoblastoid cells can be explained by low number of repeats (only ∼310) on the integrated FXN copy.
Next we investigated if R-loops are formed on expanded repeats of the integrated FXN copy in HEK293 cells using DIP analysis (Figure 3D). Similar to patient-derived FRDA lymphoblast cells, we observed 2–3-fold increase in the level of R-loops over expanded repeats region (amplicons B, C and D) in FXN-GAA-Luc cells. This suggests that R-loops are formed on transcribed expanded GAA repeats, independently of their genomic location. To confirm the specificity of this R-loop signal, we employed RNAi to knock down endogenous RNase H1 enzyme, which specifically degrades the RNA in RNA/DNA hybrids [24]. Following depletion of RNase H1 in HEK293 cells, we observed a significant increase in the R-loop signal, suggesting that endogenous RNase H1 can degrade R-loops formed over FXN gene in vivo (Figure 3E, F).
Next we wanted to test if RNase H1 over-expression can resolve R-loops formed on expanded GAA repeats. To this end we over-expressed Flag and RNase H1-Flag constructs in FXN-Luc and FXN-GAA-Luc cells. High level of RNase H1 over-expression was confirmed by RT-qPCR (Figure S4B) and western blot analysis (Figure 3G). Interestingly, following RNase H1 over-expression, we observed a reduction of R-loop signal over the GAA expansion (Figure 3H). In line with these observations, RNase H1 over-expression resulted in up-regulation of FXN transcription from the expanded allele (Figure S4C). This suggests that R-loops formed over expanded repeats in vivo can be resolved by over-expressed RNase H1 and removal of R-loops leads to increase in FXN gene expression.
Previous studies have demonstrated that expanded FXN GAA repeats are associated with increased levels of heterochromatin marks [4], [6], [7]. To investigate the relationship between R-loops and repressive heterochromatin marks formed on the expanded FXN allele, we employed histone methyltransferase inhibitor BIX-01294, previously shown to reduce the level of H3K9me2 over repeat regions [21]. Following BIX-01294 treatment, we observed a significant reduction in the levels of H3K9me2 chromatin mark (Figure 4A), similar to previous reports [21]. Significantly, following BIX-01294 treatment the level of R-loops over the expanded repeat region remained unchanged (Figure 4B). Similarly, reduction in H3K9me2 chromatin mark had no effect on FXN nascent RNA level (Figure 4C). These data suggest that H3K9me2 chromatin modification is not directly responsible for the FXN transcriptional repression and is likely to be a consequence of the reduced transcription or caused by R-loop formation.
To investigate the ability of R-loops to directly trigger FXN transcriptional repression, we took advantage of camptothecin (CPT), a specific inhibitor of DNA Topoisomerase I (Top1), an enzyme which relieves transcription-induced DNA supercoiling. Loss of Top1 enhances R-loop formation in yeast and human cells [25], [26]. Following CPT treatment, we observed an increase in R-loop formation over the expanded repeat region in FRDA cells while R-loop levels in FXN regions E–G remained unchanged (Figure 5A). This effect was consistent between different patient-derived cell lines (Figure S5A). We detected no effect on R-loop levels in control cells, demonstrating the specificity of CPT treatment to expanded repeats (Figure 5A and S5A). The ability of CPT to increase R-loop levels was not due to CPT-induced covalent links between Top1 and DNA, since Top1 knock-down also resulted in R-loop accumulation (Figure 5E, F).
Interestingly, increase in the level of R-loops coincided with increase in the amount of repressive H3K9me2 mark over FXN regions B–D surrounding the expansion in FRDA cells (Figure 5B and S5B). This also resulted in down-regulation of nascent FXN RNA in FRDA cells, but not in control cells, as observed in three independent control and FRDA cell lines (Figure 5C and S5C–D). These data suggest that R-loops at expanded repeats directly trigger FXN transcriptional repression and promote formation of repressive H3K9me2 marks. To understand the molecular mechanism of this process, we investigated the binding of G9a histone methyltransferase, which deposits H3K9me2 marks on histones, to FXN gene. Interestingly, we observed that G9a is enriched over the expanded region in FRDA cells (Figure 5D). This suggests that R-loops may recruit G9a to the expanded repeat regions, thereby promoting the formation of repressive H3K9me2 marks.
To test if R-loop formation is a general feature of trinucleotide expansion diseases, we also examined the FMR1 gene. In Fragile X syndrome patients, the FMR1 allele containing a (CGG)n>200 expansion in the 5′UTR is fully methylated and transcriptionally silenced [27]. Therefore to investigate the potential role of R-loops in FXS, FMR1 transcription was reactivated by treatment with the DNA methylation inhibitor 5-aza-2′-deoxycytidine (5-azadC). This resulted in expression of FMR1 mRNA in FXS cells to 25% of control cells, as previously reported [28]. However, FMR1 expression was unchanged in control cells (Figure 6B). Using DIP, we detected low R-loop signal in control and untreated FXS cells (Figure 6C). Following 5-azadC treatment, we observed ∼4-fold increase in R-loops over the exon 1 region upstream of the expansion in FXS cells, while no significant changes were detected in control cells. The specificity of the DIP signal was confirmed by RNase H treatment (Figure 6D). These data suggest that R-loops are transcriptionally-dependent and localise to the expanded (CGG) repeat region. Since inhibition of DNA methylation only partially reactivates expanded FMR1 allele in FXS cells, it is possible that R-loops at expanded (CGG) repeats prevent full reactivation.
To further characterize the relationship between R-loops and FMR1 expression, we carried out kinetic experiments. In particular, we studied R-loop and FMR1 mRNA levels during the process of transcriptional re-activation with 5-azadC treatment (7 days) followed by 5-azadC wash out with drug-free media for 28 days (Figure 6E, F). We observed that the R-loop levels over the exon 1 of FMR1 gene stayed at the background during activation and wash-out period in control cells (dotted line in Figure 6E). In FXS cells, the R-loops were at their peak during the re-activation procedure with 5-azadC on day 7. After removal of 5-azadC, R-loop levels gradually diminished and completely disappeared after 7 days (day 14 of the full experiment). This pattern of R-loop dynamics correlated with FMR1 expression profile (Figure 6F), suggesting that R-loops are associated with FMR1 gene regulation.
We demonstrate that R-loops are formed over endogenous expanded (GAA) and (CGG) repeats in vivo, associated with FRDA and FXS disorders, respectively (Figure 1, 6). We show that these R-loops interfere with nascent Pol II transcription on FXN gene (Figure 2E). We also demonstrate that R-loops can trigger gene silencing irrespectively of their genomic location (Figure 3). R-loops over expanded repeats are very stable in human cells (Figure 2B), possibly due to failure of their complete turn-over by endogenous enzymes, which may contribute to FRDA pathology. Interestingly, expansion-associated R-loops can be resolved by over-expressed exogenous RNase H1, which leads to transcription up-regulation of FXN expression in vivo (Figure 3).
Previous work has demonstrated that co-transcriptionally formed RNA/DNA hybrids mediate transcription elongation impairment in vitro and in yeast S.cerevisiae [29], [30], suggesting that R-loops may provide roadblocks for RNA polymerases. R-loops over expanded repeats may form a structural block, directly interfering with Pol II transcription elongation. Similar to R-loops at the 3′ends of human genes [16], expansion-associated R-loops could promote RNA Pol II termination, resulting in reduction of active Pol II downstream of the expansion, as detected in this study.
Recently it was suggested that repressive chromatin H3K9me2 modification was not directly responsible for the FXN transcriptional repression [21]. In line with this, reversal of repressive DNA methylation on FMR1 gene was not sufficient to fully restore FMR1 expression [28]. We now show that a decrease in the level of the repressive H3K9me2 chromatin mark does not result in decrease of R-loops on the expanded allele or up-regulation of FXN RNA (Figure 4). These data indicate that R-loop formation is an early event in the process of FXN transcriptional gene silencing, which happens prior to the appearance of heterochromatin marks. We also show that increasing R-loop levels lead to an increase in repressive chromatin marks and subsequent repression of FXN gene expression (Figure 5). Furthermore, we observed that recruitment of G9a methyltransferase is enhanced on expanded FXN allele (Figure 5), providing an interesting possibility that R-loops may directly recruit this enzyme to promote H3K9me2 histone mark deposition. Altogether our results suggest that R-loops act as the primary trigger for repression of expanded FXN and FMR1 alleles which may in turn act to promote heterochromatin formation. Consistent with our data, recently it has been demonstrated that promoter-bound trinucleotide repeat-containing mRNA induces epigenetic silencing in Fragile X syndrome [31]. Indeed R-loops have been implicated in the formation and maintenance of heterochromatin at centromeres in S. pombe [32]. In line with this, H3K9me2 modification is also enriched at the R-loop-containing pause region of β-actin gene (Figure S6), essential for Pol II transcriptional termination [16]. This suggests that R-loops may promote H3K9me2 modification at the 3′end of this gene, similar to expanded repeats of FXN gene. We propose that R-loops may be a common feature of many trinucleotide expansion diseases, contributing to their pathology in vivo (Figure 7). The ability of R-loops to trigger transcriptional silencing in trinucleotide expansion diseases makes them an attractive target for future therapeutic approaches to treat these devastating diseases.
In addition to uncovering the molecular mechanisms underlying FRDA and Fragile X pathologies, our work also provides interesting implications for R-loop biology. Taken into consideration this work and the work of others in the field, depending on their genomic location, R-loops may have different functions (reviewed in [11]). Therefore, stable R-loops formed over expanded triplet repeats (this study) may be different from R-loops at the 3′-ends of genes [16], [33] and R-loops formed over CpG island promoters [15], [33]. At promoters, R-loops play a protective role against epigenetic silencing. By contrast, R-loops over FXN expanded repeats correlate with a reduction in transcription elongation and the enrichment of repressive chromatin marks. This suggests that R-loops may be ‘sensed’ differently, depending on their genomic location and sequence composition. Understanding the molecular mechanisms that cells use to distinguish ‘harmful’ from ‘useful’ R-loops is an important biological question in the study of human diseases.
Epstein-Barr virus (EBV)-transformed lymphoblastoid cell lines from healthy (GM15851, GM14926, GM06895), FRDA (GM15850, GM16243, GM16209) and FXS (GM03200) patients were obtained from Coriell Institute for Medical Research. Frataxin (GAA) repeat sizes were 650/1030 (GM15850), 800/800 (GM16209) and 670/1170 (GM16243). (CGG) repeat size in FMR1 gene was 530 (GM03200). All experiments were performed with early passages of the cell lines. Lymphoblastoid cell lines were grown in RPMI 1640 medium supplemented with 15% fetal bovine serum (FBS), 100 U/ml penicillin and 100 µg/ml streptomycin at 37°C in 5% CO2. 1 µM 5-azadC (Sigma, A3656) was added to the media for 7 days. 4 µM BIX-01294 (Sigma, B9311) was added to the media for the total of 72 hours. The media was replenished every 24 hours by replacement of half of the conditioned media with fresh media and drug. 10 µM camptothecin (Sigma, C9911) was added for 6 hours. Actinomycin D (Sigma, A9415) was added to a final concentration of 5 µg/ml to the media for 6–48 hours. FXN-Luc and FXN-GAA-Luc HEK293 cells were described in [23] and cultured in DMEM medium supplemented with 10% FBS, 100 U/ml penicillin/streptomycin, 100 µg/ml hygromycin B (Life Technologies). For 5-azadC wash-out experiments, cells were treated with 1 µM 5-azadC for 7 days. On day 7, cell were washed twice with fresh RPMI 1640 medium and cultured in the absence of 5-azadC during the indicated time.
ChIP analysis on endogenous genes was carried out as previously described [34], [35]. 5 µg of the following antibodies were used: Pol II antibody (Santa Cruz, H-224), H3 (AbCam, ab1791), H3K9me2 (AbCam, ab1220), G9a (AbCam, ab40542). The immuno-precipitated DNA was used as template for real-time quantitative PCR performed using a Rotor-Gene RG-3000 machine (Corbett Research). The PCR mixture contained QuantiTect SYBR green PCR master mix (Qiagen), 2 µl of the template DNA and corresponding primers from Table S1. Cycling parameters were 95°C for 15 min, followed by 45 cycles of 94°C for 20 s, 58°–62°C for 20 s, and 72°C for 20 s. Fluorescence intensities were plotted against the number of cycles by using an algorithm provided by the manufacturer. Amount of immuno-precipitated protein at a particular gene region was calculated as ‘% of Input’ after subtracting the background signal, as determined by the ‘no antibody’ control.
20 ng of genomic DNA was used as template in a 25 µl PCR reaction, containing 2.5 U Biotaq DNA polymerase, 3 mM MgCl2, 0.4 mM dNTPs, 0.4 µM GAA104F primer, 0.4 µM GAA629R primer in 1× NH4 reaction buffer. Products were amplified using protocol from [36] with minor modifications. In particular, 5 min at 94°C were followed by 10 cycles of 94°C for 20 s, 65°C for 30 s, 72°C for 5 min. This was followed by 20 cycles of 94°C for 20 s, 65°C for 30 s and 72°C for 5 min, with the 72°C step becoming 20 s longer in each cycle. After a final step at 72°C for 10 min, PCR products were resolved on a 1% agarose gel.
DNA immuno-precipitation (DIP) analysis on endogenous genes was performed with antibody, recognising RNA/DNA hybrids, purified from S9.6 hybridoma cell lines [37], as described in [16]. In particular, lymphoblastoid and HEK293 cells were split one day before DIP. 10×106 cells were harvested, washed in PBS and incubated in cell lysis buffer (85 mM KCl, 5 mM PIPES pH 8.0, 0.5% NP-40) for 10 min on ice. Nuclei were collected by centrifugation and then incubated in nuclei lysis buffer (50 mM TRIS pH 8.0, 5 mM EDTA, 1% SDS) on ice. Proteins were digested by incubation with proteinase K (Roche) for 3 h at 55°C. Proteins and cell debris were removed by centrifugation after addition of KOAc to the final concentration of 1 M. Genomic DNA containing R-loops was then precipitated by addition of isopropanol. After washing the DNA pellet with 70% EtOH, genomic DNA was resuspended in 400 µl IP dilution buffer (16.7 mM TRIS pH 8.0, 1.2 mM EDTA, 167 mM NaCl, 0.01% SDS, 1.1% Triton X-100) and used for sonication (Diagenode Bioruptor). Bioruptor (Diagenode) settings were 3 min sonication, at the Medium setting 30 sec on/30 sec off interval and the average size of the fragments was ∼500 nt. Sonicated genomic DNA was then pre-cleared with 50 µl protein A agarose beads (Millipore) in 3 ml IP dilution buffer including protease inhibitors (0.5 mM PMSF, 0.8 µg/ml pepstatin A, 1 µg/ml leupeptin) for 1 h at 4°C. 10 µl of S9.6 antibody was added to DNA corresponding to 10×106 cells. Immuno-precipitation was carried out over night at 4°C. Subsequent washes and elution steps are identical to the procedure as described for ChIP.
The immuno-precipitated, non-precipitated, and input DNAs were used as templates for qPCR. DIP RNase H-sensitivity analysis was carried out following the genomic isolation and prior to immuno-precipitation step with the addition of 25 U of RNase H (NEB, M0297S). 100 µl nuclease digestion reaction contained 1× reaction buffer, and it was performed for 6 hours at 37°C. Amount of immuno-precipitated RNA/DNA hybrid at a particular gene region was calculated as ‘% of Input’ after subtracting the background signal, as determined by the ‘no antibody’ control. In the case of FXN gene, all the values are relative to the FXN B amplicon in control cells.
Total RNA was harvested using TRIZOL reagent (Invitrogen) followed by DNase I treatment (Roche). 1–2 µg of total RNA was reverse-transcribed using SuperScript Reverse Transcriptase III (Invitrogen) with random hexamers (Invitrogen), oligodT primer (for polyA+ RNA) or gene-specific reverse primer (Table S1). The qPCR primers for amplification of polyA+ RNAs were the following: β-actin (ex5F/ex6R), GAPDH (F/R3), γ-actin (γ-actin spliced F/R), FXN (ex3F/ex4R), FMR1 (ex14 F/ex 15R). For analysis of nascent RNA in Figures 2A and 4C FXN primer FXN D was used, while in Figure 3C FXN primer B was used. For quantitative real-time PCR, 2 µl of cDNA was analyzed using a Rotor-Gene RG-3000 real-time PCR machine (Corbett Research) with QuantiTect SYBR green (Qiagen). For analysis of nascent FXN-Luc RNA in HEK293 cells, lucR primer was used for reverse transcription, and in4F and ex5R were used for qPCR.
The RNAi was carried out as described [38]. Control siRNA duplex was 5′-UAGCGACUAAACACAUCAA -3′ (Thermo Scientific siGENOME Non-Targeting siRNA #1D-001210-01-20), RNase H1 siRNA duplex was s48357 (Ambion). mRNA target sequence for Topoisomerase I siRNA duplex was 5′-GGACUCCAUCAGAUACUAU -3′. Total protein extracts were harvested using RIPA buffer. 20 and 40 µg of total protein extracts were resolved on SDS-PAGE and detected by Western blotting. Western blots were probed with Topoisomerase I (AbCam, ab109374), actin (Sigma, A2066), RNase H1 (AbCam, ab83179) antibodies.
FXN-Luc and FXN-GAA-Luc cells were freshly split into 10 cm dishes and transfected on the following day with 10 µg Flag or RNase H1-Flag plasmids using TransFectin reagent (BioRad), following the manufacturer's instructions. Cells were harvested 48 hours after transfection. RNaseH1-Flag plasmid was cloned by replacing the GFP tag in the RNaseH1-GFP plasmid, provided by Prof.R.J. Crouch, with the Flag tag using RNaseH1-FLAG(F) and RNaseH1-FLAG(R) primers.
The Br-UTP NRO analysis was carried as described in [16]. The equivalent of 8×106 nuclei from lymphoblastoid cells were used for each Br-UTP NRO reaction.
Unless otherwise stated, the figures present the average values of at least three independent experiments +/− SEM. Asterisks (*) indicate statistical significance (* p<0.05; ** p<0.01; *** p<0.001), based on unpaired, two-tailed distribution Student's t test.
|
10.1371/journal.ppat.1004793 | Ubiquitous Promoter-Localization of Essential Virulence Regulators in Francisella tularensis | Francisella tularensis is a Gram-negative bacterium whose ability to replicate within macrophages and cause disease is strictly dependent upon the coordinate activities of three transcription regulators called MglA, SspA, and PigR. MglA and SspA form a complex that associates with RNA polymerase (RNAP), whereas PigR is a putative DNA-binding protein that functions by contacting the MglA-SspA complex. Most transcription activators that bind the DNA are thought to occupy only those promoters whose activities they regulate. Here we show using chromatin immunoprecipitation coupled with high-throughput DNA sequencing (ChIP-Seq) that PigR, MglA, and SspA are found at virtually all promoters in F. tularensis and not just those of regulated genes. Furthermore, we find that the ability of PigR to associate with promoters is dependent upon the presence of MglA, suggesting that interaction with the RNAP-associated MglA-SspA complex is what directs PigR to promoters in F. tularensis. Finally, we present evidence that the ability of PigR (and thus MglA and SspA) to positively control the expression of genes is dictated by a specific 7 base pair sequence element that is present in the promoters of regulated genes. The three principal regulators of virulence gene expression in F. tularensis therefore function in a non-classical manner with PigR interacting with the RNAP-associated MglA-SspA complex at the majority of promoters but only activating transcription from those that contain a specific sequence element. Our findings reveal how transcription factors can exert regulatory effects at a restricted set of promoters despite being associated with most or all. This distinction between occupancy and regulatory effect uncovered by our data may be relevant to the study of RNAP-associated transcription regulators in other pathogenic bacteria.
| Most transcription regulators are found only at those promoters they control. Here we show that the most prominent regulators of virulence gene expression in Francisella tularensis are found ubiquitously at promoters including those they do control and those they do not. Furthermore, we present evidence that these regulators—the RNA polymerase-associated SspA family members MglA and SspA, and the putative DNA-binding protein PigR—exert their coordinate regulatory effects only at promoters that contain a small DNA sequence element. Our findings reveal how transcription factors can associate with many promoters but only exert regulatory effects at a few. They also have implications for how SspA family members and other RNAP-associated transcription regulators might exert their effects in other pathogens.
| Francisella tularensis is a Gram-negative bacterium and the aetiological agent of tularemia, a disease that can be fatal in humans [1]. This pathogen is highly infectious, with as few as 10 organisms constituting an infectious dose, and is a potential bioweapon [2]. The ability of F. tularensis to cause disease is dependent principally upon its ability to grow within macrophages [1,3–5]. Prominent amongst those genes that are essential for the intramacrophage growth and survival of F. tularensis are those located on the Francisella pathogenicity island (FPI), which are thought to have been acquired through horizontal transfer [6–8]. Genes on the FPI encode a type VI secretion system that may secrete effector proteins into cells of the host [9,10], thereby enabling the organism to escape the so-called Francisella-containing vacuole and to replicate freely within the macrophage cytosol [4,11,12].
Expression of the genes on the FPI is dependent upon the coordinate activities of three regulators [13–18]. Two of these, MglA and SspA, belong to the stringent starvation protein A (SspA) family of proteins and form a heteromeric complex that associates with RNA polymerase (RNAP) [13,15,19]. The other is a putative DNA-binding protein called PigR (also known as FevR in F. novicida) that works in concert with MglA and SspA by contacting the RNAP-associated MglA-SspA complex directly [16–18]. MglA, SspA, and PigR also positively control the expression of virulence genes outside the FPI, and are thought to control the expression of ~100 genes in total, including many whose roles in virulence are not yet known [14–17]. The findings that MglA, SspA, and PigR are essential for intramacrophage growth and for virulence underscores the indispensible roles these regulators play in the coordinate control of virulence gene expression in F. tularensis [16,17,20].
According to the current view for how MglA, SspA, and PigR control the expression of virulence genes, PigR functions like a classical transcription activator, binding specifically to a DNA sequence present at the promoters of regulated genes; thus, contact between DNA-bound PigR and the RNAP-associated MglA-SspA complex would stabilize the binding of RNAP to those promoters that contain a PigR binding site [17,18]. However, it is not known whether the promoters of MglA/SspA/PigR-regulated genes contain a specific sequence element that confers responsiveness to PigR. If PigR were indeed to function like a classical transcription activator it would be expected to be located at only those promoters it regulates, and it would be predicted to bind to DNA recognition sites associated with target promoters regardless of whether or not the MglA-SspA complex were present in the cell. Indeed, most classical transcription activators are thought to bind specific sites on the DNA prior to interacting with RNAP [21,22]. Another prediction from the current model is that PigR interacts with the MglA-SspA complex that is associated with the RNAP holoenzyme (the form of the enzyme that contains the σ factor) during transcription initiation [17,18]. However, it is unknown whether the MglA-SspA complex is associated with the RNAP holoenzyme during transcription initiation or the RNAP core enzyme during transcription elongation, or both.
Using chromatin immunoprecipitation followed by high-throughput DNA sequencing (ChIP-Seq), we show that PigR, MglA, and SspA are present at virtually all detected promoters in F. tularensis. We also demonstrate that PigR requires MglA (and thus presumably the MglA-SspA complex) in order to specifically associate with promoters. Finally, we present evidence that the promoters of PigR-regulated genes contain a specific sequence motif that is both necessary and sufficient for PigR-mediated control. Our findings reveal that the most prominent regulators of virulence gene expression in F. tularensis are found at essentially all promoters but only positively control those that contain a specific sequence element.
In order to address the question of whether PigR specifically associates with the promoters of regulated genes, we first sought to define the locations of promoters on a genome-wide basis in F. tularensis. To do this, we determined the locations of the β′ subunit of RNA polymerase (RNAP) on the F. tularensis chromosome using ChIP-Seq. To immunoprecipitate the β′ subunit of RNAP, we constructed a strain of F. tularensis LVS in which the chromosomal copy of the rpoC gene was modified to encode β′ with a vesicular stomatitis virus-glycoprotein (VSV-G) epitope tag fused to its C-terminus (Fig 1A). This results in cells of LVS which synthesize the β′ subunit of RNAP with a VSV-G tag (β′-V) at native levels. Because β′ is a core subunit of RNAP, β′ is expected to be found at both promoter regions and within actively transcribed genes. Thus, in order to use β′-V to specifically identify the locations of promoters, we performed ChIP-Seq after treatment of the LVS β′-V cells with the RNAP inhibitor rifampicin (rif) to effectively trap RNAP at promoters [23,24] (Fig 1B). By determining the location of the β′ subunit of RNAP in cells treated with rif, we identified 526 promoter regions in F. tularensis LVS (S1 Table).
F. tularensis encodes two σ factors: σ70, the so-called housekeeping σ factor, and σ32, the so-called heat-shock σ factor [25]. As a complementary approach to identify promoters in F. tularensis, and to determine which promoters are σ70-dependent and which are controlled by σ32, we performed ChIP-Seq with cells that synthesized epitope-tagged versions of each σ factor. To do this we constructed a strain of LVS that synthesized σ70 with a VSV-G epitope tag fused to its C-terminus (LVS σ70-V), and another strain of LVS that synthesized σ32 with a VSV-G tag fused to its C-terminus (LVS σ32-V). As a control we also constructed a strain that synthesized HipB, a predicted site-specific DNA-binding protein [26], with a VSV-G tag fused to its C-terminus (LVS HipB-V). ChIP-Seq with cells of the LVS σ70-V strain identified 333 promoter regions, of which 277 (83.2%) overlap with the locations of promoters defined by determining the location of the β′ subunit of RNAP in cells grown in the presence of rif (Fig 1B, S2 Table). ChIP-Seq with cells of the LVS σ32-V strain identified only 4 promoter regions (Fig 1B, S3 Table). ChIP-Seq with cells of the LVS HipB-V strain revealed that HipB associates with 26 regions of the chromosome (S4 Table). By defining a promoter as a region with significant enrichment of σ70, σ32, or the β′ subunit of RNAP in cells grown in the presence of rif, we identified 581 promoter regions in F. tularensis LVS, 495 (85%) of which were intergenic and 86 (15%) of which were intragenic (S5 Table).
Having determined the locations of promoters in F. tularensis on a genome-wide basis, we next sought to determine at which promoters PigR, MglA, and SspA were located. To do this we utilized a previously constructed strain in which the native chromosomal copy of mglA is altered such that it specifies MglA with a TAP (tandem affinity purification) tag fused to its C-terminus [15]. We also constructed two additional strains of LVS: one in which the native chromosomal copy of pigR had been altered such that it specified PigR with a VSV-G epitope tag fused to its C-terminus (LVS PigR-V); and another in which the native chromosomal copy of sspA had been altered such that it specified SspA with a VSV-G epitope tag fused to its C-terminus (LVS SspA-V).
ChIP-Seq with cells of the LVS PigR-V strain, cells of the LVS MglA-TAP strain, and cells of the LVS SspA-V strain revealed that PigR, MglA, and SspA are located at the majority of promoters in F. tularensis and not just at the promoters of regulated genes (Fig 2, S6 Table). The finding that PigR, MglA, and SspA are found at the promoters of both regulated and non-regulated genes is illustrated in Fig 2A and 2B which show the occupancies of the β′ subunit of RNAP (in the presence of rif), σ70, MglA, SspA, PigR, and HipB at the FTL_0491, FTL_0650, and FTL_0651 promoter regions as determined by ChIP-Seq. Specifically, Fig 2A shows that PigR, MglA, and SspA are found at the promoter of the FTL_0491 gene, which is an example of a gene that is positively regulated by MglA, SspA, and PigR [14,16] (see also S6 Table), whereas Fig 2B shows that PigR, MglA, and SspA are found at the promoters of the FTL_0650 and FTL_0651 genes, which are examples of genes that are known not to be positively regulated by MglA, SspA, and PigR [14–17]. HipB was not detected at any of these promoters by ChIP-Seq (Fig 2A and 2B) indicating the specificity of the observed associations of PigR, MglA, and SspA with these promoters. In contrast, HipB is specifically enriched upstream of the hipB gene (S1 Fig), suggesting that in F. tularensis HipB may control its own expression.
The finding that PigR, MglA, and SspA are found at the majority of promoters in F. tularensis is illustrated in Fig 2C which shows the locations and degrees of occupancy of the β′ subunit of RNAP (in cells grown in the presence of rif), σ70, MglA, SspA, PigR, and HipB over a representative 400 kb region of the F. tularensis chromosome. Comparison between the regions enriched for σ70, MglA, SspA, and PigR, together with the relative degree of enrichment, revealed a striking correspondence between the four (Fig 2C). Note that the degree of occupancy of the β′ subunit of RNAP at a particular promoter can differ in the presence and absence of rifampicin (Fig 1B) [27], which may explain why the ChIP-Seq enrichment profile for β′ in cells grown in the presence of rif differs slightly from that of σ70 in certain locations (Fig 2C). The concordance among the localization of σ70, MglA, SspA, and PigR across the entire F. tularensis chromosome is demonstrated in Fig 2D which represents the 98% of promoter regions identified by ChIP-Seq of σ70 at which at least one of the three factors, MglA, SspA, or PigR, is found; the Venn diagram shows that MglA, SspA, and PigR are found at the majority of promoters identified by detection of σ70 (Fig 2D). The identification of PigR at the majority of promoters suggests that PigR is not a regulator that is only found at the promoters of specific target genes.
ChIP-Seq with cells of the LVS PigR-V strain, cells of the LVS MglA-TAP strain, and cells of the LVS SspA-V strain also revealed that PigR, MglA, and SspA are present at promoters together with σ70 and are not detected in transcribed regions. This is in contrast to the situation with the β′ subunit of RNAP, which is found both at promoters and within transcribed regions in cells grown in the absence of rif (Fig 1B). These findings suggest that PigR, MglA, and SspA might not be components of the transcription elongation complex and that PigR, MglA, and SspA likely exert their regulatory effects at the level of transcription initiation.
Interaction between PigR and the RNAP-associated MglA-SspA complex is required in order for PigR to function coordinately with MglA and SspA [18]. We therefore next asked whether PigR requires the MglA-SspA complex in order to associate with promoter regions in F. tularensis. Because the expression of pigR is dependent upon the presence of MglA [16,17], in order to address this question we performed ChIP-Seq with cells of a ∆pigR mutant strain and cells of a ∆pigR ∆mglA mutant strain that ectopically synthesized similar amounts of plasmid-encoded PigR-V. We found that when supplied from plasmid pF under the control of the strong heterologous groES promoter, PigR-V was significantly less abundant in cells of the ∆pigR ∆mglA mutant strain than in cells of the ∆pigR mutant strain (Fig 3A). The groES promoter is not positively controlled by MglA, so it is possible that the PigR-V protein is less abundant in cells of a ∆mglA mutant strain because it is less stable in the absence of the MglA-SspA complex. Therefore, to be able to compare cells containing similar amounts of PigR, we used the strong groES promoter on plasmid pF to drive the synthesis of PigR-V in cells of the ∆pigR ∆mglA mutant strain and a weakened groES promoter lacking an UP-element on plasmid pF2 to drive the synthesis of PigR-V in cells of the ∆pigR mutant strain [17] (Figs 3A and S2).
Comparison of the ChIP-Seq results obtained with ectopically produced PigR with those obtained with native PigR revealed that the ectopic synthesis of PigR does not significantly alter the genome-wide locations of this protein. This is illustrated at the promoter for the PigR/MglA/SspA-regulated FTL_0491 gene (Fig 3B), and illustrated at the promoters for the FTL_0650 and FTL_0651 genes, which are not PigR/MglA/SspA-regulated (Fig 3C). Comparison of the ChIP-Seq results obtained with ectopically produced PigR in the presence and absence of MglA revealed a striking difference; we found no specific enrichment of PigR at any promoter in the absence of MglA, or at any other region of the chromosome. This is illustrated at the FTL_0491, FTL_0650 and FTL_0651 promoters in Fig 3B and 3C. These findings indicate that MglA, and by inference the MglA-SspA complex, is required for PigR to specifically associate with promoter regions in F. tularensis.
Although PigR (together with MglA and SspA) is present at the majority of promoters, it appears to only positively regulate the expression of a fraction of the corresponding genes. We therefore reasoned that PigR might function as an activator at specific promoters through recognition of a specific sequence element.
To search for a conserved sequence motif in the promoters of genes that are regulated by PigR we first tested whether certain genes previously shown to be regulated by MglA and SspA were also regulated by PigR in F. tularensis. To do this we quantified specific candidate transcripts in both wild-type LVS cells and in cells of a LVS ∆pigR mutant strain using Nanostring (see Materials and Methods; S7 Table). Using MEME [28], we then searched for a specific motif in the promoter regions of genes that (i) were either previously shown to be positively regulated by PigR in LVS by DNA microarray [17], or shown to be positively regulated by PigR in our Nanostring assays (S7 Table), or both, and (ii) contained a region of PigR enrichment upstream from the translation start site, as determined by our ChIP-Seq studies with cells of our LVS PigR-V strain. Eleven genes fit these criteria and a 7 bp motif was found to be present in all 11 of the putative promoter regions analyzed. A logo representing this 7 bp motif, which we have named the PigR response element (PRE) is depicted in Fig 4A.
We next asked whether the PRE was found at a specific location relative to the transcription start site of a regulated promoter. To do this we first determined transcription start sites on a genome-wide basis using RNA-Seq [29,30]. This gave us 453 candidate transcription start sites (see Materials and Methods). This list was parsed further to include only those start sites found within 1 kb upstream of the translational start site of an annotated ORF, and remove from consideration those start sites associated with rRNAs and tRNAs, and those found within repeated sequences annotated as encoding transposases. To obtain a list of transcription start sites that could be independently verified as originating from a detectable promoter we further parsed this list of 197 start sites to include only those found within a region of enrichment for σ70, σ32, or the β′ subunit of RNAP (in cells grown in the presence of rif) as determined by ChIP-Seq. This gave us transcription start sites with high confidence for 110 promoters, including 3 of the 11 putative promoter regions used to initially identify the PRE through MEME (S8 Table). We then used primer extension to determine transcription start sites for 2 additional promoters of PigR-regulated genes used in the MEME analysis. Through determining the transcription start sites for the promoters driving the expression of 5 independent PigR-regulated genes we were able to infer the sequences and locations of putative -10 and -35 elements for each of these promoters and found that the PRE was either 6 or 7 bp upstream from the predicted -35 element in each case (Fig 4B). Analysis of the 110 promoters from our high quality data set revealed that only 3 of these contained a PRE (see Materials and Methods) and are known to be PigR-regulated, whereas 107 do not contain a PRE and are not known to be PigR-regulated [17] (Fig 4B, S7 and S8 Tables). This suggests that the presence and location (6 or 7 bp upstream of the putative -35 element) of the PRE is specific to PigR-regulated promoters, raising the possibility that PigR may bind directly to this site to activate transcription from those promoters that contain it.
Having identified a specific conserved sequence element in the same location in the promoters of PigR-regulated genes we sought to determine whether this sequence rendered a particular promoter responsive to PigR. To do this we first constructed a reporter strain of LVS in which one of the two copies of the PigR/MglA/SspA-regulated iglA promoter is transcriptionally fused to lacZ using a chromosomal integration vector [31]. We also made three additional reporter strains of LVS. Two of these contained different mutations at conserved base pairs in the PRE of the iglA promoter-lacZ fusion (Fig 4C), whereas the third reporter strain contained mutations in the predicted -10 element of the iglA promoter-lacZ fusion that would be predicted to abolish promoter activity (Fig 4C) [32]. Finally, we made an additional three reporter strains in cells of the LVS ∆pigR mutant strain that contained the wild-type version, a PRE mutant version, or the -10 mutant version of the iglA promoter-lacZ fusion.
The results depicted in Fig 4D show that mutations in the PRE of the iglA promoter reduce expression of the linked lacZ reporter gene only when PigR is present (i.e. in cells of LVS but not in cells of the LVS ∆pigR mutant strain). Consistent with the idea that these differences are due to a decrease in the activity of the iglA promoter, cells of the reporter strains containing mutations in the -10 element that are predicted to decrease the activity of the promoter exhibit dramatically reduced lacZ expression (Fig 4D). Note that there are two identical copies of the iglA gene in LVS because there are two copies of the FPI in this organism. Only reporter strains carrying wild-type and mutant versions of the iglA promoter-lacZ fusion integrated at the FTL_0111 locus were used in these experiments, ruling out the possibility that any of the observed differences in lacZ expression were due to differences in the location of the reporter in the different strains. Taken together, these findings suggest that residues within the PRE of the iglA promoter are important in order for PigR to exert a positive effect on expression of the iglA gene.
Having established that conserved base pairs within the PRE are important for expression of a PigR-regulated gene we next asked whether the PRE was sufficient to confer control on a promoter that did not ordinarily contain a PRE. To do this we introduced 3 mutations into the FTL_0361 promoter that generated a consensus PRE 6 bp upstream of the putative -35 element (Fig 4E). We then made reporter strains of LVS and the LVS ∆pigR mutant strain that contained the wild-type version, a PRE-containing version, or a -10 mutant version of a FTL_0361 promoter-lacZ fusion. The results depicted in Fig 4F show that addition of a PRE to the FTL_0361 promoter results in an increase in expression of the FTL_0361 promoter-lacZ fusion only in the presence of PigR (i.e. in cells of the LVS wild-type strain but not in cells of the LVS ∆pigR mutant strain). These findings demonstrate that the PRE is sufficient to confer on a promoter the ability to respond to PigR, and by inference, the ability to respond to MglA and SspA.
Using ChIP-Seq we have found that PigR, MglA, and SspA are found at the majority of promoters in F. tularensis. We have also found that PigR requires the MglA-SspA complex in order to specifically localize to promoter regions. We infer from this that interaction between PigR and the RNAP-associated MglA-SspA complex directs PigR specifically to promoter regions. Despite their ubiquitous presence at promoters, PigR, MglA, and SspA coordinately control the expression of approximately 5% of known genes and we have uncovered the molecular basis for this specificity. In particular, we have identified a 7 bp sequence element that we have called the PRE (the PigR response element), located approximately 6 bps upstream of the putative -35 element of promoters that are positively regulated by PigR/MglA/SspA. The PRE is both necessary and sufficient to confer control by PigR. Our findings indicate that although PigR, MglA, and SspA are present at essentially all promoters, they control the activities of only those promoters that contain a specific sequence element.
Finding PigR, MglA, and SspA at promoters but not within transcribed regions suggests that these proteins are associated with the RNAP holoenzyme and likely exert their regulatory effects at the level of transcription initiation. Consistent with the idea that the MglA-SspA complex interacts with the σ70-containing RNAP holoenzyme, σ70 together with the core subunits of RNAP and SspA were found to co-purify with MglA in LVS in stoichiometric amounts [15].
PigR contains a putative helix-turn-helix motif, suggesting it might exert its regulatory effects through interaction with the DNA [16,17]. Based on our findings that PigR, MglA, and SspA are present at the majority of σ70-dependent promoters in F. tularensis, together with our identification of the PRE, we propose a model for how PigR works in concert with the MglA-SspA complex to positively regulate the expression of a specific set of genes, including many that are required for virulence (Fig 5). According to this model, PigR is a transcription activator that associates with all promoters through its interaction with the RNAP-associated MglA-SspA complex. However, only at those promoters that contain a PRE does PigR make sufficiently strong contact with the DNA to further stabilize the binding of RNAP and activate transcription. In essence, the model specifies that PigR is an RNAP-associated transcription activator that functions by providing RNAP with an additional DNA-binding domain, conferring on RNAP the ability to form especially stable complexes at promoters that contain a PRE. Note that in this model, PigR/MglA/SspA-regulated promoters are depicted as being recognized by RNAP holoenzyme containing σ70 (i.e. are σ70-dependent promoters), since our ChIP-Seq studies reveal PigR, MglA, SspA, and σ70 are present at many of the same promoter regions. Note also that our model explains only how PigR, together with the MglA-SspA complex, exerts positive effects on gene expression; the small number of genes that are negatively regulated by PigR/MglA/SspA [15–17], may be controlled directly or indirectly by these factors. It is possible that some genes are regulated by PigR/MglA/SspA because they are subject to control by another regulator that is in turn regulated by PigR/MglA/SspA. However, to the best of our knowledge, pigR is the only gene encoding a putative DNA-binding protein that is positively regulated by PigR/MglA/SspA [15–17].
Our model specifies that PigR is a DNA-binding protein that associates with RNAP via the MglA-SspA complex prior to promoter binding, and is therefore associated with all promoters, as is supported by our ChIP-Seq data. According to the classical view, transcription activators that bind the DNA and contact RNAP are found only at the promoters of regulated genes and function by first recognizing their respective target sites on the DNA and then, once tethered to the DNA, by contacting RNAP [21,22]. However, there is at least one precedent in the literature for a regulator that appears to manifest ubiquitous promoter localization, while exerting effects at only a subset of promoters. In particular, CarD is an essential RNAP-associated transcription regulator in Mycobacterium tuberculosis that has been found to associate with the majority of promoters [33,34]. Although the ability of CarD to bind the DNA is necessary in order for this regulator to stimulate transcription initiation, it is not yet known whether CarD exerts its regulatory effects at promoters through recognition of a specific sequence element [34]. In addition, in E. coli, members of the MarA family of transcription activators are thought to associate with RNAP prior to DNA-binding and to contact their DNA target sites as a pre-assembled activator-RNAP complex [35–37]. However, to the best of our knowledge, genome-wide location analyses have not been performed on members of the MarA family, and so it is not yet known whether these regulators are found at the majority of promoters in E. coli.
We note that although PigR is predicted to be a DNA-binding protein, PigR has yet to be shown to be capable of binding the DNA, and need not necessarily exert its regulatory effects through direct interaction with the PRE. It is formally possible that interaction between PigR and the RNAP-associated MglA-SspA complex may enable some other portion of RNAP, such as the C-terminal domain of one of the α subunits [15,38], or perhaps the MglA-SspA complex itself, to interact productively with the PRE, resulting in transcription activation. However, in relation to the latter possibility, SspA family members do not contain any obvious DNA-binding determinants and have not been shown to bind the DNA directly [39,40]. It is important to note that although we found PigR does not specifically associate with the DNA in the absence of MglA, this does not mean that PigR is not a site-specific DNA-binding protein. PigR may need to interact with the MglA-SspA complex in order to specifically interact with the PRE, either because the protein-protein interaction promotes a structural change within PigR that is essential for DNA-binding, or because interaction between PigR and the DNA is too weak to be able to occur in the absence of stabilizing interactions provided by promoter-bound RNAP. Indeed, most sequence-specific transcription regulators bind as dimers to their cognate recognition sites, which are typically 20 bp in length. If PigR does bind the 7 bp PRE directly, this would be more reminiscent of a monomer of a dimeric regulator binding a half-site. Regardless of whether or not the ability of PigR to interact directly with the PRE is essential in order for PigR to exert its regulatory effects, our findings clearly establish the PRE as the sequence element that renders a promoter subject to control by PigR, and thus presumably MglA and SspA as well.
The virulence genes present on the horizontally acquired FPI are the ones that are most strongly regulated by PigR, MglA, and SspA, and it is tempting to speculate that the limited size of the PRE (at 7 bp) may have facilitated the expansion of the PigR/MglA/SspA regulatory network to include these. Only three mutations were required in order to generate a consensus PRE within the FTL_0361 promoter (Fig 4E). More than three mutations would likely have been required had the PRE been closer to 20 as opposed to 7 bp. The relatively short length of the PRE means that relatively few changes would be required to place a particular promoter under the control of PigR/MglA/SspA, including any promoter that might have been acquired from a foreign source through horizontal transfer.
Our ChIP-Seq studies suggest that PigR interacts with the MglA-SspA complex at the majority of promoters. This raises the possibility that through interaction with the MglA-SspA complex, PigR may modulate the activity of any other regulator that functions through interaction with the RNAP-bound MglA-SspA complex. Indeed, it has been suggested that PmrA, another important regulator of virulence gene expression in Francisella, might function through interaction with MglA and SspA [41]. It is therefore conceivable that at some promoters PigR may modulate the activity of PmrA, or vice versa, through competition for a binding surface on the MglA-SspA complex.
The role of the MglA-SspA complex in positively controlling the expression of virulence genes appears to be to simply serve as a contact site on RNAP for PigR. Evidence for other SspA family members serving as contact sites for transcription activators comes from studies of bacteriophage P1 late gene expression; in E. coli, SspA evidently functions as a co-activator of P1 late gene expression by making simultaneous contact with RNAP and the phage-encoded sequence-specific DNA-binding protein Lpa [39]. However, in the case of Lpa, it is not known whether this regulator is associated with the majority of promoters in E. coli or just those driving expression of the P1 late genes. SspA family members have been shown to be important for the virulence of a variety of pathogens [42–47]. Serving as a contact site on RNAP for a transcription activator may represent a common mechanism by which SspA family members control the expression of virulence genes in numerous pathogens.
F. tularensis subsp. holarctica LVS and its derivatives were grown at 37°C in either Mueller Hinton (MH) broth (Difco), supplemented with glucose (0.1%), ferric pyrophosphate (0.025%), and Isovitalex (2%), or on cysteine heart agar (Difco) medium supplemented with 1% hemoglobin solution (VWR); when appropriate, kanamycin was used for selection at either 5 μg/ml or 10 μg/ml. Escherichia coli strain XL1-blue (Stratagene) was used for plasmid construction and, when appropriate, kanamycin was used to select for resistance at 50 μg/ml. E. coli containing plasmid pBSK iglA-lacZ, or its derivatives, were grown at 30°C.
A modified version of pEX18Kan (provided by Shite Sebastian and Simon Dillon, Harvard Medical School, Boston, Massachusetts, United States) was used as the vector for VSV-G tagging integration constructs. We have used pEX18Kan for deletion constructs [15,17]; it utilizes a ColE1 origin of replication, which is nonfunctional in LVS, and contains the Tn903 kanamycin resistance gene (Epicentre) driven by the LVS groES promoter. The plasmid pKL01 was generated by first amplifying the last 400 base pairs (bp) of the FTL_1743 locus (rpoC), minus the stop codon, by PCR. The 5′ primer contained DNA specifying a KpnI site upstream of the gene fragment. The 3’ primer included DNA containing a NotI site and one extra base pair, encoding a 3 amino acid alanine linker. The linker is followed by DNA specifying the 11 amino acid vesicular stomatitis virus-glycoprotein (VSV-G) epitope tag, followed by a stop codon and DNA specifying an EcoRI site. The corresponding PCR product was digested with KpnI and EcoRI and cloned into pEX18Kan that had been digested with KpnI and EcoRI, generating pKL01. We largely removed the sacB gene by digesting with MscI and EcoRV and re-ligating the vector together, resulting in pKL02.
VSV-G tagging integration constructs for rpoD, rpoH, sspA, pigR, and hipB were generated by amplifying the last 250–400 bp (depending on gene size) of the gene using a 5′ primer containing a KpnI site and a 3′ primer containing a NotI site, which allows each fragment to be fused with DNA specifying the alanine linker and VSV-G epitope tag. Fragments were subcloned into pKL02 that had been digested with KpnI and NotI. Plasmid pKL05 contains the DNA specifying the 3’ end of rpoD and was used to generate strain LVS σ70-V. Plasmid pKL04 contains the DNA specifying the 3’ end of rpoH and was used to generate strain LVS σ32-V. Plasmid pKL08 contains the DNA specifying the 3’ end of pigR and was used to generate strain LVS PigR-V. Plasmid pCS05 contains the DNA specifying the 3’ end of hipB and was used to generate strain LVS HipB-V. Plasmid pKL07 contains the DNA specifying the 3’ end of sspA. Because expression of the putative sspA operon could potentially be interrupted by plasmid integration, pKL07 was modified to contain an outward facing promoter after plasmid integration. To do this, another groES promoter was amplified from LVS genomic DNA and cloned upstream of the sspA gene fragment, into the BamHI and PstI sites, resulting in plasmid pKL13, which was used to generate strain LVS SspA-V.
The pBSK iglA-lacZ plasmid (provided by Thomas Kawula, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina, United States) utilizes a ColE1 origin of replication, which is nonfunctional in LVS, contains a kanamycin resistance determinate (aphA1), and contains a transcriptional fusion between the iglA promoter and the lacZ gene [31]. The pBSK iglA-lacZ plasmid contained two tandem PacI sites, so pMO1, which contains DNA specifying the wild-type iglA promoter with a single PacI site before the lacZ gene, was generated by digesting pBSK iglA-lacZ with PacI and NotI and recloning the wild-type iglA promoter fragment into the plasmid backbone. The pMO1 plasmid was used to generate LVS PiglA-lacZ and LVS ∆pigR PiglA-lacZ. Mutations in the PRE of the iglA promoter were generated using splicing by overlap extension PCR [48]. The corresponding PCR products were digested with PacI and NotI and cloned into pBSK iglA-lacZ that had been digested with PacI and NotI to replace the wild-type iglA promoter. The pMO2 plasmid contains DNA specifying the iglA promoter containing the PRE mutant 1 and was used to generate LVS PiglA-M1-lacZ and LVS ∆pigR PiglA-M1-lacZ. The pMO3 plasmid contains DNA specifying the iglA promoter containing the PRE mutant 2 and was used to generate LVS PiglA-M2-lacZ. The pMO4 plasmid contains DNA specifying the iglA promoter containing the -10 mutant and was used to generate LVS PiglA-10M-lacZ and LVS ∆pigR PiglA-10M-lacZ.
The promoter region of FTL_0361 was amplified from F. tularensis LVS genomic DNA using a 5′ primer containing DNA specifying a NotI site upstream of the promoter and the 3’ primer including DNA containing a PacI site downstream of the promoter. The resulting PCR product was digested with PacI and NotI and cloned into pBSK iglA-lacZ that had been digested with PacI and NotI to replace the iglA promoter, generating pMO5. The pMO5 plasmid was used to generate LVS PFTL_0361-lacZ and LVS ∆pigR PFTL_0361-lacZ. Plasmids containing mutations in the FTL_0361 promoter were generated in the same manner as plasmids containing mutations in the iglA promoter. The pMO6 plasmid contains DNA specifying the FTL_0361 promoter containing the PRE and was used to generate LVS PFTL_0361-PRE-lacZ and LVS ∆pigR PFTL_0361-PRE-lacZ. The pMO7 plasmid contains DNA specifying the FTL_0361 promoter containing the -10 mutant and was used to generate LVS PFTL_0361-10M-lacZ and LVS ∆pigR PFTL_0361-10M-lacZ.
Electroporation of integration plasmids into LVS was performed as described [49]. Cells in which a single homologous recombination event had occurred between the integration vector and the chromosome were selected on cysteine heart agar with 1% hemoglobin and either 5 μg/ml (for VSV-G tagging integration vectors) or 10 μg/ml kanamycin (for lacZ reporter integration vectors). Strains containing the correct integration were confirmed by colony PCR, by Western blotting and/or Southern blotting. Strain LVS βʹ-V, which synthesizes the βʹ subunit of RNAP with a C-terminal VSV-G tag, was generated by electroporation of plasmid pKL02 into LVS. Strain LVS σ70-V, which synthesizes the σ70 protein with a C-terminal VSV-G tag, was generated by electroporation of plasmid pKL05 into LVS. Strain LVS σ32-V, which synthesizes the σ32 protein with a C-terminal VSV-G tag, was generated by electroporation of plasmid pKL04 into LVS. Strain LVS SspA-V, which synthesizes the SspA protein with a C-terminal VSV-G tag, was generated by electroporation of plasmid pKL13 into LVS. Strain LVS PigR-V, which synthesizes the PigR protein with a C-terminal VSV-G tag, was generated by electroporation of plasmid pKL08 into LVS. Strain LVS HipB-V, which synthesizes the HipB protein with a C-terminal VSV-G tag, was generated by electroporation of plasmid pCS05 into LVS.
Strains containing the iglA-lacZ transcriptional fusion and derivatives were integrated at the FTL_0111 iglA locus, as determined by Southern blotting; the probe was synthesized using the PCR DIG Probe Synthesis Kit (Roche), hybridized to digested chromosomal DNA that had been transferred to a positively charged nylon membrane, and detected using CDP-Star (Roche). Strains PiglA-lacZ and LVS ∆pigR PiglA-lacZ, which contain lacZ under the control of the wild-type iglA promoter, were generated by electroporation of pMO1 into LVS and LVS ∆pigR, respectively. Strains LVS PiglA-M1-lacZ and LVS ∆pigR PiglA-M1-lacZ, which contain lacZ under the control of the iglA promoter containing the two mutations in the PRE (PRE mutant 1), were generated by electroporation of pMO2 into LVS and LVS ∆pigR, respectively. Strain LVS PiglA-M2-lacZ, which contain lacZ under the control of the iglA promoter containing three mutations in the PRE (PRE mutant 2), was generated by electroporation of pMO3 into LVS. Strains LVS PiglA-10M-lacZ and LVS ∆pigR PiglA-10M-lacZ, which contain lacZ under the control of the iglA promoter containing two mutations in the -10 element, were generated by electroporation of pMO4 into LVS and LVS ∆pigR, respectively. Strains LVS PFTL_0361-lacZ and LVS ∆pigR PFTL_0361-lacZ, which contain lacZ under the control of the FTL_0361 promoter, were generated by electroporation of pMO5 into LVS and LVS ∆pigR, respectively. Strains LVS PFTL_0361-PRE-lacZ and LVS ∆pigR PFTL_0361-PRE-lacZ, which contain lacZ under the control of the FTL_0361 promoter containing the PRE, were generated by electroporation of pMO6 into LVS and LVS ∆pigR, respectively. Strains LVS PFTL_0361-10M-lacZ and LVS ∆pigR PFTL_0361-10M-lacZ, which contain lacZ under the control of the FTL_0361 promoter containing mutations in the -10 element, were generated by electroporation of pMO7 into LVS and LVS ∆pigR, respectively.
Plasmids pF and pF-PigR-V have been described previously [17] and were used as a negative control vector and to drive ectopic expression of PigR with a C-terminal VSV-G tag, respectively. The pF-PigR-V plasmid contains DNA encoding the PigR protein fused to the VSV-G epitope, which is driven from the groES promoter; the pF plasmid does not contain the pigR gene or DNA encoding the VSV-G epitope tag. Plasmid pF2-PigR-V synthesizes PigR-V under the control of a weakened groES promoter lacking its putative UP-element and was made by replacing sspA in the plasmid pF2-SspA [15] with DNA encoding PigR-V. These plasmids were electroporated into either cells of the previously described LVS ∆pigR mutant strain [17], or cells of a LVS ∆pigR ∆mglA mutant strain; LVS ∆pigR ∆mglA was created by using the pEX2-∆mglA vector [15] in the ∆pigR background, by allelic exchange and confirmed by Southern blotting.
ChIP-Seq was performed with cells of the following strains: LVS βʹ-V; LVS σ70-V; LVS σ32-V; LVS PigR-V; LVS SspA-V; LVS HipB-V; LVS (as a mock control); LVS containing plasmid pF (as a mock control); LVS ∆pigR containing plasmid pF2-PigR-V; and LVS ∆pigR ∆mglA containing plasmid pF-PigR-V. In order to perform ChIP-Seq on MglA, we used cells of the LVS strain synthesizing MglA with a C-terminal TAP tag (LVS MglA-TAP) at native levels, which has been described previously [15]. Cells were grown at 37°C in 100 mL of supplemented MH to mid-log (OD600 0.3–0.4), and when indicated, rifampicin (Sigma) was added to a final concentration of 50 μg/mL for 30 minutes before crosslinking. Cells were incubated in a final concentration of 1% formaldeyhyde (Sigma) for 30 minutes, after which glycine (Sigma) was added to a final concentration of 250 mM. ChIP was performed in biological triplicate (excepting β′ + rifampicin, the LVS pF empty vector control, and LVS ∆pigR ∆mglA pF-PigR-V, which were performed in duplicate, and σ70, which was performed in quadruplicate) with either 40 mL or 80 mL of culture using anti-VSV-G agarose beads (Sigma) for cells synthesizing VSV-G tagged transcription factors or IgG Sepharose beads (GE Healthcare) for cells synthesizing TAP-tagged MglA essentially as described previously [50], except that a water bath sonicator (Biorupter, Diagenode) was used to lyse cells and shear chromosomal DNA to 200 to 500 bp. Immunoprecipitated DNA was quantified using the Quant-iT PicoGreen dsDNA Assay Kit (Invitrogen). The same protocol was performed with the untagged LVS strain as a mock immunoprecipitation (mock IP) control.
Illumina libraries were constructed with approximately 2 to 160 ng immunoprecipitated DNA using either the TruSeq DNA Sample Prep Kit (Illumina) or the NEBNext Ultra DNA Library Prep Kit for Illumina (NEB), generally following the supplied protocols. In using the TruSeq DNA Sample Prep Kit, adapters were diluted 1:10 before ligation and libraries were gel-purified after 11 cycles of amplification. When using the NEBNext Ultra DNA Library Prep Kit, adapters were diluted 1:10 before ligation and libraries were size-selected using Agencourt AMPure XP beads prior to 12 cycles of amplification. Libraries were quantified using the Quant-iT PicoGreen dsDNA Assay Kit (Life Technologies) and sequenced by Elim Biopharmaceuticals, Inc. (Hayward, CA), using an Illumina Genome Analyzer Ilx generating 36 bp single-end reads or an Illumina HiSeq 2500 generating 50 bp single-end reads. Sequencing reads have been submitted to the NCBI Sequence Read Archive (SRA, http://www.ncbi.nlm.nih.gov/Traces/sra) with the accession number SRP055716.
For each strain, the reads were mapped to the F. tularensis subsp. holarctica LVS genome (NCBI locus AM233362) and the sequence of the integrated plasmid, if applicable, using bowtie2-2.0.6 [51]. Regions of enrichment were called using QuEST, version 2.42 [52]. The three mock IP biological replicates, consisting of approximately 30.6 million reads, were merged and used as a background control for each biological replicate. The two pF empty vector control IP biological replicates, consisting of approximately 42.7 million reads, were merged and used as a background control for the ectopic PigR-V experiments. Peaks in each biological replicate are regions that fit the following criteria: they are 1.5-fold enriched for reads over background, with a positive peak shift and strand correlation, and a q-value of less than 0.01. Peaks for each immunoprecipitated protein were defined as the maximal region identified in at least two biological replicates. Promoter regions are defined as the maximal regions of enrichment of βʹ plus rif, σ70 or σ32. Tracks were visualized using the Integrative Genomics Viewer (IGV), version 2.3 [53]. Peak analyses were carried out using Perl scripts, samtools, version 0.1.17 [54], and BEDtools, version 2.17.0 [55].
Cell lysates were separated by SDS-PAGE on 4–12% or 12% Bis-Tris NuPAGE gels in MES or MOPS running buffer (Life Technologies). Either the iBlot dry blotting system or the XCell II Blot Module (Life Technologies) was used to transfer proteins to either PVDF or nitrocellulose. Membranes were blocked with SuperBlock Blocking Buffer (Pierce) with 0.25% Surfact-Amps 20 (Pierce) for 1 hour to overnight. Membranes were then probed with polyclonal anti-VSV-G (diluted 1:1,500; Sigma) or anti-GroEL (diluted 1:160,000; provided by Karsten Hazlett, Albany Medical College, Albany, New York, United States) for one hour, washed (10 minutes incubations in TBST plus 0.25% Surfact-Amps 20, 4 times) and re-blocked for 1 hour. After membranes were incubated with polyclonal goat anti-rabbit (diluted 1:10,000; Pierce) and washed, proteins were detected using SuperSignal West Pico Chemiluminescent Substrate (Life Technologies).
Cells of the LVS wild-type strain and cells of the LVS ∆pigR mutant strain (described in 17) were grown to mid-log in biological triplicate. 1 mL of each sample was pelleted (20,000 rcf for 5 minutes), resuspended in 500 μL Qiagen buffer RLT, frozen on dry ice, and stored at -80°C. Equal amounts of lysate, normalized to OD600, in 4 μL Qiagen buffer RLT and 1 μL water were submitted to the Epithelial Cell Biology Core Facility (Boston Children’s Hospital) for processing using the Nanostring nCounter Prep Station and Digital Analyzer according to the manufacturer’s instructions. For each replicate, total transcript counts were normalized using internal controls with background subtraction, as per manufacturer’s instructions. Transcript abundance was determined by averaging biological triplicates. Criteria indicating a significant change in gene expression are a 2-fold change in transcript abundance and p-value < 0.05 in a two-tailed Student’s t-test.
Genes with significant changes in expression in cells lacking PigR in comparison to wild-type cells (greater than 3-fold by microarray [17] or 2-fold by Nanostring) were examined for promoter regions with detectable PigR, identifying 11 genes (FTL_0026, iglA, pdpA, FTL_0131, FTL_0207, FTL_0449, FTL_0491, FTL_1218, FTL_1219, FTL_1678, FTL_1790). The 400 bp region surrounding the maximal PigR binding site, upstream from 11 genes, excluding coding regions, was searched for a common motif using MEME, version 4.9.1 [28]. We have named the second result, which was present in all 11 promoters and consisted of the consensus sequence TGCTGTA, the PigR response element (PRE).
LVS cells were grown in aerated liquid culture at 37°C in supplemented MH to mid-log (OD600 0.3–0.4), and RNA was isolated from 10mL of cells in triplicate as described previously [18]. RNA-Seq was used to identify transcription start sites. In particular, from each sample we prepared three cDNA libraries derived from the 5’ ends of RNAs. The first library was generated from RNAs carrying a 5’ triphosphate (prepared as described in [56]), the second library was generated from RNAs carrying a 5’ monophosphate (prepared as described in [56]), while the third library was generated from RNAs carrying either a 5’ triphosphate or a 5’ monophosphate. The third library was prepared by omitting a single step (treatment with Terminator 5’ exonuclease) from the procedure used to generate RNAs carrying a 5’ triphosphate. To identify high quality transcription start sites we first identified 3,120 genomic loci where 50 or more sequencing reads aligned in one of three libraries generated from RNAs carrying a 5’ triphosphate. Of these 3,120 loci, we identified 452 sites that were significantly enriched in a comparison of libraries generated from RNAs carrying a 5’ triphosphate with libraries generated from RNAs carrying a 5’ monophosphate and significantly enriched in a comparison of libraries generated from RNAs carrying a 5’ triphosphate or a 5’ monophosphate with libraries generated from RNAs carrying a 5’ monophosphate. These 452 sites were further filtered to remove those associated with rRNAs, tRNAs, or repeated sequences annotated as transposases. Among the remaining sites, we identified 110 high quality start sites as those found within 1 kb of a translational start site for an annotated ORF and located within a promoter region defined by ChIP-Seq (S8 Table).
Primer extension was used essentially as described previously [57] to determine the putative transcription start sites for pdpA and FTL_1219.
For each transcription start site identified by RNA-Seq, putative -10 and -35 elements were predicted based on homology to the E. coli consensus sequence. For each promoter, 11 bp of sequence, 5 bp upstream from the -35 element and extending upstream, was submitted to FIMO [58] to search for the PRE. Only three promoters were found to contain the PRE with a p-value <0.001, all of which are known to be PigR-regulated. None of the remaining promoters are known to be PigR-regulated and none of them contained the PRE (p<0.001).
Cells were grown to mid-log phase, and β-galactosidase activity was assessed essentially as described previously [17]. Assays were performed at least twice in triplicate on separate occasions. Representative data sets are shown. Values are averages based on one experiment.
|
10.1371/journal.pgen.1007240 | Methods for fine-mapping with chromatin and expression data | Recent studies have identified thousands of regions in the genome associated with chromatin modifications, which may in turn be affecting gene expression. Existing works have used heuristic methods to investigate the relationships between genome, epigenome, and gene expression, but, to our knowledge, none have explicitly modeled the chain of causality whereby genetic variants impact chromatin, which impacts gene expression. In this work we introduce a new hierarchical fine-mapping framework that integrates information across all three levels of data to better identify the causal variant and chromatin mark that are concordantly influencing gene expression. In simulations we show that our method is more accurate than existing approaches at identifying the causal mark influencing expression. We analyze empirical genetic, chromatin, and gene expression data from 65 African-ancestry and 47 European-ancestry individuals and show that many of the paths prioritized by our method are consistent with the proposed causal model and often lie in likely functional regions.
| Genome-wide association studies (GWAS) have revealed that the majority of variants associated with complex disease lie in noncoding regulatory sequences. More recent studies have identified thousands of quantitative trait loci (QTLs) associated with chromatin modifications, which in turn are associated with changes in gene regulation. Thus, one proposed mechanism by which genetic variants act on trait is through chromatin, which may in turn have downstream effects on transcription. In this work, we propose a method that assumes a causal path from genetic variation to chromatin to expression and integrates information across all three levels of data in order to identify the causal variant and chromatin mark that are likely influencing gene expression. We demonstrate in simulations that our probabilistic approach produces well-calibrated posterior probabilities and outperforms existing methods with respect to SNP-, mark-, and overall path-mapping.
| Discerning the genetic and molecular basis of complex traits is a fundamental problem in biology. Genome-wide association studies have revealed that the majority of variants associated with disease lie in noncoding regulatory sequences [1, 2]. Identifying the target genes of these variants and the mechanisms through which they act remains an open problem [3]. Recent efforts to systematically characterize how genetic variation impacts more granular molecular phenotypes have yielded thousands of single nucleotide polymorphisms (SNPs) that associate with local and distal histone modifications—termed histone quantitative trait loci (hQTLs) [4–7]. Furthermore, recent studies have identified many expression quantitative trait loci (eQTLs) that co-localize with hQTLs, implying there may exist a shared genetic influence on epigenetic traits and gene expression [8–11]. Therefore, one proposed mechanism by which regulatory variants may affect gene expression and thereby impact traits is through changes in chromatin state [10]. However, this putative chain of causality whereby the effects of SNPs on expression are mediated by chromatin modifications has yet to be established. This is further compounded by the complex space of plausible causal directions connecting transcription factor binding, DNA methylation, chromatin variation, and gene expression. Since laboratory experiments are very costly, there is a need for statistical methods that can accurately prioritize the causal SNP and chromatin mark within an implicated region under a plausible causal model. However, even if the causal direction is given, pinpointing the exact SNP and mark within a genomic region is very challenging due to the confounding effects of linkage disequilibrium (LD) among SNPs and correlations among marks [5, 6, 10, 12–14].
Methods to investigate the relationships between the genome, the epigenome, and expression have largely focused on quantifying the overlap between hQTLs and eQTLs [10, 14, 15]. Previous studies have sought to identify hQTLs by selecting the SNP with the strongest p-value for association to a local chromatin mark and to local gene expression [10, 14, 15]. Moreover, various methods exist for the fine-mapping of SNPs that may be concurrently affecting two traits, including eCAVIAR [16] and Coloc [17]. Although these methods can be applied to jointly analyze SNP, chromatin, and expression data, they do not model the causal path whereby SNPs impact expression through chromatin alteration.
Here we propose a fine-mapping framework, pathfinder, that explicitly models the hierarchical relationships between genome, chromatin, and gene expression to predict both the causal SNP and the causal mark within a gene region that are influencing expression of a given gene. Our framework assumes a causal model where a SNP impacts a chromatin which in turn alters gene expression. In our framework we refer to a “causal” SNP as any SNP that disrupts inter-individual variation of chromatin state either through a direct biological mechanism (e.g., chromatin accessibility) or indirectly through an unobserved biological mechanism. Similarly, we refer to a “causal” chromatin mark as either a mark that biologically alters expression or that tags an underlying epigenetic regulatory mechanism of expression. Our framework takes as input the strength of association (as quantified through the standard Z-scores) between all SNP/mark pairs and all marks to expression as measured in a given set of individuals. To explicitly account for the correlation structure among SNPs and marks, we use a Matrix-variate Normal distribution to model all Z-scores jointly. By construction, this allows our probabilistic model to assign posterior probabilities for each SNP, mark, and path (where paths include all possible SNP-mark combinations) to be causal in the region. A key advantage of our approach is that it produces well-calibrated posterior probabilities for causality. Thus, pathfinder can be used to prioritize variants and marks for validation experiments.
In simulations we compare against several existing methods, demonstrating that pathfinder outperforms alternative approaches with respect to both accuracy and calibration. This is largely because our comparators do not take into account mark-expression associations. In some cases, these additional associations may help distinguish between two potentially causal paths that have comparable evidence for causality. For example, in cases where a SNP is associated with expression of a local gene and is also associated with two local chromatin marks, knowledge of the impact of each mark on gene expression may help distinguish between two possible paths for causality. Finally, we analyze genotype, chromatin and expression data from 65 African-ancestry and 47 European-ancestry individuals. We show that the top causal SNPs proposed by pathfinder tend to lie in more functional regions and disturb more regulatory motifs than expected by chance. We also present evidence that most of the top paths reported by pathfinder demonstrate consistency with our proposed sequential model, thus strengthening the case for our method’s applicability to empirical biological data.
Here we introduce a hierarchical statistical method for fine-mapping of causal SNPs and chromatin marks (e.g., histone modifications) that may be concordantly influencing gene expression within a genomic region. We build upon previous insights that a vector of Z-scores is well-described by a Multivariate Normal (MVN) distribution parameterized by LD [13, 18, 19] to model association statistics between chromatin marks and gene expression. We analyze all chromatin peaks across four mark types (DHS, H3K4me1, H3K4me3, and H3K27ac) jointly in the same framework; we refer to a “mark” as a chromatin peak at a particular location, and “mark types” as DHS, H3K4me1, H3K4me3, and H3K27ac. To simultaneously take into account both SNP LD and the correlations between chromatin marks, we use the Matrix-variate Normal distribution to jointly model association statistics between all SNPs and marks within a region. Our method takes as input SNP-mark and mark-expression associations within a region centered around a particular gene, as well as correlations among all SNPs (LD) and correlations among all considered marks. Pathfinder enumerates over all possible causal paths, considering one causal SNP and one causal mark for each path, and outputs a posterior probability for each path to be causal, which can subsequently be used to prioritize SNPs and marks for validation. We compute marginal probabilities for individual SNPs (or marks) to be causal by summing the posterior probabilities over all paths that contain the SNP (or mark). For simplicity, in this work we refer to a “causal” mark as a mark that either causally drives inter-individual variation of gene expression or is correlated to an underlying causal mechanism (e.g. transcription factor binding), though it may not be biologically causal for expression.
The advantage of our method over existing approaches is that it integrates mark-expression associations which may help distinguish between two paths with otherwise comparable evidence for causality. We illustrate a scenario in Fig 1. Consider a genetic region where SNP g1 has a strong association with two local marks h1 and h2, as well as a significant association with gene expression. Using only SNP-mark and SNP-expression effects, we are unable to discern whether SNP g1 influences expression through mark h1 or h2. However, if we consider mark-expression effects, we see that mark h1 has a strong association with gene expression where mark h2 does not. This additional information helps support the hypothesis that there is a causal path from SNP g1 to mark h1 to gene expression.
We used simulations to compare pathfinder’s performance against alternative methods with respect to SNP-, mark-, and path-finding efficiency as well as the calibration of its posterior probabilities. We generated genetic, chromatin, and gene expression data for 10,000 50kb regions, each centered around a single gene, over 100 individuals, using SNP LD and mark correlations derived from 65 Yoruban (YRI) individuals (see Methods). We define a “mark” as an individual peak location for any mark type in the dataset (DHS, H3M4me1, H3K4me3, or H3K27ac). For each gene, we randomly assigned a single causal pathway from one SNP to one mark to gene expression. We then ran our methods on all regions individually and assessed their ability to correctly prioritize the true causal path in each region (Methods).
We compare against an independent fine-mapping approach (whereby we fine-map SNP-mark associations and mark-expression associations independently and take the product of the resulting probabilities to produce posterior probabilities for paths), a Bayesian network analysis [20], a naive ranking (where we rank SNP-expression and mark-expression associations to prioritize SNPs and marks within a region; for path-finding, we rank the product of these two), a formal colocalization method [17], and finally, against overlaps between eQTLs and hQTLs within a region centered around a gene of interest (see Methods). Unlike the first four approaches, the overlap methods do not produce rankings, but yield candidate sets of causal SNPs, marks, and paths. For this reason, we present these results in a separate analysis using an alternative metric for comparison.
We find that pathfinder has consistently better performance than the other ranking approaches with respect to all three features—SNP-, mark-, and path-mapping within a region (Fig 2). For example, association ranking, Coloc, Bayesian network analysis, and independent fine-mapping accumulate 55%, 62%, 47%, and 13% of the top paths on average in order to recapture 90% of the causal paths, whereas our method only requires 8% of the top paths. Note that SNP-expression association ranking is equivalent to running a basic eQTL analysis, which does not take into account chromatin data, in order to identify causal SNPs. A similar improvement in accuracy was observed for the size of the credible sets, defined as the number of SNPs required to capture a given percentage of the causal variants (S1 Table).
Next, we evaluated pathfinder’s performance compared against standard analyses that investigate overlaps between hQTLs and eQTLs within a genomic region. In such experiments, the variant with the strongest association to each local chromatin mark is selected, as well as the variant with the strongest association to local gene expression. In addition, marks are filtered to ensure a 10% FDR (see Methods). This produces a set of candidate marks, as well as one candidate SNP per mark, and one SNP deemed causal for gene expression in the region. Implicitly, the overlap of these variants suggests a set of candidate SNPs, marks, and paths for the region. For the same set sizes, pathfinder identifies 96% of the causal marks versus 74% in the standard overlap approach (Fig 3). SNP-finding accuracy is comparable between the two methods.
We next assessed the calibration of the posterior probabilities for causality output by pathfinder. Our method has slightly deflated credible sets for SNP- and path-finding, but well-calibrated credible sets for mark-finding (Fig 4). In contrast, the independent fine-mapping approach has consistently inflated credible sets—that is, it captures more causal paths than expected, but also has drastically larger credible set sizes. For example, when accumulating 90% of the posterior probabilities over all regions, pathfinder captures 88% of the true causal paths within the top 380 candidate paths, whereas independent fine-mapping captures 94% of the causal paths within the top 1026 candidate paths. Similar outcomes were attained for the 50% and 99% credible sets (S1 Fig). Overall, pathfinder’s credible sets are less biased and narrower than those obtained through the independent fine-mapping approach.
Finally, we investigated the effects of simulation and method parameters on pathfinder’s accuracy. Firstly, we varied the causal SNP and mark effect sizes such that the variance explained of mark and gene expression ranged from 0.1 to 0.5. As anticipated, increased heritability leads to better performance (See Fig 5A–5C). Secondly, in order to assess the impact of SNP LD and mark correlations on SNP- and mark-finding performance, we stratified our existing simulations based on the mean correlation of the causal SNP or mark to all other SNPs or marks (See Fig 5D–5I). We grouped our simulations into three categories: low, medium, and high correlations. As anticipated, SNP-finding performance decreases slightly as SNP LD increases. Notably, mark-finding performance is actually improved at higher SNP LD. This is due to the redundancy in information about SNP-mark associations at the causal mark when these effects are exhibited across multiple correlated SNPs. SNP- and mark-finding performance, however, do not seem to be significantly affected by mark correlations in our simulations—at least not at the level of variation exhibited in our data. In addition to stratifying our existing simulations by LD, we also assessed the impact of using European rather than African LD in the same regions, as European LD is known to be more extensive. Here we retained the YRI mark and expression data in order to isolate the effect of SNP correlations. The credible set sizes computed from the CEU dataset do not substantially differ from those obtained in YRI (S2 Table). This result demonstrates that the more extensive LD observed in European individuals will not significantly affect pathfinder’s performance. Thirdly, we evaluated the effect of the prior variance tuning parameter on fine-mapping performance (See Fig 5J–5L). The prior variance is an estimate of the variance explained by the causal SNP and mark in the region, as we do not know a priori what the causal effect sizes are. We show that the optimal range for the prior variance parameters is between 5 and 10, in simulations with a variance explained of 0.25 on both levels. Overall, performance does not seem to change drastically in response to variations in the prior variance, even significantly outside of this optimal range.
Our hierarchical model makes several key assumptions that may sometimes be violated in empirical data. Firstly, pathfinder assumes that a single causal SNP and a single causal mark are driving the associations within a region, where in reality there may exist multiple true causal SNPs or marks [13, 19]. Secondly, pathfinder assumes that SNP effects on gene expression are mediated by a chromatin mark, which may not be the case in real data. We therefore assessed the performance of our method when these two assumptions are violated in various ways, diagrammed in Fig 6.
First, we investigate violations 1–3, which include multiple causal pathways throughout the region. Path-mapping accuracy, measured by the proportion of causal paths identified, is reduced in all three scenarios (Fig 6). Note that the number of causals identified does not necessarily decrease, but rather the proportion, as there are more causal paths in each region. SNP- and mark-finding accuracy under these violations are also compromised, but with two notable exceptions. In the multi-causal-SNP scenario, mark-finding accuracy increased in comparison with the single-SNP simulations; for example, only 8% of marks were selected (versus 18% in the single causal simulations) to capture 90% of the causal marks. In the multi-causal-mark scenario, SNP-finding accuracy increased. Intuitively, this is due to the redundancy in the signal that is captured by the Matrix-variate Normal distribution.
We next investigate violations 4–5, in which an additional SNP or mark influences gene expression directly. We observe in these two scenarios that performance is reduced for SNP-, mark-, and path-finding, but not drastically. For example, in order to capture 90% of the causal paths, pathfinder must select on average 25% and 28% of paths under violations 4 and 5, respectively (compared with 15% under standard simulations). Because anti-correlated marks (e.g. activating and repressing marks) often tend to act in the same region, we also assess pathfinder’s behavior specifically when two marks have opposite effects on expression. As expected, pathfinder’s performance does not decline in the presence of anti-correlated peaks (S2 Fig).
Finally, we discuss pathfinder’s performance under violations where the causal order is modified (violations 6–7). Under violation 6, where a single causal SNP affects gene expression directly, which in turn affects a single mark, pathfinder actually captures a higher proportion of the affected marks and overall paths. For example, in order to capture 90% of the causal paths, pathfinder must select on average only 3% of the top-ranked paths (compared with 15% under standard simulations). In violation 7, where the SNP has independent effects on the mark and the gene expression, we show that pathfinder’s accuracy in finding the causal mark and path is significantly reduced. Note that in this case, the “path” is not truly a path but a SNP/mark pair, as effects of the SNP on mark and gene expression are independent. Our power in distinguishing between these two models depends on the prior variance explained parameter. Under violation 7, the variance explained in gene expression by the causal mark is much smaller than expected, thus reducing our confidence in the true causal configuration. We conclude that under the SNP→expression→mark violation, pathfinder will identify causal paths very confidently even if they do not follow the assumed SNP→mark→expression model. Therefore a high posterior probability for a path may not be sufficient evidence for causality. On the other hand, when SNP effects on mark and expression are independent, pathfinder is less likely to produce false positives. For these reasons, we recommend a pre- or post-filtering step to retain only those regions that show some prior evidence for the SNP→mark→expression model using a conditional analysis or partial correlation approach (Methods).
For completeness, we also assess existing methods under these simulations (S3 Fig). Most notably, the simple association-ranking approach shows a distinct improvement under violations 6 and 7, in which SNPs have a direct effect on gene expression. This is expected as pathfinder assumes the causal effect to be mediated by chromatin. A similar improvement can be observed for Coloc under violation 7, in which the SNP affects both chromatin and gene expression directly.
We evaluated the behavior of our hierarchical fine-mapping method when applied to empirical data. We performed these analyses on data from 65 YRI individuals whose genotypes were obtained through 1000 Genomes, and whose PEER-corrected H3K4me1, H3K4me3, H3K27ac, DHS, and RNA expression levels in lymphoblastoid cell lines (LCLs) were obtained from [10]. In each region, we analyzed all four mark types jointly (H3K4me1, H3K4me3, H3K27ac, and DHS) by including all peaks spanning the region for each mark type. Each peak of each mark type was therefore treated as a single chromatin mark. We filtered the 14,669 regions using a two-step regression analysis to yield 1,317 regions that showed evidence for the sequential model of SNPs affecting histone marks which in turn affect gene expression (see Methods). pathfinder’s runtime scales approximately as s3t3, where s and t are the number of SNPs and marks within a region, respectively. On average, each 50kb region contained 160 SNPs and 25 marks. Most runs were completed in under a few minutes. The most dense region contained 331 SNPs and 66 marks and took approximately 21 minutes (S4 Fig).
In Table 1, we report the average 50%, 90%, and 99% credible set sizes produced when running pathfinder on real data. We compare against basic eQTL mapping, where we fine-map SNPs to gene expression ignoring chromatin data. We show that the credible set sizes are significantly narrower when running pathfinder with all three levels of data, consistent with our findings in simulations. For example, eQTL mapping requires an average of 45.3 SNPs in order to capture 90% of the posterior probability for SNP causality, whereas pathfinder only requires 28.4 SNPs. If we define a gene to be fine-mapped if 99% of the posterior probability mass for SNP causality is contained within the top 10 SNPs or fewer, then standard eQTL mapping fine-maps 46 of the genes in our data, whereas pathfinder fine-maps 73 of the genes. Notably, pathfinder also requires only 1.8 marks on average in order to capture 90% of the posterior probability for causal marks. In 82% of the regions where the top two marks capture more than 90% of the posterior probability, these two marks are two distinct peaks of the same mark type.
The mean variance explained observed in the top path chosen by pathfinder, across all conforming regions, were 0.38 (s.e. 0.01) for the SNP-mark effect and 0.20 (s.e. 0.01) for the mark-expression effect (S5 Fig). These effects are reasonably consistent with the 25% variance explained we used in simulations at each level (see Simulations). The correlation between the SNP-mark and mark-expression effect size magnitudes in the top selected paths across all regions was 0.03 (p = 0.400). That is, the strength of the SNP-mark effect did not seem to correlate with the strength of the mark-expression effect. We assessed the relative impacts of each type of histone mark by computing the proportion of probability mass assigned to each mark type in aggregate over all regions (S3 Table). H3K4me3 is the most informative mark type in this data, capturing 31% of the total probability mass despite being the least prevalent of all four mark types, constituting only 13% of all marks.
We also report the size of pathfinder’s credible sets when applied to empirical CEU data rather than YRI in Table 2. These two datasets are not directly comparable, as the types of epigenetic marks and their quantities differ substantially. Nonetheless, we demonstrate that pathfinder’s performance on the CEU dataset does not drastically diverge from its behavior in YRI. Data pre-processing strategies such as PCA and PEER correction may substantially impact the number of mark-expression correlations that are retained [21]. We find that credible set sizes for PEER-corrected data are narrower, giving a slight but significant improvement in performance (S4 Table).
As our pre-filtering step was designed to preserve only regions in which SNP effects on gene expression are mediated by chromatin, we expected a large majority of the analyzed regions to show evidence for this mechanism. To confirm this, we investigated whether the top paths prioritized by our method demonstrate consistency with this causal model. We defined a set of top paths as those which were ranked first in a region and whose posterior probabilities for causality were assigned by pathfinder to be greater than 0.1. This resulted in 480 total top paths. Out of 480 top paths, only 12 had a significant (p < 0.05/480) partial correlation between SNP and gene expression after controlling for chromatin. However, 193 paths had a significant partial correlation between SNP and chromatin after controlling for gene expression. This finding suggests that the top paths are more consistent with the SNP→mark→expression model than with a SNP→expression→mark model.
Next we examined the relationship between the product of the effect sizes between SNP-mark and mark-expression against the overall SNP-expression association (Fig 7). We expect this relationship to be correlative; if truly mediated by the mark in question, the overall SNP-expression effect size should be proportional to the product of the two contributing effect sizes. Note that we weight our correlation by the reported posterior probability for each path, such that the paths we have more confidence in will contribute more to this metric. We find a high correlation (r = 0.91) between these effect size vectors for our top paths, as compared with a correlation of r = 0.36 when running the same analysis on random paths within each region. This result indicates that pathfinder is identifying many pathways that are likely to be following its causal model.
In Table 3, we list the top ten paths prioritized by pathfinder across all real data regions. Most SNPs implicated in these paths are known to alter several regulatory motifs and often lie in an enhancer region or a promoter region of the genes whose expression they affect. 59% (s.e. 2%) of the SNPs implicated in the top paths fall into active ChromHMM states (1–7) in LCLs, including active TSS, flanking active TSS, transcription at gene 5’ and 3’, strong transcription, weak transcription, genic enhancers, and enhancers. Only 47% (s.e. 2%) of random paths fall into these active states (p = 0.001834). Moreover, on average, SNPs in the top paths disturbed 5.35 (s.e. 0.26) regulatory motifs, whereas random SNPs chosen at the same regions only disturbed 4.40 (s.e. 0.20) motifs on average (p < 0.001). We did not, however, observe a similar change in transcription factor binding affinity at these motifs (δ = 5.26 vs δ = 5.27, (p = 0.511)). As an example, in Fig 8A–8D, we display the genomic context for the top region reported by pathfinder, including average mark signals for DHS, H3K4me1, H3K4me3, and H3K27ac, stratified by genotype, in a 4kb region centered around the TSS of the NDUFA12 gene. The implicated SNP lies within the NDUFA12 TSS. Fig 8E plots the gene expression signal against that of the top mark, stratified by genotype. In S6 Fig, we show associations for the top region reported by pathfinder, spanning a 50kb region centered around the NDUFA12 TSS.
Next we examined the spatial relationships between the SNP, mark, and TSS implicated in the top paths reported by pathfinder (Fig 9). SNP to mark and mark to TSS distances were significantly lower in our selected paths compared with randomly chosen paths at the same regions. The average distance from SNP to mark in pathfinder’s top paths was approximately 11.7kb, compared to 15.3kb in randomly chosen paths (p < 0.001). The average distance from mark to TSS in selected paths was approximately 8.6kb, compared to 9.7kb in randomly chosen paths (p = 0.026). SNP to TSS distances were not significantly different in top versus random paths (p = 0.108), with top SNPs lying on average 11.7kb away from the TSS and random SNPs lying 12.4kb away. 5% of top SNPs lied within 2kb of the TSS while 15% lied within 2kb of the corresponding peak. 23% of peaks in the top paths lied within 2kb of the gene TSS. S7 Fig displays all three distances where top paths are stratified by mark type.
To further validate the top paths chosen by pathfinder, we determined the extent to which SNPs in these paths overlap with eQTLs that have been identified in LCLs using the larger scale Geuvadis data set [22]. 21% of the top paths contained SNPs that were also identified as eQTLs from the Geuvadis data set. In comparison, when randomly choosing paths at the same regions, only 11% overlapped with eQTLs (p < 0.001). Simply choosing the SNP with the highest association with gene expression in each region (equivalent to standard eQTL-mapping) resulted in an overlap of 24% with existing eQTLs. These results contradict the improvement in accuracy demonstrated in simulations when using pathfinder. We suspect this discrepancy is due either to imperfect locus ascertainment (i.e., a number of loci may include SNPs that directly affect gene expression rather than indirectly through chromatin) or the fact that the Geuvadis eQTLs were also selected using standard fine-mapping approaches and we may thus expect a stronger agreement between the two resulting eQTL sets.
We also investigated the extent to which pathfinder’s top SNPs overlap with eQTLs that have been experimentally validated through differential expression in an LCL dataset [23]. Here, we define the set of validated eQTLs to be those whose p-values for differential expression passed a threshold of 0.01. We find that 2.2% (or 13) of pathfinder’s top SNPs overlap with this validated set, where choosing the SNP with the highest association with gene expression in each region resulted in an overlap of 2.3% (also 13 SNPs).
Finally, we investigated whether any of the top paths reported by pathfinder could be found within GWAS hit regions for various autoimmune diseases, as our data were collected from LCLs. These autoimmune diseases included Celiac disease, Crohn’s disease, PBC (Primary Biliary Cirrhosis), SLE (Systemic Lupus Erythematosus), MS (Multiple Sclerosis), RA (Rheumatoid Arthritis), IBD (Irritable Bowel Disease), and UC (Ulcerative Colitis). We restricted to GWAS hits with variants associated to the trait with p < 5 × 10−8. We found that 19 of our 480 top paths were contained in a GWAS-implicated region. In Table 4, we report the paths that localized within autoimmune GWAS regions. In order to determine whether our top paths are truly enriched in GWAS regions, we established how many of these paths appear in an equivalent number of random regions that have not been implicated by an autoimmune GWAS. We centered each random region around a SNP that was matched for a similar MAF and LD score as the GWAS tag SNP. We ran this analysis 100 times to define a null distribution for the number of top paths found in a background region. We found that 19 out of 480 top paths was not a significant enrichment (p = 0.44).
In this work we proposed a hierarchical fine-mapping framework that integrates three levels of data—genetic, chromatin, and gene expression—to pinpoint SNPs and chromatin marks that may be concordantly influencing gene expression. A key contribution of our approach is the ability to model the correlation structure in the association statistics using a Matrix-variate Normal distribution. Our approach is superior to existing methods, demonstrating the advantage of using a probabilistic approach that takes into account the full sequential model. Moreover, pathfinder produces well-calibrated posterior probabilities, and is thus a reliable method for the prioritization of SNPs and marks for functional validation.
We conclude by addressing some of the limitations of our method. Most notably, our method is based upon the SNP→mark→expression assumption. In many genomic regions that show simultaneous evidence for SNP to mark and SNP to gene expression effects, this model will not necessary hold true. In simulations, we show that under the SNP→expression→mark violation, pathfinder may identify causal paths very confidently, leading to false positives under the proposed model. When a SNP is in fact independently influencing a mark and gene expression, pathfinder is less likely to produce false positives. However, the risk of mis-appropriating our method in this way can be reduced by requiring genomic regions to show evidence for our causal model. We recommend a pre-filtering step before running pathfinder on real data that we outline in Methods. In our empirical data analyses, we demonstrate that this two-step regression robustly filters out non-conforming regions. We also acknowledge that, though there are multiple lines of evidence for SNPs influencing expression through local hQTLs, recent works have also emphasized the importance of interactions with distal hQTLs. Thus, developing a systematic way to incorporate data in distal regions with evidence for interactions with a local eQTL would be a fruitful direction. Moreover, pathfinder assumes that the true causal SNP and mark within a region are present in the data, which may not always be the case. In this scenario, pathfinder will instead place its confidence in the SNP or mark that best correlates with the missing causal SNP or mark in question. Similarly, many epigenetic marks are not themselves causal for gene expression, but are simply correlated to a causal event (e.g., transcription factor binding). It is also often the case that multiple marks at promoter and enhancer regions are concordantly acting to impact gene expression. In these cases, individual marks are not necessarily causal in themselves, but may be viewed as a cause for inter-individual variation or simply correlated to a causal factor. In this light, pathfinder aims to identify the epigenetically modifying region so that it can be tested experimentally and/or characterized functionally (for example, to identify the effector transcription factor). We also note that pathfinder currently uses an approximation whereby the observed Z-score at the causal SNP is used to estimate the true NCP at the causal SNP (Methods). We leave this to be addressed in future work; this correction will likely further improve the calibration of our method’s credible sets. We note that pathfinder only uses individuals for which we simultaneously have genetic, chromatin, and gene expression measurements, thus ignoring eQTL data that has been measured in larger sample sizes. However, eQTL data from larger samples could potentially be used as a prior for expectation of SNP causality or perhaps for validation after running pathfinder on real data. Finally, although our analyses showed that H3K4me3 marks are the most informative for fine-mapping, small data set sizes analyzed in this work prohibit us in making definitive conclusions on which mark is most useful leaving such avenues for future work.
For each individual, let h be the signal value for the causal histone mark and G be their vector of genotypes at a region containing s SNPs. Let E be the individual’s mRNA expression level for the gene at this region and H be a vector representing all t marks at the region, which contains h. Here we analyze all individual peak locations across all available mark types in a joint framework. As such, each of t individual marks represents one peak location for a particular mark type. Our causal framework can be modeled as:
h = G β g + ϵ g (1) E = H β h + ϵ h (2)
where ϵ g ∼ N ( 0 , 1 - σ g 2 ) and ϵ h ∼ N ( 0 , 1 - σ h 2 ). The vector βg represents the allelic effects on the causal histone mark whose entries will be non-zero only at the causal SNP. The vector βh represents the histone mark effects on expression levels whose entries will be non-zero only at the causal histone mark. σ g 2 and σ h 2 represent the variance explained at the SNP-mark and mark-expression levels.
We simulated data for 100 individuals over 10,000 50KB regions, using genotypes and LD from 65 YRI individuals obtained through 1000 Genomes [26]. SNP and mark correlations in our simulations were taken from the true correlations exhibited in these regions derived from these individuals. To determine causal status, we randomly chose one SNP and one mark to be causal in each region, thus defining a causal path through the data. Subsequently, we standardized genotypes and simulated values for chromatin marks and gene expression over all 100 individuals.
In order to simulate correlations between histone marks as observed in our empirical data, we drew mark values from an MVN as N ( H i n d , ϵ g Σ h ), where the means, Hind = HcΣh,c, represent the induced values on non-causal marks due to correlations with the causal mark. The mean mark values for the causal mark were generated for each of the 100 individuals as Hc = βgGc, where Gc is the genotype of the individual at the causal SNP, the effect size βg was drawn from a normal distribution, N ( 0 , σ g 2 ), with variance set to the desired variance explained by SNPs on marks σ g 2 = 0 . 25, with the error term ϵg set to 1 - σ g 2. Finally, the individuals’ values for gene expression are computed as E = βhHc + ϵh, where Hc is the causal mark value as computed from the MVN, the effect size βh was set to the desired variance explained from mark to expression σ g 2 = 0 . 25, with the remaining error term given by N ( 0 , 1 - σ g 2 ).
For simulations in which there were multiple causal SNPs or marks, we randomly drew m or p, the number of causal SNPs or marks, from a binomial distribution where the expected number of causals per region was set to 1. However, we only included simulations with two or more causals. For multi-causal-SNP simulations, we then randomly selected m causal SNPs in the region and simulated chromatin marks and gene expression as described previously, but drew the effect sizes of each SNP as N ( 0 , σ g 2 / m ), such that the total expected variance explained remained at 0.25. For multi-causal-mark simulations, we randomly selected p causal marks in the region and simulated chromatin marks by defining the means, Hc, of each causal mark independently as described for the single-causal simulations. We then computed gene expression by drawing the effect size, βh, of each causal mark from N ( 0 , σ g 2 / p ) such that the total expected variance explained remained at 0.25.
We benchmark our method against five alternative approaches. Firstly, we compare against the standard overlap analysis whereby hQTLs and eQTLs are independently identified within a region centered around a gene. We follow the protocol outlined in [14]. In this experiment, we computed the best SNP association in each region with every mark measured in the region as well as with the gene expression value for that region. We determined adjusted p-values for each top association by performing permutation tests. We then accounted for multiple testing at the mark level by determining the minimum FDR at which each adjusted p-value would be considered significant. This was estimated via the qvalue package [27]. This procedure resulted in a set of significant SNP-mark associations, as well as one SNP-expression association within the region, as only the top SNP association is retained for each biological phenotype. We then evaluated the number of causal SNPs, marks, and paths that were ultimately included in these candidate sets.
Secondly, we compared against the approach of independently fine-mapping the two levels of data (SNP-mark and mark-expression), and multiplying together pairs of posterior probabilities to produce probabilities of causality for paths. For these independent fine-mapping experiments, we used a simple approach that assumes a single causal variant, approximating posterior probabilities for causality directly from Z-scores [28].
In addition, we compared against a basic ranking approach, where we independently computed SNP-mark, mark-expression, and SNP-expression associations for every SNP and mark within a region. For SNP and mark prioritization, we simply produced a ranking of the SNP-expression and mark-expression posterior probabilities for causality, respectively. For path prioritization, we produced a ranking of the product of SNP-mark and SNP-expression posterior probabilities.
We next compared against a bayesian network model which computes directed association strengths between all possible pairs of nodes in a given network [20]. The method takes as input raw genotype and phenotype values. As nodes, we included all SNPs and marks, as well as the gene expression value, within a region. We allowed only for node pairings directed from SNP to mark or from mark to gene expression. For SNP and mark prioritization, we ranked association strengths over all directed SNP-expression edges and mark-expression edges, respectively. For path prioritization, we produced a ranking of the product of SNP-mark and mark-expression strengths.
Finally, we compared against Coloc, which is designed to identify SNPs that are likely to be causal for multiple traits at once. Specifically, Coloc outputs a posterior probability that a SNP is causal for two arbitrary traits simultaneously. We adapted Coloc for our purposes by running the method on all SNPs independently. For each SNP, the two given traits were (1) gene expression, and (2) a mark value. Thus, we ran Coloc independently for all SNP-mark combinations. This produced a set of posterior probabilities indicating, for each SNP-mark combination, the likelihood that the SNP is causal for both the mark value and gene expression simultaneously. For path prioritization, we ranked these probabilities over all SNP and mark combinations. For SNP and mark prioritization, we marginalized over all marks and SNPs, respectively, producing posterior probabilities for each SNP and mark to be causal independently.
The real data analyses were done on 65 YRI individuals whose genotypes were obtained through 1000 Genomes and standardized. PEER-normalized [29] H3K4me1, H3K4me3, H3K27ac, DHS, and RNA expression marks in lymphoblastoid cell lines (LCLs) for these individuals were obtained from [10]. For each gene in the dataset, we computed associations for every SNP-mark, SNP-gene, and mark-gene pair within a 50kb window centered around the gene TSS. On average, each region contained 160 SNPs and 25 marks (across the four mark types—H3K4me1, H3K4me3, H3K27ac, and DHS—whose peak values we analyzed together in each region). Overall, from 14,669 50kb regions, we filtered for regions that exhibited evidence for our sequential model where SNPs affect chromatin marks, which in turn affect gene expression. Specifically, for each region we performed a two-stage regression where we first regressed gene expression on all chromatin marks, and (2) regressed the proportion of expression explained by the chromatin marks on each SNP. If at least one SNP had a low p-value for association (p < 0.05/n.snps) to the proportion of gene expression explained by chromatin data, we kept this region for our real data analysis. After this filtering procedure, we retained 1,317 regions.
We obtained motif annotations from HaploReg [25] and ChromHMM annotations from the NIH Roadmap Epigenomics Consortium [30]. When comparing annotations of top prioritized paths with those of random paths, we established corresponding background paths by choosing a random SNP/mark combination at every region where a top path was reported.
For GWAS analyses, we explored regions whose tag SNP was associated to an autoimmune trait with p < 5 × 10−8. Associations were obtained from recent literature for eight autoimmune phenotypes [31–36]. For each of pathfinder’s top reported paths, we determined whether the corresponding SNP was contained within any of the GWAS regions in our dataset. In order establish a null distribution for this statistic, we ran the same analysis for random regions in the genome not overlapping with the GWAS regions in our dataset. Specifically, for each GWAS region, we randomly selected a SNP in the same chromosome matched for MAF (ϵ = 0.01) and LD score (ϵ = 0.001) with the GWAS tag SNP. We established a window around this matched SNP corresponding to the window size of the GWAS region. Finally, we determined the number of top paths that fell within these random regions. We repeated this experiment 100 times to establish the null distribution of this measurement and calculated a p-value using a Z-test.
|
10.1371/journal.pcbi.1002729 | The Dynamics of Naturally Acquired Immunity to Plasmodium falciparum Infection | Severe malaria occurs predominantly in young children and immunity to clinical disease is associated with cumulative exposure in holoendemic settings. The relative contribution of immunity against various stages of the parasite life cycle that results in controlling infection and limiting disease is not well understood. Here we analyse the dynamics of Plasmodium falciparum malaria infection after treatment in a cohort of 197 healthy study participants of different ages in order to model naturally acquired immunity. We find that both delayed time-to-infection and reductions in asymptomatic parasitaemias in older age groups can be explained by immunity that reduces the growth of blood stage as opposed to liver stage parasites. We found that this mechanism would require at least two components – a rapidly acting strain-specific component, as well as a slowly acquired cross-reactive or general immunity to all strains. Analysis and modelling of malaria infection dynamics and naturally acquired immunity with age provides important insights into what mechanisms of immune control may be harnessed by malaria vaccine strategists.
| Human malaria infections resulting in serious complications and death occur predominantly in young children, and resistance is gradually acquired with repeated exposure. Malaria parasites have two major stages within the human host during its life cycle: an initial liver stage, and the subsequent blood stage, where parasites replicate in and destroy red blood cells. The mechanisms of acquired resistance to severe malaria may involve immunity to both the liver and blood stage parasites. However the relative contribution of each type of immunity is not yet understood. To gain novel insight, we have analysed data from a malaria exposed cohort from western Kenya. We used mathematical modeling to understand what form of immunity is consistent with the observed rates of reinfection in adults and children seen in the field study data. We found that the reinfection pattern can be completely explained by blood stage immunity. Moreover, the blood stage immunity must consist of rapidly-induced strain-specific immunity that clears individual infections, and general immunity that accumulates slowly and decreases the average parasite growth rate with age. Understanding the dynamics of naturally acquired immunity and infection provides important insights for effective vaccine development.
| Plasmodium falciparum (Pf) malaria in holoendemic areas is characterized by high level parasitaemia and symptomatic infections in early childhood, followed by the development of semi-protective immunity that allows the persistence of low level asymptomatic infections and appears to reduce the likelihood of becoming infected if bitten by an infective mosquito. The mechanisms mediating anti-infection and anti-disease immunity are complex but are thought to include innate and adaptive immune responses that limit both the liver and blood stage of the parasite life cycle in the human host [1], [2].
One approach to understanding acquired immunity to Pf-malaria has been to study correlates of protection by measuring point-prevalence levels of immunity and prospectively assessing the infection status and clinical disease. A number of studies have elaborated upon this approach by first treating patients with anti-malarial drugs to eliminate blood-stage malaria infection, and then observing the time to natural (re)-infection in an endemic setting [3]. By measuring immune parameters at baseline, and observing their association with time-to-infection, it may be possible to identify the immune responses most important for protection from malaria. A significant difficulty with these studies is that most immune responses to malaria increase with age and after cumulative exposure to malaria antigens, and so it is often unclear whether measured responses actually mediate protection or are merely a historical marker of past exposure [4], [5]. Both antibody specificity and isotype may play a role in protection [6], [7] and indeed the specific assay used to measure immune function can lead to contradictory or inconsistent conclusions. For example, antibodies that are reactive in an ELISA assay tend to increase with age, but are often not correlated with protection when corrected for age [8]. Antibodies detected using a functional assay that measures inhibition of parasite growth show an increase with age, and an association with protection from clinical disease, but not from infection [9]. Surprisingly, growth inhibitory antibodies (that can restrict parasite growth in vitro) are associated with a delay in time-to-infection for individuals within a given age group, but the level of inhibition decreases with age [3], [10]. Independent of these experimental studies, modeling of malaria infection has also attempted to understand the possible cross-reactivity and molecular targets of malaria immunity, using a heuristic approach based on a qualitative assessment of the data
An alternative approach to understanding anti-malarial immunity is to study the dynamics of infection and then predict how these dynamics would be influenced by acquired immunity. That is, by comparing the infection dynamics observed in susceptible (children) in contrast to resistant (adult) populations, we can use a reverse-engineering approach to understand the differences observed in infection and growth, and predict what immune mechanisms are compatible with such an outcome. Here we use such a modeling approach to understand the effects of naturally acquired immunity on the dynamics of P. falciparum malaria infection in a cohort of 197 people from a holoendemic region of western Kenya.
The details of cohort study have been described elsewhere [3]. Briefly, upon entry into the study (day 0) study participants (n = 201) were treated with Coartem®, which is expected to eradicate blood stage infection but is not effective against liver stage parasites [11]. Therefore, if a study participant was parasitaemic at week two post-treatment, this was considered an emerging liver stage infection and they were eliminated from further study. After treatment, blood smears were monitored weekly for 11 weeks by examination of thick and thin blood smears using light microscopy for the presence of Pf-malaria parasites. In addition, if weekly samples were not collected after week two post-treatment, then the study participant was eliminated; thus leaving 197 remaining for analysis. The cohort was divided into four age groups based on the immuno-epidemiology of malaria; C1 (children 1–4 years old (y.o.) were the least likely to have developed anti-malarial immunity and thus most susceptible to rapid infection; C2 (5–9 y.o.) have begun to develop partially protective immunity though may have limited strain-specific immunity; C3 (10–14 y.o.) have developed anti-malarial immunity that should begin to contain a broader repertoire of strain-specific immunity due to cumulative exposure; and A (adults>15 y.o.) who have developed anti-malarial immunity that decreases the rate of infection, parasite growth and provides coverage across many strains. Fig. 1A shows the infection curves for each age group over the 11 weeks of the study.
The early stages of infection involve an infected bite, infection of a liver cell, and maturation of the liver stage parasite over 6–7 days [12]. After this, the infected liver cells rupture, releasing approximately 20,000 merozoites [13]. These merozoites then initiate the blood stage infection, and successive rounds of blood stage replication follow, each lasting approximately 2 days. These earliest events during human liver and early blood stage infection are not measured in our cohort, and we first detect parasites when their density in blood reaches our detection threshold (∼40 parasites/µl). Both the rate of initiation of blood stage infection as well as the rate of growth of blood stage immunity will affect the time to first detection of infection (Fig. 1B). Liver stage immunity acts to block some infected bites from reaching the blood stage, and thus the initiation of the blood stage infection will occur with a frequency less than or equal to the real infected biting rate, and occur approximately seven days after the infected bite. Once blood stage infection is initiated, parasites grow at a rate determined by how many merozoites successfully invade new RBCs, and grow to maturity over the two-day life-cycle. We might consider this the ‘parasite multiplication rate’ (PMR), which reflects the fold increase in parasitaemia over the two day life cycle of an infected RBC. Thus, the concentration (C) of parasites during their growth in the blood stage can be described by formulawhere A is the initial concentration of parasitized RBC in blood produced by the liver stage (estimated elsewhere [13]), r is the PMR and t is the time in days passed after initiation of blood stage. Incorporating the rate of initiation of blood stage infection (k), we expect the proportion of uninfected individuals remaining at a given time (S(t)) to follow the equation:(1)where T is the detection threshold and τ is the first possible moment of blood stage infection after treatment.
In our cohort, the youngest children are the most-susceptible to infection (have the least immunity), and thus we use this group as a baseline from which to observe the effects of age-acquired immunity in the older cohorts. The infection curve of the youngest children seems to conform well with the simple dynamics described in equation (1) (Fig. 1.C). Thus, assuming that the earliest initiation of the blood stage (τ) is day 7 (due to the pharmacodynamics of lumafantrine), we found the rate of initiation of blood stage infection (k) at 0.066/day, with 95% confidence interval (CI) of (0.056, 0.076), and PMR (r) of 5.9, CI (3.145, 8.671) for the youngest age group. Using this fit to the infection of the youngest children, we then attempt to find what mechanisms of immunity can alter this infection curve to produce the curves observed in older age groups.
Previous work suggests that immunity may act anywhere in the pre-erythrocytic stage to block sporozoite invasion [14] or kill infected liver cells. Wherever it acts, the primary results of pre-erythrocytic immunity will be to (i) reduce the proportion of infected bites that lead to blood stage infection (reduce k), and/or (ii) to reduce the number of infected liver cells that successfully mature to release merozoites and initiate the blood stage (reduce A).
From a modeling perspective, anything that reduces the initiation of successful blood stage infection has the same effect – it simply decreases the slope of our reinfection curve (which remains in exponential form). Fig. 2.A. illustrates the predicted effects of decreasing the rate of initiation of blood stage infection and shows the best fits of a model that incorporates this change to the cohort data. The fits are essentially a family of curves commencing at the same time, and decreasing at different rates. The best fit parameters and goodness of fit statistics are in the [Text S4], (model 1). This is quite different from the observed infection rates in two respects: (a) it does not capture the greater delay until infection is observed in older children and adults, and (b) it does not capture the convex part of the adults' and older children's infection curve.
One may think that greater delay in detection of blood stage infection in adults may be explained by other factors such as increased blood volume in adults, or a decrease in the number of infected liver cells surviving to maturity in adults (due to the effects of cytotoxic T cell immunity killing [13] or growth inhibition [15]). However, this is not adequate to explain the observed delay. The number of infected liver cells is rarely higher then 10–14, [13], [16], [17], and the blood volume differs by only a factor of approximately 5, thus even combined they could explain at best a delay of 4.7 days (using the PMR of the youngest age group 5.9). It is clear that, in addition to underestimating the delay, pre-erythrocytic immunity alone fails to account for the change in shape of the curve associated with age and cumulative exposure.
Once the blood stage of infection is successfully initiated, immunity may act on either the growth of the parasite within the infected RBC or directly on the released merozoites to reduce the rate of successful subsequent invasion events. Immunity acting at these stages will have the effect of reducing the parasite multiplication rate, and reduced PMR will lead to a greater delay until infection is detected. However, reducing the PMR alone affects the infection curves simply by shifting the curves to the right. That is, if the rate of successful initiation of the liver stage is unaffected, the effect is only to change the time until parasite density grows above the threshold but not to affect the infection rate.
We attempted to fit our model assuming only the PMR changed with age (Fig. 2.B). The best fit parameters and goodness of fit statistics are in the [Text S4], (model 3).. It is once again clear that although growth-inhibiting immunity can account for the delay in infection, it fails to capture the shape of the infection curve. Even combining the effects of liver and blood stage immunity fails to capture the observed infection curves. For example, in the adult infection curves, no combination of liver and blood stage immunity can account for the delayed, slowly starting shoulder of the infection curve, followed by an early period of high infection, followed by a further slowing in infection rate (Fig. 1).
The simple models above consider the effects of a fixed growth rate within a given age group. However, the growth rate of parasites is likely to vary between individuals as a result of a combination of parasite strain and host factors such as intrinsic immune maturation [18], [19]. Therefore, rather than consider all infections in an age group having the same growth rate [PMR = 6, for example], we investigated the effects of variation in the parasite growth rate around the population mean within each group. For simplicity we have assumed a normal distribution in PMR, and the same standard deviation (as a proportion of the mean) for each age group. We note that it is not essential that all groups have the same coefficient of variation, however fitting the same parameter to all reduces the number of parameters fitted by three (See [Text S1] for a complete description of the mathematical model).
Using a simple assumption that all age groups have the same biting rate and the same type of positive normal distribution of growth rates (with the standard deviation the same proportion of the mean PMR) we then fitted a theoretical infection function (2) to the observed infection proportion, allowing only the mean PMR to vary between age groups (note that a higher biting rate in adults has been suggested by some [20] and can also be accommodated in the model).(2)g (.) is the probability density function of the delay to detection,(3)fN(μ,pμ) (x) is the probability density function of a normal distribution with parameters μ and pμ, and μ is the mean PMR in a given age group. Constant p is a positive number same for all age groups. FE(k)(x) is the cumulative density function of an exponential distribution with parameter k – the average number of bites per day, which is also the same for all age groups.
The derivation of the theoretical infection function that incorporates distribution in the time to infective bites and PMR within the host for different age groups is shown in [Text S1]. The best fit parameters and goodness of fit statistics are in the [Text S4], (model 4).
Remarkably, allowing only the average PMR to vary between age groups captures the main features of the natural infection profiles. We see both the increasing delay with age, as well as the early rapid increase in infection rates followed by an apparent slowing down of the rate of infection (Fig. 3A). With this model of a distribution in PMRs, we can now understand the unusual shape of the adult infection curve: The early slow shoulder of the curve represents the small fraction of infections that grow rapidly, and are detected early (the right hand side of the distribution in Fig. 3B). The rapid phase of infection is around the mean of the PMR curve, and the apparent slowing represents the very slowly growing infections, which are not detected during the 11 weeks of analysis. Because of the distribution of growth rates, there is a proportion of infections where the PMR <1 (shaded in Fig. 3B), implying the number of parasites decreases at each round of infection (each currently infected RBC infects less than one RBC in next round). In children, with mean PMR of ≈3.8, only a small proportion of infections have PMR <1. However, for the adults, with mean PMR ≈1.35, a large proportion of bites (≈24%) have PMR <1, which is why we see an apparent slowing of the infection rate later in the study. We note that by contrast simply allowing a distribution of the level of liver stage immunity alone does not improve the fit of the liver stage model (see [Text S3] for a detailed description of the model ).
Our analysis suggests that a model that can fit the data well is one that assumes a distribution in PMRs within each age group, and a decrease in mean PMRs with age. A key question that follows is – could such a distribution with age arise from a known mechanism of immunity? Similarly, we have considered above only the relationship between PMR and time to detection of infection. However, decreased PMR would also decrease the observed levels of parasitaemia with age. Is the change in PMR required to produce the reinfection curves consistent with the change in parasitaemia levels with age? To answer this, we developed a stochastic model of within-host immunity and parasite growth to explore the effects of naturally acquired immunity. We focused on the impact of immunity affecting parasite growth, and further allowed that such immunity might either be strain-specific, or ‘general’ (affecting all strains equally). In this context, a strain-specific immunity may be defined by expression of variable proteins such as var genes or immunologic targets proteins such as merozoite or sporozoite surface antigens) that can vary across strains. General immunity is a response (be it immune or physiological – such as hyper-splenism) that acts equally on all strains. A number of previous models of malaria immunity have considered the effects of partially cross-reactive immune responses [21], [22], [23], [24]. However, we effectively model only the extremes of ‘completely strain-specific’ or ‘completely cross reactive’ responses for simplicity.
In our model, strain-specific immunity arises as the result of the infection with a certain strain and neutralises the parasites of this strain only. General immunity arises as the result of infection with any strain and it can neutralise parasites of any strain. The acquisition rate of both types of immunity is proportional to concentration of parasites in the blood. Without infection, the existing level of immunity decays at some constant rate. Thus the model has four parameters of immunity: for both strain-specific immunity and general immunity we have a rate of increase in immunity per unit parasite (denoted by α and γ respectively), and a rate of loss of immunity in the absence of parasite denoted (denoted by β and δ respectively). We also require three parasite parameters: the mean number of bites per day k (biting rate), the baseline multiplication rate r of the parasites, and the number of different strains n. The basic equations of parasite – immunity dynamics of the model are:(4)Here Pi is the concentration of parasites of the strain i, Si is the strength of the i th strain specific immunity, . G is the strength of the general immunity. The function h(.) is either one (if the concentration of parasites<threshold) or zero (if the concentration of parasites≥threshold), and allows for the elimination of parasites and the decay of immunity when the level of parasites drops below a threshold Z (in this case 0.005 parasites/microlitre ).
A detailed description of the mathematical model is in the [Text S2].
Using this model we can simulate the ‘life history’ of an individual in a malaria endemic area. In the model, we start with the assumption that all individuals are identical and lack immunity at birth (t = 0), and are then exposed to bites from different strains arriving at random times, with a random order of strains (choosing from our 50 notional strains), to which they then develop immunity (ignoring the transient contribution of maternally-derived immunity). Thus, the timing of bites and which strain is inoculated is stochastic, but between bites the dynamics of parasites and immunity is deterministic (as in equation 2).
Fig. 4 shows the dynamics of parasite infection and acquisition of strain specific and general immunity by an individual, starting from birth to one year [left panel] and five to six years [right panel]. The top panels show the dynamics of parasite infection, as different parasite strains (indicated by different colours) initiate blood stage infection, grow, and then induce strain-specific immunity, leading to their clearance. Given the long half-life of strain-specific immunity and the absence of within-host parasite antigen variation in our model, each new parasitemia peak represents infection with a new strain. In addition to inducing strain-specific immunity (bottom panels, coloured), these infections also induce general immunity (solid black line), which accumulates over time. By simulating the life history of a small population of individuals (n = 50), we can then apply the concept of ‘treatment-time-to-infection’ trials to the simulated individuals. That is, by removing all blood stage parasites of individuals in different age groups and observing the time until parasite levels reach our detection threshold, we can simulate (re)-infection (Fig. 5 and 6). Fig. 5 shows the dynamics of parasitaemia during (re)-infection from the field study data (top) and the simulation (bottom), for four subjects from each of the different age groups. Remarkably, the simulation captures a number of the factors observed in our observational study; firstly, the natural infection curves show increasing delay with age and an increasing proportion of individuals remaining uninfected. Secondly, the observed reduction in parasite levels in blood with age is also captured in the model, indicating that the decreased PMR required to produce the reinfection curves is consistent with the decreased PMR required to produce the observed reduction in parasitaemia with age. (as higher immunity means parasites are controlled at a lower parasitaemia) (Fig. 5). Fig. 6A shows the infection curves for the whole simulated population, and Fig. 6B compares the mean parasitaemia for the field study data and simulation for different ages. Importantly, the major factor that can account for both delayed infection and lower parasitaemia in adults is simply a reduced average growth rate of parasites with age and naturally acquired immunity.
A number of modelling studies have previously been applied to understanding the dynamics of malaria infection and the impact of immunity. Some of these studies have focused upon the dynamics of experimental infection of neurosyphilis patients in the USA during the early 20th century with both P. falciparum [25] and P. vivax [26]. Such infections with P. falciparum often show a dynamic of repeated recrudescence, in the absence of reinfection. Modelling both infection rates and infection dynamics has also been applied to more recent data on infection in naturally exposed individuals [27], [28]. The recent discovery of the molecular mechanisms of antigenic variation in P.falciparum has also driven studies modelling the dynamics of immune interaction with an antigenically variable pathogen [21], [22], [23], [24]. In our study, we utilise field data on malaria infection dynamics after a short treatment course in a holoendemic region of Kenya. The estimated P. falciparum infection rates are extremely high, especially amongst young children, but consistent with the high entomological infection rates measured in this area [29]and with the blood-stage infection rates previously observed by others in similar studies [30], [31]. The prior treatment of patients gives an ideal opportunity to study malaria incidence in the absence of recrudesence from prior infection. However, there is a potential problem equating the measured time to reinfection in treated, asymptomatic individuals with incidence rates of new blood-stage infection in individuals living in highly endemic areas, and then using these findings to make general statements about acquired immunity to malaria. However, given that children in this area are treated on average twice a year [32], this is likely as natural an infection dynamic as can be readily studied.
Fitting of a simple model of infection dynamics to the field data on P. falciparum infection demonstrates that the observed rates of infection with age are consistent with a form of immunity that reduces the PMR of the parasite in blood. This mechanism of blood-stage immunity reducing PMRs is consistent not only with the dynamics of infection after treatment in different age groups, but also with the observed reduction in peak parasitaemia with older age groups (Fig. 6). Although clinical malaria was not studied in this model, one might imagine a reduction in parasite replication with age may also play some role in both the observed reduction in episodes of severe malaria as well as an apparent shift from acute to chronic infections with age. The fact that the model fits well without invoking a need for liver-stage immunity does not exclude a role for liver stage immunity, but suggests it is not a major force shaping the observed infection dynamics. By contrast, liver-stage immunity alone would be expected to have relatively little impact on either the delay or peak parasitaemia. The estimated PMR in this study and their change with age are very consistent with published studies of parasite growth in unexposed adults in the UK (estimated to have a high PMR of approx 8–14/cycle [12], [13]), compared with the reported very low PMRs in malaria endemic areas (1.6–3/cycle) [33], [34].
Although the initial model was developed on the basis of a number of heuristic assumptions, the model provides some useful insights into the possible mechanisms of blood-stage immunity induced from natural malaria infections. Firstly, ‘parasite clearing’ immunity must be highly strain-specific: That is, we often see one parasite peak being cleared immediately before another peak arises (Fig. 5, C1, top). Given the time taken for clearance and growth, this means that we must have one parasite being cleared while another grows. Thus the immune response mediating clearance of individual parasite peaks must have the characteristics of being both highly strain-specific, and rapidly induced. Although this strain-specific response must be rapidly induced, the duration of strain-specific immunity may be varied in the model, depending on the number of strains used and the biting rate. That is, if there are only a small number of immunologically distinct strains of parasite, then the duration of immunity needs to be short – to allow reinfection. However, if there are a large number of immunologically distinct strains, the half-life of strain-specific immunity could be considerably longer and still consistent with the data. Thus we are not able to speculate on the relative duration of strain-specific versus general immunity, as has been discussed elsewhere [21], [22], [23], [24].
The long timescale over which age-related immunity is acquired suggests that whatever factor mediates this must have a long half-life (since the time taken for the immunity level to reach its maximum is related to the half-life of immunity). If specific immunity were the only mechanism, then it would need to be long-lived, and the age-associated acquisition of immunity would simply be the progress from having immunity to a few strains, to having immunity to most strains. However, the intermediate phenotype here (having immunity to a proportion of strains) is incompatible with the age and parasitaemia data. That is, if we had immunity to 50% of strains, and no immunity to the other 50%, this would mean that 50% of bites would be blocked (changing the reinfection curve as in Fig. 2A) and that parasitaemia of the remaining strains would be ‘normal’. This is clearly not consistent with the data, suggesting that although strain-specific immunity is required to explain the clearance of individual parasite peaks, it is not sufficient to explain age-associated changes in reinfection and parasitaemia.
The alternative to a strain-specific immune response to individual infection episodes is a general anti-parasite response that decreases the growth of all strains. In contrast to specific immunity, general immunity takes many years to acquire and accumulates after repeated infections. Thus, each infection should both induce only a small rise in general immunity, and it must be long-lived. It is also clear that general immunity alone cannot produce the observed infection dynamics, as this could not simultaneously drive one parasite strain down while the other grows. In broad terms, we might consider that an increase in general immunity with time drives down the mean PMR of an age group. The distribution in growth rates within an age group arises due to strain-specific immunity, with individual strains (in different individuals) having different growth rates due to the variable levels of strain-specific immunity, which wanes over time after infection with that strain.
Clinical studies of immune correlates of protection from malaria are often complicated and difficult to reconcile (reviewed in [8]). Overall, antibody levels have been found to generally increase with age and exposure, when measured by the ability to bind parasite antigens in ELISA. It has generally been difficult to demonstrate a clear correlation with protection from parasitaemia, as this is so strongly confounded by age. These antibody levels are also long-lived, appearing to persist with a half-life up to 5–10 years [35]. However, it is difficult to differentiate between binding antibodies exerting a direct protective effect, or merely acting as biomarkers of past infections.
Our model does not address the particular mechanisms or specificity of immunity. However, one prediction of the model is that each patent infection is cleared by a highly strain-specific response, as it is common that the clearance of one strain occurs while infection with another strain is still growing. This is consistent with early studies suggesting that growth-inhibitory antibodies may be highly strain-specific when tested in clinical isolates [36]. Interestingly, some growth-inhibitory antibodies appear as a correlate of protection within an age group, but paradoxically decrease with age [3], [10]. It is interesting to note that strain-specific immunity also paradoxically decreases with age in our model, due to the increase in general immunity.
We note some difficulties with defining the term ‘strain’ in malaria, as recent deep sequencing studies have suggested a wide range of genotypes [37]. Moreover, even once the genetic structure of malaria strains is resolved, the immunological specificity and cross-reactivity of responses may be highly complex. We do not use the term ‘cross-reactive immunity’ when describing a factor that limits the growth of all strains, but instead prefer the term ‘general immunity’. The concept of cross reactivity may tend to imply specific antibodies that cross-react between different strains. However, we do not wish to exclude other factors such as an enlarged spleen may tend to decrease the PMR of all strains, and may also play a role in protection.
The majority of cases of severe malaria occur amongst young children in endemic areas, and age and persistent exposure provide some level of immunity to both clinical symptoms, as well as measured parasitaemia [38]. Although this age-associated acquisition of immunity is known, clinical studies of immune responses to malaria often struggle to identify the mechanisms of protection [38]. We have employed an alternative strategy in an attempt to deconstruct the mechanisms of naturally acquired immunity based on the dynamics of infection in otherwise healthy asymptomatic individuals. We find that simple pre-erythrocytic immunity acting alone is inconsistent with the data, as is a highly strain-specific form of blood-stage immunity, as both would have the overall effect of merely decreasing the number of successful infections. Instead, the observed dynamics of infection and parasitaemia levels in a cohort of individuals from a holoendemic area of western Kenya are consistent with the reduction in parasite multiplication rate with age, leading to both a delay in the time until infection is detected, as well as reduce the peak parasitaemia. In terms of the clinical outcomes observed, our study suggests that the reduced rate of clinical malaria and reduced rate of apparent infection (measured at a particular threshold of parasitaemia detection) may both be driven by a reduction in parasite growth rate, independent of any real change in infection rate. Further studies are required to deconvolute immunity that inhibits parasite growth distinct from its ability to infect hepatocytes and erythrocytes.
|
10.1371/journal.pcbi.1006324 | The role of curvature feedback in the energetics and dynamics of lamprey swimming: A closed-loop model | Like other animals, lampreys have a central pattern generator (CPG) circuit that activates muscles for locomotion and also adjusts the activity to respond to sensory inputs from the environment. Such a feedback system is crucial for responding appropriately to unexpected perturbations, but it is also active during normal unperturbed steady swimming and influences the baseline swimming pattern. In this study, we investigate different functional forms of body curvature-based sensory feedback and evaluate their effects on steady swimming energetics and kinematics, since little is known experimentally about the functional form of curvature feedback. The distributed CPG is modeled as chains of coupled oscillators. Pairs of phase oscillators represent the left and right sides of segments along the lamprey body. These activate muscles that flex the body and move the lamprey through a fluid environment, which is simulated using a full Navier-Stokes model. The emergent curvature of the body then serves as an input to the CPG oscillators, closing the loop. We consider two forms of feedback, each consistent with experimental results on lamprey proprioceptive sensory receptors. The first, referred to as directional feedback, excites or inhibits the oscillators on the same side, depending on the sign of a chosen gain parameter, and has the opposite effect on oscillators on the opposite side. We find that directional feedback does not affect beat frequency, but does change the duration of muscle activity. The second feedback model, referred to as magnitude feedback, provides a symmetric excitatory or inhibitory effect to oscillators on both sides. This model tends to increase beat frequency and reduces the energetic cost to the lamprey when the gain is high and positive. With both types of feedback, the body curvature has a similar magnitude. Thus, these results indicate that the same magnitude of curvature-based feedback on the CPG with different functional forms can cause distinct differences in swimming performance.
| When animals move, they receive sensory inputs, which in turn are used to modulate the movement. Relatively little is known about how these inputs affect performance during steady locomotion. Using a computational model of a swimming lamprey, we investigated two different types of feedback, both consistent with experimental data. Both have strong, but different, effects on swimming speed and energy consumption, suggesting that sensory feedback is crucial not just for responding to perturbations, but also for high performance steady locomotion.
| To move effectively, all animals must activate their muscles to move their bodies. In nearly every animal studied, this pattern of muscle activation is produced by a relatively small neural circuit called a central pattern generator (CPG) [1, 2], which is influenced by the mechanics of their bodies and the physical world around them. The CPG integrates sensory information, particularly from proprioceptors that sense body movement, and adapts the pattern of muscle activation accordingly [3].
The lamprey has been a model organism for investigating sensorimotor integration during locomotion because its spinal cord, which contains the CPG circuit, can easily be isolated and stimulated to produce a pattern of muscle activity that strongly resembles the pattern in intact swimming animals [4]. Moreover, the primary proprioceptors in lampreys are mechanosensory cells, called edge cells, and are located on the spinal cord [5], in contrast to the proprioceptive muscle spindles and Golgi tendon organs of mammals, which are located in the periphery [6]. These edge cells synapse onto several different classes of ventral horn interneurons that make up the CPG [7]. There are two classes of edge cells: one that inhibits contralateral activity and another that excites ipsilateral activity [7]. Through these connections, edge cells can entrain the CPG to a bending input [5, 8–10] and reset the rhythm after a brief perturbation [11]. In land animals, muscle spindles have similar effects on the CPG [12].
Previous studies of edge cells, like those for many other mechanoreceptive sensory cells, have examined their effects in open loop conditions. In these experiments, researchers record the response to different types of mechanical stimuli (e.g., [10, 13]) or they record the effect of edge cells on the CPG due to mechanical inputs (e.g., [8, 9, 11, 14]). However, when a lamprey swims, the edge cells are part of a closed loop system [15]. The CPG activates the muscles, which causes the body to bend, stimulating the edge cells, which then affect the CPG (Fig 1). Edge cells, therefore, not only allow the animal to respond to unexpected perturbations outside the normal pattern of activity, but they can shape the pattern itself. For example, they can cause the rhythm to entrain to a mechanical resonance [16, 17], which may contribute to more efficient swimming.
Here we describe a multiscale computational model of the lamprey closed-loop neuromechanical system. In previous work [16, 18–20], we developed a computational framework that includes a Hill-type muscle model [21] with calcium dynamics [22–24] that drives a passive structural model coupled to a viscous, incompressible Newtonian fluid using an immersed boundary formulation [25–27]. These previous studies examined the effects of varying the stiffness of the body [20], the frequency of activation [20] and the effects of muscle nonlinearities [19]. We found that the same muscle dynamics result in different kinematics when body stiffness is varied, with maximum speed and maximum acceleration occurring at different stiffnesses. Changing the frequency of muscle activation resulted in a change in active lengthening, a phenomenon observed in natural fishes and thought to be related to efficient swimming [20]. Removing non-linear dynamics known to be a part of the muscle force development indicated that without these non-linearities, the same body kinematics can be achieved with very different energy requirements [19]. This previous work focused on the kinematics and effects from changing different aspects of the mechanical parts of the body. Our current study focuses on modeling the central pattern generator and sensory feedback in an effort to understand how information from body mechanics might be communicated to the electrical signal driving the motion.
Other neuromechanical models of undulatory swimming have been developed (e.g., [28, 29]). Iwasaki, Chen and Friesen [30] studied how sensory feedback modulates oscillations in the central pattern generator in leeches. They include models of the CPG, motoneurons, muscle dynamics, sensory feedback and body-fluid interactions, and show robust swimming with adaptivity from sensory input. Both leeches and lamprey exhibit sensory feedback through stretch receptors. Iwasaki et al [30] showed that this feedback in leeches was an essential component in leech swimming. Ekeberg and Grillner [28] used a detailed Hodgkin-Huxley model of neural activation, but a resistive force model, while Gazzola et al. [31] used a fluid model based on elongated body theory with a simple activation wave and proprioceptive feedback modulation. However, because the resistive model of fluid mechanics does not accurately capture the timing or magnitude of forces on the body [20], we chose to incorporate a high-fidelity Navier-Stokes fluid model that can resolve bending dynamics of the lamprey body along with flow features such as the vortex wake shed by the swimmer. Although we are using a two-dimensional fluid model, comparisons with robotic models of anguilliform swimmers and real biological swimmers show excellent agreement in many metrics of flow [18]. The closed-loop model presented here adds a coupled-oscillator representation of the CPG (after [32–34]) with sensory input from stretch receptors that depend on body curvature.
We construct the CPG using chains of nonlinear phase oscillators. Some of the first models to capture key characteristics of experimental studies of the lamprey CPG were phase oscillator models, which define each oscillator based on a single variable, namely its phase [32]. These types of models are simple, yet reproduce some important features of the CPG responses to perturbations [11] along with its entrainment to bending stimuli [8, 33, 35, 36].
Although edge cells have been studied for a long time and their electrical activity is known to be affected by bending, their response properties are only beginning to be characterized. Current experiments indicate that they respond to both the magnitude of bending and the bending rate [13], but their effects on the CPG are not well understood, particularly in a closed-loop system with relatively small deviations from a steady sinusoidal oscillation. Here, we simulate different functional forms of mechanosensory feedback, and investigate how the choice of the feedback model affects swimming performance in the closed-loop neuromechanical system.
Even without feedback, the computational model naturally approximates a subtle feature of fish swimming. When fish are swimming steadily, anterior muscles on one side of the body become active with a small lag after the body is already bending toward that side [37], so that the muscles are always active while shortening and generate positive mechanical power. Although it may sound like this effect is backwards, since muscle force produces curvature, it occurs due to the periodic nature of the oscillation and only appears after several swimming cycles. This phase lag emerges naturally in the computational model [20] and allows the muscles to maintain the ongoing oscillation. Close to the tail, the lag becomes negative, so that muscle activity precedes bending, starting while they are still lengthening under the pressure of fluid forces [37]. This change in phase lag along the body has been suggested to improve the efficiency of fish swimming by stiffening the tail so that anterior muscle power is more effectively transmitted to the fluid [38, 39]. To quantify the lag, we compare the relative wave speeds of the mechanical curvature wave and the neural activation wave. When the curvature wave is slower than the neural activation wave, the posterior muscle ends up being active during lengthening (see, e.g., [23]), as is seen in fishes [37].
We examined two classes of feedback based on body curvature. In the first, which we term directional feedback, the CPG receives an input that contains the magnitude and direction of curvature. We find that this feedback alters the relative timing of left and right side activity, changing the duration of activity without altering the overall frequency. This type of effect is similar to the well-known phasic effects of edge cells in lampreys [11, 40] and of muscle afferents in cats and other tetrapods [3, 12]. In the second class of feedback, termed magnitude feedback, the CPG receives an input that only contains the magnitude of the curvature, not its direction. Depending on the sign of the feedback gain, we find that this type of feedback is either generally excitatory or inhibitory, speeding up or slowing down the overall oscillation without altering the left-right phasing. This feedback produces an effect that is similar to the excitatory effect of edge cells in the lamprey [41, 42], and of proprioceptive input on vertebrate CPGs in general [3].
In summary, we demonstrate that proprioceptive feedback to the CPG that uses both magnitude and direction of body curvature alters the duty cycle of muscle activation, whereas proprioceptive feedback that uses only the magnitude of body curvature alters the frequency of the activation cycle. Below we will show that these two effects, in the full integrated model, alter swimming kinematics and energetics in complex ways. These results show that the model CPG is capable of generating very different responses depending on whether or not the proprioceptive feedback provides directional information.
In this section, we describe each model component (Fig 1) (CPG, muscle, body, fluid and sensory edge cells) and indicate how these subsystems interact in a way that leads to swimming.
As illustrated in Fig 2, the body model of the lamprey includes components that support both passive forces that represent the mechanics of the body and active forces due to contractile muscle segments that bend the body to enable swimming. The signal from the CPG activates the muscles segments. The passive and active forces along the lamprey are coupled to the surrounding viscous, incompressible fluid. Fig 2A shows that the body is comprised of three filaments, one for each lateral side Xj(s, t), j = 1, 2 and one representing the midline X3(s, t). In all simulations presented, the length of the midline of the lamprey was chosen to be L = 12.56 cm in the average range of lamprey lengths used in experiments [22].
There are two forms of tapering in our model. First, the body becomes narrower toward the tail (Fig in S2 Fig) Second, we account for the decrease in the physiological cross-sectional area of muscle in the posterior body. Total muscle force is proportional to its cross sectional area. Therefore, we scale muscle force accordingly in the tapered region, assuming that the 2D representation of the lamprey has elliptical cross-sections out of the plane.
The first eighth of the body represents the passive head region with no muscular activity, while waves of muscular contraction can act on the rest of its length to propel the animal through the fluid. Fig 2B shows an enlarged portion of the body, and illustrates the discretization of the segmented filaments that will be used in an immersed boundary framework. The stiff midline filament is discretized into 640 segments and the two lateral filaments are discretized into 320 segments each. Connections between the nodes along the midline and the crosslinks that connect the midline to the lateral sides are modeled as passive Hookean springs. The use of Hookean springs rather than damped springs is possible without the danger of lateral oscillations due to the inherent damping effects of the surrounding viscous fluid. The links along the lateral sides, enlarged in Fig 2C, represent a model of the muscle and the skin. The muscle model supports active muscle contractions and is based upon the kinetic model presented in the section “Muscle Model” below. The skin is modeled using springs that only resist extension, but not compression, similar to collagen fibers (bottom spring in Fig 2C). Given the individual stiffness constants of the structural springs comprising the discretized lamprey body, its overall average macroscopic bending stiffness E can be computed [20, 43]. This structural model was used in [20] to explore the role of body stiffness in the swimming performance of the lamprey for a prescribed wave of muscle activation (i.e. no CPG). Note that for the simulations presented in this manuscript, we use a fixed bending rigidity (E = 0.76 MPa) throughout.
Muscles are activated by an electrical signal from the CPG. Our previous model of lamprey swimming prescribed a wave of neural activation to simulate the input of the CPG, but this neural activation was not informed by the evolving dynamics [20].
In contrast, here we simulate the full Navier-Stokes fluid dynamics, where fluid effects on evolving body curvature are captured, along with the dynamics of the vortex shedding in the wake of the tail movement. Although we are using a two-dimensional fluid model, comparisons with robotic models of anguilliform swimmers and real biological swimmers show excellent agreement in many flow metrics [18].
The CPG is modeled using a double chain of sinusoidally coupled phase oscillators (one for the left lateral side and one for the right, see Fig 3):
d θ k , i d t = ω + ∑ j = 1 n α i , j sin ( θ k , j - θ k , i - ψ i j ) + α c sin ( θ k , i - θ k * , i + π ) + η ( κ i ) . (1)
Here, θk,i represents the phase of the ith oscillator in the chain, on the kth side, where k = 1 represents the right side, and k = 2 represents the left side. The notation k* indicates the opposite side, in other words, if k = 1, then k* = 2. The natural frequency of these oscillators is denoted by ω in Eq (1), and throughout this work we choose ω = 2π rad/sec.
The ith oscillator on the kth side is coupled to the other oscillators on the same side with strength given by
α i , j = { A a exp ( - | i - j | λ a ) i - j < 0 A d exp ( - | i - j | λ d ) i - j > 0 , (2)
where |i − j| is the number of segments between between oscillator i and oscillator j. As in [36], we choose the parameters Aa = 1.0 rad/sec, Ad = 10 rad/sec, λd = 5, and λa = 40. The asymmetry in the coupling strengths Aa and Ad is consistent with experimental data [36, 44]. We tune the system with a constant phase shift between neighboring oscillators, so that ψ i j = ( i - j ) ψ ¯. Because in these simulations there are 280 segments on each lateral side that comprise the active portion of the lamprey body, we choose ψ ¯ = 2 π 280, which corresponds to the natural phase shifts found experimentally in [40, 45].
The term αc sin(θk,i − θk*,i) in Eq (1) represents the observed antiphase behavior between left and right segments which couples oscillator i on the kth side to oscillator i on the opposite side. The coupling strength αc = 81.87 rad/sec was chosen to be ten times the strongest descending coupling connection. We choose the connection across a segment to be stronger than the lateral connections because it has been observed that activation waves travel in anti-phase, and that co-contraction is not favored in undulatory swimming [36].
Finally, the term η(κi) in Eq (1) captures proprioceptive feedback to the CPG system. Here κi is the curvature of the lamprey midline at segment i, and the choice of the functional form of η, the feedback, will be explored below.
The signal that the CPG sends to activate the muscle segment is modeled based on the phase of the oscillator on that side. If the muscle segment is sent an activation signal by the CPG, free calcium in that muscle segment is taken up by the thick filaments, generating contractile force. As in Williams [22], we use a mass-action kinetic model of calcium dynamics in each muscle segment and couple it to a Hill-type model of muscle force generation, modified based upon data from lamprey experiments. In Hamlet et al. [19], we explored the implications of muscle nonlinearities on the swimming performance of the lamprey model (with a prescribed activation wave taken as input) by varying the dependence of muscle force generation on segment length and shortening velocities. Here, however, we choose the same parameters in the muscle model in each simulation. We briefly describe the elements of muscle force generation as in Hamlet et al. [19].
Each muscle segment is modeled as a modified Hill-type muscle model that consists of a contractile element (CE), that develops force by contracting in response to an activation signal, and an elastic stretch element (SE), that stores and releases energy from force developed by the contractile element. Here, we assume the contractile force depends upon both length and velocity of the stretch, as well as the amount of bound calcium in the system. The stretch element is described by a spring-mass-damper system driven by the contractile force. To account for tapering along the body, the magnitude of the force is scaled by the ratio of the cross-sectional area of the segment and the maximum cross-sectional area of the body.
The contractile force has the form
P c = P 0 α ( v c ) λ ( l c ) C b , (11)
where α is a nonlinear function of vc, the velocity of shortening of the CE; λ is a nonlinear function of lc, the length of the CE; Cb is the bound calcium given by Eq (5); and P0 is a constant scale factor indicating the maximal contractile force. For more details on this muscle model, we refer the reader to [19]. In particular, all simulations described below assume work-dependent de-activation and calcium-dependent passive stiffness of muscle elements, corresponding to the case VLWσ in [19].
The CPG responds to sensory feedback. Here we model proprioceptive (body-sensing) feedback from edge cells. As discussed above, edge cells are mechanoreceptors [2, 5] that sense stretch along the body and send signals that serve to inhibit the contralateral side of the body and excite the ipsilateral side (relative to the edge cell’s position) with increasing stretch. Recent results of Massarelli et al. [33] have shown that edge cells also respond to rate of change of stretch. Here, for simplicity, we restrict our feedback model to depend only on stretch, which is done by monitoring body curvature. Details regarding the functional form of the input from the edge cells to CPG are not known at this time. To represent the feedback from edge cells to the CPG, we add a feedback term in Eq (1) of the form η(κi) where κi is the curvature of the lamprey midline at segment i. As a starting point, we model this as an additive response. At each time step, curvature is calculated along the centerline of the body from head to tail at a point (x(s), y(s)) by
κ = x ′ y ″ − y ′ x ″ ( ( x ′ ) 2 + ( y ′ ) 2 ) 3 / 2 (12)
where ′ denotes derivative with respect to the spatial parameterization s, x is the longitudinal direction and y is the transverse direction. In practice, these spatial derivatives are replaced by finite differences, and we smooth this curvature at segment i by replacing it by the average curvature of the neighboring segments from i − 5 to i + 5. The last 5 segments of the body receive no input from the curvature of the body. The value of curvature varies both along the lamprey body and in time as the body bends into an S-shaped swimming pattern. Positive curvature represents bending to the right. When curvature magnitude is high, one side of the body is stretched and the other is compressed.
We chose to compare two different functional forms of the feedback:
η ( κ i ) = η m | κ i | , (13)
which we call magnitude feedback (denoted M), and
η ( κ i ) = ( - 1 ) k η d κ i , (14)
which we call directional feedback (denoted D), and where i is the segment number, k is the side of the body (1 is left and 2 is right), and ηm and ηd are the constant gain parameters with units of cm rad/sec. We will discuss the implications of these feedback forms in the Results section below.
The neutrally-buoyant lamprey body supports passive forces due to its elastic linkages as well as active forces along the lateral sides due to muscle contractions. The strength and timing of the contractions of individual muscle segments evolve with the CPG, the calcium dynamics, and the Hill-type muscle model that are each coupled to the evolving body shape. We adopt an immersed boundary framework [27] that couples the forces supported along the three filaments of the lamprey body, together denoted by X(s, t), to a surrounding viscous, incompressible fluid:
ρ [ ∂ u ( x , t ) ∂ t + u ( x , t ) · ∇ u ( x , t ) ] = - ∇ p ( x , t ) + μ ∇ 2 u ( x , t ) + f ( x , t ) (15) ∇ · u ( x , t ) = 0 (16) f ( x , t ) = ∫ Γ F ( X ( s , t ) , t ) δ ( x - X ( s , t ) ) d s (17) ∂ X ∂ t = u ( X ( s , t ) , t ) = ∫ Ω u ( x , t ) δ ( x - X ( s , t ) ) d x . (18)
Eqs (15) and (16) are the incompressible Navier-Stokes equations, where u is the fluid velocity field, p is the pressure, ρ = 1 g · cm−3 is the fluid density, μ = 1 mPa · s is the fluid viscosity, and δ is the two-dimensional Dirac delta function. Eq (17) communicates the Lagrangian forces F(X(s, t), t) supported on the lamprey (Γ) to Eulerian forces defined at any point x in the fluid domain (Ω). Eq (18) enforces the no-slip condition that says the velocity of a material point of the lamprey body is equal to the fluid velocity evaluated at that point. Note that the forces F(X(s, t), t) contain contributions due to the deformation of the passive elastic structure as well as actuated muscle contractions.
The immersed boundary formulation in Eqs (15), (16), (17) and (18) uses a continuum description of both the fluid domain and the immersed lamprey body. In practice, the filaments of the lamprey are represented by discrete nodes, and the fluid equations are discretized on adaptive finite-difference grids. To communicate forces and velocities between the Eulerian fluid grid and the Lagrangian nodes of the lamprey, a grid-dependent, regularized approximation of the Dirac delta function is used [25, 27]. Because both the passive elastic linkages and the contractile muscle segments produce equal and opposite forces at their endpoints, the neutrally-buoyant lamprey is a free-swimmer that generates zero total force and torque at each instant of time. We interface our structural model with the parallelized, adaptive mesh implementation of the immersed boundary method (IBAMR) (developed by Griffith [25]) that allows us to achieve the spatial resolution necessary to resolve boundary layers at physiologically-appropriate Reynolds numbers. The adaptive mesh refinement uses a finer fluid mesh near the nodes of the lamprey as well as in fluid regions where a vorticity exceeds a specified threshold. The fluid domain was defined to be 7.5 body lengths long and 3.0 body lengths high. In the simulations presented here, we allowed five levels of refinement, with the coarsest discretization a 32 cell grid in the x-direction over the entire domain, and the finest level effectively a 512 cell grid in the x-direction. No-stress boundary conditions on the computational domain were implemented. The simulations were run for a total of 10 seconds of simulated time which was determined to be a sufficient amount of time for each case to achieve and maintain a steady swimming speed. Each simulation was run on a linux cluster comprised of 8-core 2.4-2.8GHz AMD Operon processors with 128GB of RAM per core. The simulations were each run on 8 processors on a single node and ran on the order of a day to complete 10 seconds of simulated time. Variation in runtimes was due primarily to differences in the evolving adaptive mesh refinements in the Navier-Stokes solver. Initially, the fluid velocity was at rest and the lamprey body was in a horizontal position, with all elastic linkages at their rest lengths. Convergence studies of this lamprey-fluid immersed boundary model were presented in [18].
Because of the coupling between the body and the fluid, the simulation naturally develops a phase lag in which muscle activation comes after body curvature, even without feedback [20]. This effect is due to the periodic nature of the swimming oscillation and stabilizes several swimming cycles after the simulation starts. For the current model, we computed the phase lag between muscle activation and curvature, as shown in Fig 6. We identified the time of peaks in curvature (t c u r v e , i L, t c u r v e , i R) and the closest muscle activation time (t a c t , i L, t a c t , i R) (Fig 6A) for each point along the body and the left and right sides, where i indicates the tail beat cycle number. The phase lag is defined as
ϕ i L = t a c t , i L - t c u r v e , i L t a c t , i + 1 L - t a c t , i L , (19)
for each side of the body, so that a positive value of the phase lag indicates that the muscle is active while shortening and a negative value indicates that muscle is active while lengthening. In this convention, phase runs from 0 to 1.
We then computed the activation wave speed va and the curvature wave speed V. Activation waves travel in anti-phase down the left and right chains of oscillators, at the same speed. For simplicity, we write the equations below for phase θ (chosen on one side) and curvature κ defined continuously on position s and time t. In the analysis, we use central differences to approximate the temporal and spatial derivatives. The activation wave speed is defined based on the oscillator phase θ (Eq (1)),
v a ( s , t ) = ∂ θ ( s , t ) / ∂ t ∂ θ ( s , t ) / ∂ s . (20)
To compute curvature wave speed accurately, it is best to estimate a curvature phase using the Hilbert transform H{⋅} [46], an operator that estimates the complex representation of a real-valued oscillatory signal like curvature κ(s, t), such that |H{κ(s, t)}| is the curvature amplitude and ϕ = ∠H{κ(s, t)} is the phase of the curvature, both at a position s. Then the curvature wave speed is
V = ∂ ϕ ( s , t ) / ∂ t ∂ ϕ ( s , t ) / ∂ s (21)
and the wavespeed ratio is 〈V〉/〈va〉, where 〈⋅〉 indicates a mean over time and space. Specifically, we take the mean over the region s ∈ (0.5…0.8) and over the last 7 tail beats.
When the ratio is less than one, it indicates that the curvature wave is travelling more slowly than the activation wave (as in Fig 6). The difference in wave speed means that the curvature wave increasingly lags behind the activation wave as the two move down the body, with the phase lag then becoming larger in magnitude and more negative toward the tail, the same pattern seen in fishes [37] and in our previous model [20].
As a baseline control case, we examined our model without feedback. The muscles in the model are activated by a CPG, which is approximated as a set of phase oscillators that receive input from each other and from the body’s curvature through a feedback function. When the feedback function is zero (η(κi) ≡ 0), there is no feedback and the phases evolve with no dependence on body curvature. This corresponds to the gain ηm = 0 or ηd = 0 in each of the two feedback models that we propose. For this control case, because the phases of segments i and j on each lateral side of the lamprey body are initialized to differ by the prescribed phase shift ψij and the phases of segments on opposite sides are initialized to differ by π, all phases move at the constant velocity ω. The initial conditions chosen for the phase oscillators did not affect the steady state swimming results, as long as the chains were not initialized with exactly the same phases. To maintain consistency in the simulations, the same initial conditions were used for each simulation.
To produce a muscle activation signal, the phases are used to construct an on-off activation following Eq (3). The cutoff value τ was tuned so that duration of the activation signal in the control case matches that of freely swimming lampreys [4] and of our previous models [19, 20]. This activation signal has a temporal period of one second and a duty cycle of 0.36, meaning that each segment is activated for 36% of the period. Because we tuned the neural activation signal to be the same as in our previous models, and because the body and fluid mechanics are the same, the emergent kinematics and energetics of this control case match the computational results in our previous work [19, 20].
Fig 7 shows a snapshot of the lamprey body and vorticity wake for this control case. The inputs to this model system are the passive bending rigidity of the lamprey, the fluid viscosity and density, the maximum tetanic force that each muscle could exert, the baseline frequency and coupling parameters of the CPG, and the activation cutoff value. The outputs are the flow patterns (shown as red and blue vorticity in Fig 7), the evolution of the CPG phase values, and the movement of the body (black lines in Fig 7). From the body movement, as in previous studies, we compute the swimming speed and the amplitude, wavelength, and speed of the mechanical wave [19, 20, 39], the ratio of the wave speeds of the mechanical and activation waves, [37, 47], and the muscular work done by swimmer. Based upon swimming velocity, body length and the density and viscosity of water, the Reynolds number of this control swimmer is about Re = 8000.
In previous studies using this model, we examined how the swimming performance depended upon body stiffness [20] and muscle nonlinearities [19]. While one could have used shear as a body-sensing measure, in the closed-loop model presented here, we explore how different assumptions about feedback from body curvature on the evolving signal from the CPG affect this swimming performance.
In the lamprey, proprioceptive feedback is mediated by mechanoreceptive cells in the spinal cord called edge cells [5]. When the spinal cord bends, one side is stretched, which activates edge cells on that side [5, 13]. The greater the curvature, the stronger the activation [13]. The activated edge cells then inhibit activity on the opposite side and excite activity on the same side [7]. To approximate these effects in our model, we assume that the feedback signal is proportional to the magnitude and direction of curvature. For these simulations, labeled D, η(κi) = (−1)k ηd κi where k is the side of the body (1 is left, 2 is right). Thus, the natural case is represented by ηd > 0. If κi is positive, representing bending to the right, and ηd is positive, oscillators on the right side of the body receive a positive signal, representing excitation, while those on the left side receive a negative signal, representing inhibition.
This class of feedback affects the duty cycle during steady swimming, without affecting the frequency. Each swimmer was initialized in a straight configuration in a still fluid near the right of the domain. Fig 8(A) shows the computed activation signal over two seconds of simulated time on each side of a representative segment of the lamprey body for gains ηd = −15 cm · rad · s−1 (top), ηd = 0 cm · rad · s−1 (middle), and ηd = 15 cm · rad · s−1 (bottom). For instance, for the positive gain value, the amount of time that a muscle segment is activated is less than that for the control case. Fig 8(B) shows the corresponding force development on the opposing segments for each of the simulations. The high duty cycle in the negative gain case lead to considerable co-contraction, as seen by the overlapping blue and red curves on the top panel in Fig 8B. The amount of co-contraction decreases as gain increases.
Fig 9 shows how these differences in force development on individual segments affect the overall swimming kinematics and dynamics. Whether the gain is positive or negative, each swimmer with directional curvature feedback (in black) swims more slowly than the control (gray). The dynamics of the negative gain swimmer compared to the control is shown in S1 Movie, and the dynamics of the positive gain swimmer compared to the control is shown in S2 Movie. In the case of negative gain, the long duration of co-contraction of opposing segments reduces the achieved tail-beat amplitude of the swimmer. In the case of positive gain, the short duration of activity leads to less total force, which also reduces the achieved tail beat amplitude. These effects are quantified in detail below.
In the lamprey, activating edge cells produce a generalized excitatory effect on the CPG [41, 42] that does not depend on the direction of the curvature. To approximate this effect, we examined feedback that depended only on the magnitude of curvature. In this case, labeled M, the feedback advances the phase of the oscillators equally on both sides. For a fixed muscle segment, the feedback term η(κi) = ηm|κi| increases the frequency of the activation signal if ηm is positive and decreases it if ηm is negative. Fig 10(A) shows the computed activation signal over two seconds of simulated time on each side of a representative segment of the lamprey body for gains ηm = −0.08 cm · rad · s−1 (top), ηm = 0 cm · rad · s−1 (middle), ηm = 0.08 cm · rad · s−1 (bottom). For each gain, we see that the activation signals on the left and right sides are in opposite phase (when one side is on, the other is off) and the duty cycle is approximately constant. Unlike the D case, we see that frequency of activation increases with gain. Fig 10(B) shows the magnitude of the contractile force exerted by the opposing segments that results from the activation signals shown in Fig 10(A) for each gain value. We see the maximal contractile force achieved is basically unchanged in each case, but the frequency increases with gain so that the force is maintained for a shorter period of time. This reduces the total force developed in each cycle. Hence, in the positive gain case the total activation time is lower, so muscles achieve peak force for less time (Fig 10B, bottom panel).
With this class of feedback, swimmers with positive gain swim faster than the control case, while those with negative gain swim slower. Fig 11 shows the body shape and flow patterns for two swimmers at time t = 8 s, one with gain ηm = −0.08 cm · rad · s−1 (Fig 11A) and the other with gain ηm = 0.08 cm · rad · s−1. The dynamics of the negative gain swimmer compared to the control is shown in S3 Movie, and the dynamics of the positive gain swimmer compared to the control is shown in S4 Movie. While each swimmer progressed towards the left as the waves of neural activation moved from head to tail, the swimmer with negative gain progresses more slowly than the control because of the decreased beat frequency. Its drift upwards is due to transient fluid effects from the strong start-up vortex that is shed during the first tail beat. Because of the lamprey’s reduced beat frequency and the subsequent reduced swimming velocity, this vortex remained near the body, pushing it upwards. The emergent waveforms of the three swimmers differ only slightly in amplitude and wavelength.
In the examples shown above, we see that magnitude feedback (M) affects the frequency of the activation signal, while directional feedback (D) alters the duty cycle. Here we examine how these changes in activation signal depend upon the gain constants ηm or ηd. As the phase model is heuristic, there are no direct experimental measurements that can be used to set the values for these constants. We performed computational experiments to determine the range of values of the gains around the control case of ηm = 0 or ηd = 0 that resulted in sustained, periodic swimming. For instance, for large absolute values of gain, the phase velocities could either get too large, or, in the case of directional feedback, the phase velocities could become negative, so that realistic activation signals propagating along the left and right muscle segments could not be achieved. For the case of magnitude feedback, we found the range to be |ηm| < 0.09 cm · rad · s−1, and for directional feedback the range was found to be |ηd| < 20 cm · rad · s−1. To compare the effects of the different classes of feedback at a range of gains, we scaled the gain relative to the maximum range for that class. In other words, we scaled ηm by 0.09 cm · rad · s−1 and ηd by 20 cm · rad · s−1.
The frequency of the CPG activation signal on each muscle segment directly affects the tailbeat frequency of the computational swimmer. Fig 12A shows the tailbeat frequency as a function of the non-dimensional gain parameter. We see that in the magnitude feedback case, tail beat frequency increases linearly with gain. However, in the directional feedback case, the frequency is unchanged from 1 Hz. In contrast, Fig 12B shows the average duty cycle along the lamprey body as a function of non-dimensional gain. While the duty cycle is nearly constant at 0.36 in the magnitude feedback case, it decreases linearly with gain in directional feedback. We examine the implications of these effects on body kinematics and swimming performance below.
In living fishes, both frequency and duty cycle vary. Nearly all fishes increase tail beat frequency as they increase swimming speed while maintaining a fairly constant tail beat amplitude [31, 50]. Duty cycle varies along the body and across fish species [4, 37], and a recent study shows that, in bluegill sunfish, it decreases with increasing swimming speed [51].
Our results suggest that feedback structure may modulate frequency and duty cycle, but fish can alter activation frequency and duty cycle in a variety of other ways. Frequency may be modulated by changing the descending drive from the brain [52, 53]. Less is known about how duty cycle may be modulated. Currently only the phase of the oscillators and not the amplitude of the activation are affected by feedback. This type of model will capture changes in frequency of activation and deactivation, but not the relative strength of those activations. In the future we will incorporate models in which the amplitude of the CPG output may be affected by feedback input, which would require shifting from a phase model to a model that incorporates neural properties (e.g. [33, 54]).
The response to sensory feedback is much more sensitive to the magnitude of curvature, rather than the magnitude and direction. This is due to an antisymmetry in the directional signal across each segment of the body. Positive curvature feedback from one side is partially cancelled by negative curvature feedback from the other side, or vice versa. When feedback is only due to the magnitude of the curvature, the feedback signals from both sides tend to add, making the overall system much more sensitive to this type of feedback.
Both forms of feedback change the duration of the activation signal, either symmetrically for both left and right sides (magnitude feedback) or asymmetrically (directional feedback). Because the muscle model requires time to develop force, the shorter the activation duration, the lower the average force. For example, compare the top and bottom panels in Fig 10B. These contractile forces are coupled to the body’s passive elastic forces and to the viscous incompressible fluid surrounding the lamprey. Together, this coupling leads to an emergent swimming waveform. Fig 13 shows the tail beat amplitude and body wavelength for the computational lamprey as a function of non-dimensional gain.
Fig 13A shows that the two types of feedback have qualitatively different effects on amplitude. In the magnitude feedback case the amplitude steadily decreases as the gain increases. The decrease in tail amplitude is due to the increase in activation frequency, which reduces the average force developed at a segment in each cycle. In the directional feedback case, tail amplitude is no longer monotonic with gain, but instead has a maximum at a slightly negative gain. With the most negative gains, the considerable amount of co-activation offsets the increased force development so the net force available for bending is reduced, resulting in a lower amplitude. At high positive gains, the duration of muscle activity is very low, which means that the muscle cannot produce much force, which also results in a low amplitude (see Fig 8). In between, at gain ηd = −3 cm · rad · s−1 (−15%), the amount of co-activation is optimal, resulting in the largest amplitude.
For both classes of feedback, wavelength follows a very similar pattern as amplitude (Fig 13B). It decreases with increasing gain for magnitude feedback, and has an optimum for directional feedback. Because the lamprey body is nearly inextensible, it may seem inconsistent for amplitude and wavelength to follow the same pattern. For a traveling wave with a constant amplitude along the body, wavelength should be inversely proportional to amplitude. However, here we are reporting only tail amplitude. In fact, the wave envelope as a whole depends nonlinearly on the feedback, and inextensibility is maintained. For example, at negative gains for magnitude feedback, as tail amplitude increases the head amplitude decreases, which allows the wavelength also to increase even as the body length stays the same.
The body deformation couples with the surrounding fluid to give rise to forward movement. Natural lampreys swim at about 0.1 L/s (body lengths per second) during migration [48]. Maximum sustained speeds are about 2.5 L/s [49] and they are capable of bursts up to about 5 L/s [49]. Fig 14 shows the swimming speed of the computational lamprey as a function of non-dimensional gain for both the directional feedback cases. The swimming speed for magnitude feedback increases as gain increases, but begins to level off at the most positive gain values. This effect is due to the combined effects of tail beat frequency and amplitude. As gain increases, frequency increases but amplitude decreases, both approximately linearly. In swimming eels, swimming speed is best correlated with the lateral tail velocity, or fA, where f is frequency and A is amplitude [39]. Thus, if U ∝ fA for our swimmer, the relationship should be nonlinear, which is what is seen in Fig 14.
Fig 14 also shows that for the directional feedback case, the effects of gain on swimming speed are similar to those on amplitude, since frequency is unchanged. However, even though the shapes of the curves are similar, the swimming speed is not a simple function of the tail amplitude, because the maximum swimming speed does not occur at the same gain as the maximum amplitude. The maximum swimming speed, 0.52 L · s−1, occurs at a gain of ηd = −4 cm · rad · s−1 (−20%), corresponding to a duty cycle of 0.389, while maximum amplitude occurs at a gain of −3 cm · rad · s−1 (−15%), corresponding to a duty cycle of 0.3831.
A measure of the effectiveness of a swimming stroke is the Strouhal number St which is the ratio of the speed of the tail to the forward swimming motion
S t = 2 f A U (22)
where f is the tailbeat frequency, A is the tailbeat amplitude and U is the swimming speed. Lower St indicates a more effective swimming motion, because the tail’s side-to-side velocity is lower relative to the forward velocity. Fig 15 shows that for both directional feedback cases, St decreases as non-dimensional gain increases, although the total change in St is larger for the magnitude cases than the directional cases. The swimmer is thus becoming a more effective swimmer as gain becomes larger and positive. Even so, compared to swimming fishes, the Strouhal number is still relatively high. Fishes tend to swim with St between 0.2 and 0.4 [55, 56], which corresponds to an efficiency optimum [56]. We note that since our model is 2D, agreement between computed Strouhal numbers and those of natural animals is qualitative, however the 2D model is an essential step to understanding swimming in a 3D fluid.
Another metric associated with effective swimming is the neuromechanical phase lag [37, 38, 57]. At anterior points on the body, once the swimmer reaches steady state, the bending precedes muscle activity, so that muscle activity is always while the muscle is shortening. When a muscle is active while shortening, it produces positive mechanical power. Closer to the tail, the muscle becomes active earlier in the cycle, sometimes to the extent that the muscle is active while lengthening (see Fig 6 for an illustration of this effect) [47]. When the posterior muscles are active while lengthening, they absorb energy, producing negative mechanical work. At the same time, these muscles also stiffen the posterior body, which helps to transmit body forces to the fluid more effectively [20, 38]. In Fig 6, we quantify this effect by computing the ratio of the mechanical wave speed and the activation wave speed. When this ratio is less than one, it mean that the posterior muscle is active earlier in the cycle, similar to what is seen in fishes.
Fig 15B shows that for the magnitude feedback case, the wavespeed ratio decreases with increasing gain. This change is approximately in proportion to the increase in frequency (Fig 12A) and the associated decrease in body wavelength (Fig 13B). In this feedback case, the activation wave has the same wavelength, regardless of gain, which explains the decrease in the wavespeed ratio.
In the directional feedback cases, the wavespeed ratio is not monotonically associated with gain (Fig 15B). Instead, it comes closest to one at ηd = −20%, similar to the effect of gain on curvature wavelength (Fig 13B). At high positive or negative gain, the ratio decreases, indicating a larger phase lag near the tail for these cases. In this case, however, the effect is not solely due to the change in curvature wavelength. At high negative gain, the activation wave is longer (1.24 L), while at high positive gain, the wave is shorter (0.72 L). The change in activation wavelength accounts for the smaller decrease in the ratio at positive gains with directional feedback (Fig 15B).
These results are partially consistent with the hypothesis we proposed for the earlier model, that higher phase lags develop when the internal mechanical forces are low relative to the fluid forces [20]. For the magnitude feedback case, increasing frequency results in lower average muscle force (Fig 10), even as swimming speed increases, which suggests that the internal forces are declining relative to the external forces. For the directional feedback case, the same effect is seen for positive gain: duty cycle decreases, so the average muscle force decreases (Fig 8B), leading to a decrease in internal muscle forces relative to external fluid forces.
The small wavespeed ratio for negative gains in the directional case seem to follow a different pattern. The internal muscle forces increase on average for negative ηd (see Fig 8) and there is more co-activation, which should lead to an effectively stiffer body overall [58]. The activation wavelength, though, gets longer at more negative gain, even as the curvature wavelength gets shorter. This leads to a decreasing wavespeed ratio.
Williams et al. [47] measured the wavelengths of the mechanical and activation waves in lampreys. Based on her published results, lampreys swim with a wavespeed ratio of approximately 0.7 [47], a value that is substantially lower than we find in any of our simulations (Fig 15B). Experimental results suggest that lampreys may respond differently to mechanical inputs in the posterior part of the body [9, 10], a pattern that may increase the phase lag near the tail [9]. In our simulations, sensory inputs are processed equivalently, regardless of their location. To accurately reproduce the phase lag of living lampreys, we may need to alter the feedback pattern along the body.
We approximated the cost of transport (metabolic energy per unit length per unit distance). Muscles produce positive mechanical work when they are active and shortening, and require a metabolic energy input that is proportional to the mechanical work. Muscles produce negative mechanical work when they are active and shortening. This process requires much less metabolic energy, but is still proportional to the amount of negative work (Ruina et al. [59]). Thus, following Hamlet et al. [19], we compute the cost of transport by summing the positive work and a fractional amount of the absolute value of the negative work done by the lamprey’s muscles over each segment at each time step over a full cycle, and then divided the result by the cycle period and mass of the lamprey (as in [19]). For comparison, we also normalized the results by the cost of transport for the control case.
All of the kinematic and muscular factors described above combine to propel the swimmer forward with a particular energy cost to go a certain distance. For both feedback cases, the cost of transport decreases with increasing gain. Fig 16 shows the normalized cost of transport as a function of non-dimensional gain. The effect is most pronounced for the directional feedback case. At the highest positive gain, the swimmer uses 24% less energy to go the same distance as the control case. For the magnitude feedback case, the effect is less linear. At the highest positive gain, the swimmer uses 8% less energy, but at the most negative gain, the swimmer uses 18% more energy.
For both feedback cases, the average muscular force decreases at higher positive gain (Figs 8B and 10B), resulting in lower average muscular work. The Strouhal number also decreases, indicating that these swimmers are hydrodynamically more effective. Thus, the overall cost of transport decreases as gain increases.
At negative gain, the cost increases, but the cause of the increase is different for the two feedback cases. For magnitude feedback, the increase in cost is due to the same effect that causes the decrease for positive gain. As gain decreases and becomes negative, the muscular work increases and Strouhal number increases, indicating a less effective swimmer with a higher overall energetic cost. For directional feedback, the increase in cost is due to the increase in co-activation of the left and right side muscles. As gain becomes more negative, more and more of a muscle’s force is used to oppose its antagonist, rather than deforming the body and producing forward propulsion.
Wavespeed ratios less than one, which correspond to negative phase lags near the tail, have been thought to indicate more effective swimming [38], possibly with lower cost of transport. Our simulations suggest that this is not always true. In the directional feedback case, the wavespeed ratio becomes smaller as gain becomes negative, but the overall cost goes up, due to the increase in co-activation.
Cost of transport does not explicitly include speed. In both feedback cases, at high positive gain, the swimmer uses less energy to go a unit distance. However, the swimmer with magnitude feedback goes 17% faster than the swimmer with directional feedback, so it takes less time to go that distance. Animals that need to minimize energy cost may need to factor in both the overall cost of transport and also the speed.
Many studies of feedback examine its role in responding to perturbations. Our simulation, in contrast, examines the role of feedback in maintenance of a steady effective locomotor pattern. Relatively few other studies have taken this approach. For example, Ekeberg and Grillner [28] developed a neuromechanical model of lamprey swimming that incorporated feedback from edge cells, proportional to curvature, similar to our directional feedback case. They found a “slight increase” in spatial wavelength, within the range of uncertainty in the parameters, when edge cell feedback was included [28], but they did not investigate whether the effects might be different if the gain was higher. It seems likely that their model was in the range of low gain feedback in which wavelength does not change substantially with gain (i.e., −20 < ηd < 40 in Fig 13B).
Gazzola et al. [31] simulated a minimal model of a flexible swimmer that included proprioceptive feedback. They were interested in whether feedback alone, without a CPG, could allow the model to swim steadily. They included passive elastic forces based on a beam model, active muscular forces as an additional internal torque, and hydrodynamic forces based on elongated body theory. They modeled proprioceptive feedback by adding a term to the internal torque. This feedback term was directly proportional to local curvature, except with a time delay. With a CPG, but without feedback, they found that the swimmer had a resonant behavior: it had peaks in swimming speed at several activation frequencies. With feedback, but without the CPG, they found that the swimmer tended to converge to the nearest resonant peak, as long as they gave it some starting deformation. They found that increasing the feedback gain (χ in their Eq [9]), tended to increase the amplitude and the swimming speed. It is difficult to compare that finding to our results, because of the lack of a CPG model in their feedback case, but their feedback term is most similar to our directional feedback model. Unlike their result, we found that amplitude and swimming speed decreased as we increased feedback gain. This difference may be related to changes in level of co-activation in our model, a property that cannot be simulated directly using an added torque as in Gazzola et al. [31].
Very recently, Thandiackal and Ijspeert [60] simulated a similar swimmer, using a CPG modeled with phase oscillators and a feedback term that depended on the local fluid dynamic pressure. Their fluid dynamic model is based on elongated body theory [61], which has been validated, but is much simpler than our Navier-Stokes model. They modeled their feedback approach on the lateral line sensory system in fishes, while ours is modeled on the edge cells in lampreys. In their model, the feedback had a very strong effect on both the mechanical and activation wavelength. In our model, without feedback, the system will have an nominal activation wavelength of 2π/ψ (see our Eq (1) and their Eqn. 2.3). In their model, as they varied ψ over a wide range, feedback tended to push the swimmer to have about one full wave on its body. If the nominal wavelength was small, feedback increased the wavelength, and if the nominal wavelength was large, feedback decreased the wavelength [60]. We did not examine the effects of changing ψ, but we also found that feedback could have strong effects on wavelength.
Our feedback model approximates two different known effects in the lamprey’s sensorimotor control system: first, the phasic effect of edge cells [7] that tends to enhance the ongoing locomotor rhythm [2], and second, the generalized excitatory effect of sensory feedback [3, 41]. Even in the absence of perturbations, we find that both effects are important during steady locomotion.
In our model, both forms of feedback primarily affect the duration and frequency of muscle activation along the body. Increasing the frequency of activation (through magnitude feedback) increases the swimming speed, as seen in fishes [39, 50]. In fishes, however, amplitude usually remains approximately constant as frequency and speed increase [50]. In our model, the amplitude decreases as frequency increases (Fig 13), which means that speed does not increase linearly (Fig 14). This suggests that fish probably have to increase the overall activation strength as frequency goes up, in order to increase swimming speed.
Increasing the duration of muscle activity (through directional feedback) changes the amount of co-activation of muscles on opposite sides of the body. This has a more complex effect on swimming than changing frequency. In particular, the fastest swimmers have a fairly substantial amount of co-activation (Fig 14, black curve). Co-activation can change the effective stiffness of a joint [62], and our previous results showed that there is an optimal passive stiffness for maximum swimming speed [20]. It may be that the co-activation we see here alters the effective stiffness to maximize swimming speed. However, co-activation also comes at a cost. Swimming at the optimal speed requires more energy than swimming at lower speeds with less co-activation (Fig 16, positive gains).
Sensory feedback may have an important ongoing effect on steady locomotion, even in the absence of perturbations. With both forms of curvature feedback, we see that the energetic cost of transport decreases with gain. The decrease in cost may also be aided by the decrease in force with gain (Fig 10), which in turn decreases the muscular work at higher gains. However, speed increases with increasing gain (Fig 14) and Strouhal number decreases, which indicate that swimming is becoming more effective at higher gain, leading to an overall decrease in energetic cost.
While it is well-established that the lamprey’s edge cells are stretch receptors, exactly how this information on stretch feeds back to the CPG is not known. As a starting point to explore the closed-loop locomotor system, here we chose a simple phase-oscillator model of the CPG as well as simple models of feedback where curvature has an additive effect on the evolution of the phase oscillators. In the phase oscillator model, the feedback can only affect the phase of the CPG signal and not its amplitude. More detailed CPG models that capture the neural system (e.g. [45, 63]) will be necessary to analyze these effects. Within the context of the simple model presented here, future investigations will study the effect of feedback with a time delay, as well as the effect of Reynolds number on the system where, perhaps, an animal may need to change it’s control strategy as it grows. In addition, here we only examined the closed-loop system during steady locomotion in the absence of perturbations. Future studies will investigate how the closed-loop swimmer reacts to perturbations in the fluid environment as well as how curvature feedback affects it’s ability to actively turn.
|
10.1371/journal.ppat.1006153 | Toxin Mediates Sepsis Caused by Methicillin-Resistant Staphylococcus epidermidis | Bacterial sepsis is a major killer in hospitalized patients. Coagulase-negative staphylococci (CNS) with the leading species Staphylococcus epidermidis are the most frequent causes of nosocomial sepsis, with most infectious isolates being methicillin-resistant. However, which bacterial factors underlie the pathogenesis of CNS sepsis is unknown. While it has been commonly believed that invariant structures on the surface of CNS trigger sepsis by causing an over-reaction of the immune system, we show here that sepsis caused by methicillin-resistant S. epidermidis is to a large extent mediated by the methicillin resistance island-encoded peptide toxin, PSM-mec. PSM-mec contributed to bacterial survival in whole human blood and resistance to neutrophil-mediated killing, and caused significantly increased mortality and cytokine expression in a mouse sepsis model. Furthermore, we show that the PSM-mec peptide itself, rather than the regulatory RNA in which its gene is embedded, is responsible for the observed virulence phenotype. This finding is of particular importance given the contrasting roles of the psm-mec locus that have been reported in S. aureus strains, inasmuch as our findings suggest that the psm-mec locus may exert effects in the background of S. aureus strains that differ from its original role in the CNS environment due to originally “unintended” interferences. Notably, while toxins have never been clearly implied in CNS infections, our tissue culture and mouse infection model data indicate that an important type of infection caused by the predominant CNS species is mediated to a large extent by a toxin. These findings suggest that CNS infections may be amenable to virulence-targeted drug development approaches.
| Coagulase-negative staphylococci (CNS) are the leading cause of sepsis in hospitalized patients, causing a significant number of deaths. This situation is further worsened by a limitation of therapeutic options due to the fact that most CNS infectious isolates are resistant to methicillin. CNS sepsis has been assumed to be due to on over-reacting immune response triggered by invariant bacterial surface structures. By using tissue culture and animal infection model-based evidence, we here show that in contrast to that notion, the PSM-mec toxin produced by methicillin-resistant strains of the leading CNS species Staphylococcus epidermidis has a strong impact on the severity of sepsis and its outcome. This is the first report to link a toxin to the pathogenesis of the most frequent bacterial cause of sepsis. Notably, these findings pave the way for anti-virulence strategies against this widespread and deadly type of infection.
| Bacterial sepsis is a frequent cause of death in hospitalized patients. Coagulase-negative staphylococci (CNS) are the leading cause of nosocomial sepsis, especially in neonates [1–3]. CNS sepsis most often originates from the infection of indwelling medical devices, such as in catheter-related bloodstream infections (CRBSIs) or central line-associated blood stream infections (CLABSIs) [4]. Most prominent among CNS infections are those due to the skin commensal Staphylococcus epidermidis [5]. However, the bacterial factors contributing to the development of sepsis, in particular in CNS, are poorly understood.
Given that toxins have long been assumed to be widely absent from CNS [6], sepsis caused by S. epidermidis and other CNS, similar to other Gram-positive bacteria, has so far been believed to be due predominantly to an overwhelming immune reaction directed against invariable, pro-inflammatory cell surface molecules, such as teichoic acids and lipopeptides [7]. Recently, the notion that CNS do not commonly produce toxins had to be revised with the discovery of the pro-inflammatory and cytolytic phenol-soluble modulin (PSM) staphylococcal toxin family [8]. However, due to the difficulties associated with genetic manipulation of S. epidermidis and other CNS, the roles of PSMs in CNS infections, including most notably sepsis, have hitherto remained unexplored.
Most S. epidermidis blood infections are caused by methicillin-resistant strains (MRSE), with methicillin resistance rates even exceeding those found among S. aureus [9]. Methicillin resistance is encoded on so-called staphylococcal chromosome cassette (SCC) mec mobile genetic elements, which are believed to have originated from CNS, from where they were transferred to S. aureus [10]. While other PSMs are core-genome encoded [8], one PSM toxin, called PSM-mec, is encoded within SCCmec elements of subtypes II, III, and VIII [11, 12]. The psm-mec gene is embedded in a short regulatory (sr) RNA, which in S. aureus has been reported to down-regulate the production of other PSMs and thereby decrease virulence [13, 14]. While this effect has been claimed to generally explain lower virulence of hospital-associated as compared to community-associated MRSA strains [13], it is quite moderate and extensively strain-dependent [11, 13]. Recently, the psm-mec locus has been introduced on a plasmid into some CNS that naturally lack psm-mec, and was reported to trigger gene regulatory changes [15]; but the roles that the PSM-mec peptide or the psm-mec srRNA naturally play in CNS including S. epidermidis are unknown.
Here we analyzed the role of the psm-mec locus in S. epidermidis sepsis by using tissue culture and animal infection models. Our findings show for the first time that a toxin can have a strong impact on CNS sepsis, setting the stage for anti-virulence strategies directed against this frequent and deadly infection.
To analyze the impact of the psm-mec locus on S. epidermidis sepsis, we produced isogenic psm-mec deletion mutants (Δpsm-mec) in two MRSE strains, a clinical isolate (SE620) and the genome-sequenced strain RP62A. PSM-mec production in these strains is representative of clinical PSM-mec-positive MRSE (S1 Fig), which we determined in a clinical S. epidermidis strain collection from Norway to occur in ~ 2/3 (59/91) of the ~ 50% (91/180) methicillin-resistant S. epidermidis. We also introduced a point mutation in the start codon of the psm-mec gene in the genome of strain SE620 to differentiate between effects mediated by the PSM-mec peptide versus those due to the psm-mec srRNA (psm-mec*). Notably, the stability of the psm-mec RNA was not significantly altered by introduction of the 1-basepair start codon mutation (S2 Fig).
We first analyzed those mutants in a murine sepsis model. Mortality was significantly reduced in the Δpsm-mec mutants of both strains (Fig 1A and 1B). There was no significant difference between the Δpsm-mec mutant and the psm-mec* start codon mutant (Fig 1A). Furthermore, CFU in the blood and the kidneys were strongly reduced in the Δpsm-mec mutants of both strains and the psm-mec* start codon mutant (Fig 1C–1F). These results demonstrate a strong contribution of the PSM-mec toxin to bacteremia and mortality due to S. epidermidis sepsis, while the psm-mec srRNA did not show any impact.
We showed previously that synthetic PSM-mec peptide is strongly pro-inflammatory and has moderate to strong cytolytic capacity [12]. To analyze the contribution that the psm-mec locus has to pro-inflammatory and cytolytic capacity in the S. epidermidis background, we measured cytokine concentrations during experimental murine sepsis and determined cytolytic capacity of the bacterial strains toward human neutrophils in vitro. Cytokine concentrations during sepsis are the result of a systemic reaction due to several immune cell types, and are thus best determined in vivo, while cytolytic capacity can be most accurately measured in vitro.
The PSM-mec peptide, but not the psm-mec srRNA, had a strong and significant impact on the production of cytokines during murine sepsis (Fig 2). At 12 h after infection, the mouse IL-8 homologue CXCL1 was significantly reduced when mice were infected with the Δpsm-mec or psm-mec* start codon mutant of strain SE620, to about half the concentration measured in mice infected with the wild-type strain (Fig 2A). Concentrations of IL-1β and TNF-α were even more strongly reduced to levels not significantly different from those measured in mock (PBS) infected animals (Fig 2B and 2C). In the RP62A background, the phenotypes were similar, with differences being more pronounced during earlier stages of the infection (measured at 2 versus 12 h) (Fig 3). These results showed that the cytokine storm that commonly accompanies bacterial sepsis is strongly dependent on the PSM-mec toxin in S. epidermidis.
In addition to being pro-inflammatory, the PSM-mec toxin has pronounced cytolytic capacity [12]. Cytolysis by PSMs is believed to be most important for infection when bacteria are engulfed in the phagosome of neutrophils and other phagocytes [16, 17]. Survival of bacteria when incubated with human neutrophils and survival in whole human blood was significantly higher with the S. epidermidis wild-type strain than with Δpsm-mec or psm-mec* start codon mutants, as was killing of neutrophils when incubated with whole bacteria (Fig 4), emphasizing the role of the PSM-mec toxin in evasion of neutrophil killing and resistance to the strong bactericidal capacities of immune defense mechanisms in human blood. Together, these results indicate that the both the pro-inflammatory and cytolytic capacities of the PSM-mec peptide contribute to the development of S. epidermidis sepsis.
In S. aureus, the psm-mec locus has also been implicated in biofilm-forming capacity, although effects were generally minor and highly strain-dependent [12]. Similar to S. aureus, biofilm formation in S. epidermidis was affected only slightly by the psm-mec locus, and as this was seen only in one strain, similarly strain-dependent (Fig 5). In that strain, SE620, the effect was due to the PSM-mec peptide, not the psm-mec srRNA. These findings indicate that during indwelling medical device-associated blood stream infections by S. epidermidis, the impact of PSM-mec generally is by contributing to the development of sepsis, as we have shown here, rather than by promoting biofilm formation on the device itself.
Our results showed that the psm-mec srRNA is not involved with sepsis or other relevant virulence phenotypes in S. epidermidis. As a previous study suggested that the psm-mec srRNA leads to gene regulatory changes in S. epidermidis [15], based on the introduction of a psm-mec expressing plasmid into S. epidermidis, we also directly investigated whether the psm-mec locus has a gene regulatory impact in S. epidermidis. The most important regulatory effect of the psm-mec locus in S. aureus, by which the sometimes negative impact of the psm-mec locus on virulence in S. aureus was explained, has been reported to consist in the alteration of the expression of other, core genome-encoded PSMs [18]. PSM expression was altered only to a very low extent in the psm-mec-negative as compared to the wild-type S. epidermidis strains, with changes only significant for some PSMs and never exceeding a factor of ~ 1.5 (Fig 6). This demonstrates that there is only a very minor effect of the psm-mec srRNA on PSM expression when analyzed directly in the S. epidermidis background. Furthermore, we analyzed genome-wide gene expression in the psm-mec mutants of both strains by microarray analysis (Tables 1 and 2). For microarray analysis, strains were grown to the maximum of PSM-mec expression as determined by qRT-PCR (10 h) (Fig 7). While we observed gene regulatory changes that were due to the psm-mec srRNA, they mostly comprised metabolic (e.g., riboflavin and purin/pyrimidine synthesis) rather than virulence genes, and were inconsistent between the two strains. Notably, the results of the previously claimed impact of the psm-mec locus on virulence would be negative [13, 15, 18], contrasting the positive effect we observed in the mouse sepsis model. Such a gene regulatory mechanism can thus be ruled out as underlying psm-mec-mediated development of S. epidermidis sepsis.
Our results may explain the highly inconsistent phenotypes that have been attributed to psm-mec in S. aureus [11–13], inasmuch as the psm-mec locus may exert effects in the background of S. aureus strains that differ from its original role in the CNS environment. One such possibility that remains to be investigated is that the highly expressed psm-mec mRNA interferes with other DNA or RNA sequences in S. aureus. Furthermore, the psm-mec srRNA barely exceeds the limits of the psm-mec gene [14], which contrasts the only other case of an srRNA with an embedded peptide toxin in staphylococci, namely the well-described regulatory RNAIII of the staphylococcal accessory gene regulator (Agr) system. RNAIII significantly exceeds the boundaries of the embedded PSM peptide gene, hld [19]. Together, these observations suggest that the psm-mec srRNA does not serve a well-defined general purpose in virulence gene regulation.
In conclusion, our study reveals that sepsis due to MRSE is mediated to a large extent by the PSM-mec peptide toxin, representing the first example of a toxin being made responsible for the development of CNS sepsis. Our study was largely based on the investigation of isogenic psm-mec mutants in clinical strains of S. epidermidis, using tissue culture and animal infection models. Future clinical work is needed to assess whether PSM-mec and/or other toxins contribute to sepsis in humans. Importantly, our results suggest that CNS sepsis may be amenable to virulence-targeted therapeutic approaches, such as those targeting the quorum-sensing system Agr [20], which strictly regulates PSM expression [21], or monoclonal antibody-based therapy directed against the toxin.
Strain RP62A is a genome-sequenced clinical MRSE isolate [22]. Strain SE620 is an MRSE clinical isolate from Norway [23]. Isogenic Δpsm-mec deletion mutants and the psm-mec* start codon mutant were produced with the constructs previously used for S. aureus [12, 14], using a strategy with the allelic exchange vector pKOR1 [24]. The psm-mec locus and adjacent DNA do not differ between S. aureus and S. epidermidis [25]. For construction of the psm-mec* mutant, the start codon mutation was created by introducing a ClaI restriction site (introducing ATC instead of the ATG start codon) using primer PSMEClarev GAGGGTATGCATATCGATTTCACTGGTGTTATTACAAGC and primer PSMECladir (reverse complement of PSMEClarev). Two PCR fragments were amplified using those primers and primers psmEatt1 and psmEatt2, respectively [12], cut with ClaI, ligated, and cloned into pKOR1. The resulting plasmid was used for allelic replacement as described [24]. Growth patterns of the mutants were indistinguishable from those of the wild-type (S3 Fig). Strains were grown in tryptic soy broth (TSB), unless otherwise noted.
Female, 6–10 weeks old, C57BL/6NCRl (Charles River) mice were used. The mice were injected via the tail vein with 5 x 108 CFU in 100 μl phosphate-buffered saline (PBS) of the indicated bacterial strains grown to mid-exponential growth phase and monitored for disease development every 8 h for up to 120 h. This dosis was determined to be minimally necessary to achieve mortality and production of inflammatory cytokines (S4 Fig). Animals were euthanized immediately if showing signs of respiratory distress, mobility loss, or inability to eat and drink. Cytokine concentrations were measured at 2 and/or 12 h, as indicated, using commercially available ELISA kits (IL-1β, TNF-α, BD BioSciences; CXCL1, R&D Systems).
For survival in whole blood experiments, about 108 bacteria in 100 μl Dulbecco’s PBS from mid-exponential growth phase were added to 500 μl heparinized human blood and mixtures were incubated for 6 h. Aliquots were taken at 2-h intervals, and CFU were determined by plating and incubating plates overnight at 37°C.
For neutrophil interaction experiments, neutrophils were isolated from the venous blood of human volunteers as described [26]. Bacteria from mid-exponential growth phase were mixed with neutrophils at an MOI (bacteria/neutrophils) of 10:1. Bacteria/neutrophil mixtures were incubated at 37°C, 5% CO2, 90% humidity for 6 h. At 2-h intervals, 50 μl of Triton X-100 was added to the 200-μl bacteria/neutrophil suspensions, aliquots were plated, and plates incubated at 37°C overnight for CFU counting. Alternatively, the rate of neutrophil lysis promoted by the bacteria was determined after 4–h incubation using a lactate dehydrogenase (LDH) assay at an MOI of 100:1.
Biofilm formation was assessed in a semi-quantitative 96-well microtiter plate assay as previously described [27], using TSB + 0.5% glucose.
Relative PSM concentrations in culture filtrates were determined as described using reversed-phase high-pressure liquid chromatography/electrospray mass spectrometry (RP-HPLC/ESI-MS) [28].
Quantitative RT-PCR was performed as previously described [29] with the following oligonucleotides: psm-mecF, TGCATATGGATTTCACTGGTGTTA, psm-mecR, CGTTGAATATTTCCTCTGTTTTTTAGTTG, psm-mec probe, ATTTAATCAAGACTTGCATTCAG. Expression was measured relative to that of 16S RNA. Cultures were grown to the maximum of psm-mec expression as determined by qRT-PCR (10 h). Total RNA and cDNA were prepared as described [30]. Biotinylated S. aureus cDNA was hybridized to custom Affymetrix GeneChips (RMLChip 3) with 100% coverage of chromosomal genes from strains S. epidermidis RP62A and scanned according to standard GeneChip protocols (Affymetrix). Each experiment was replicated 3 times. Affymetrix GeneChip Operating Software was used to perform the preliminary analysis of the custom GeneChips at the probe-set level. Subsequent data analysis was performed as described [30]. The complete set of microarray data was deposited in NCBIs Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/) and is accessible through GEO Series accession number GSE85265.
To determine psm-mec mRNA stability in strains S. epidermidis SE620 and the psm-mec* start codon mutant, bacteria were cultured at 37°C for 10 hours. At time t = 0 min, rifampicin (50 mg/ml stock in DMSO) was added to the cultures to a final concentration of 100 μg/ml. One-ml aliquots were taken and immediately centrifuged at 4°C to pellet cells, which were then frozen at -70°C. Remaining cultures were further incubated at 37°C with shaking; one-ml aliquots were taken at the indicated times and RNA was subsequently isolated from all cell pellets as described [29]. Samples were analyzed by qRT-PCR using primers psm-mecR and psm-mecF with a SuperScript III Platinum SYBR Green One-Step qRT-PCR kit (Invitrogen) according to the manufacturer’s instructions. Expression was measured relative to that of 16S RNA.
Statistical analysis was performed using GraphPad Prism Version 6.0. Comparisons were by 1-way or 2-way ANOVA for comparisons of three and more, and by unpaired t-tests for comparisons of 2 groups. Error bars show ±SEM.
The animal protocol (LB1E) was reviewed and approved by the Animal Care and Use Committee at the NIAID, NIH, according to the animal welfare act of the United States (7 U.S.C. 2131 et. seq.). All mouse experiments were performed at the animal care facility of the NIAID, Building 50, in accordance with approved guidelines. All animals were euthanized by CO2 at the end of the studies. Human neutrophils were isolated from blood obtained under approved protocols at the NIH Blood Bank or with a protocol (633/2012BO2) approved by the Institutional Review Board for Human Subjects, NIAID, NIH. All subjects were adult and gave informed written consent.
|
10.1371/journal.pgen.1004359 | A Dominant-Negative Mutation of Mouse Lmx1b Causes Glaucoma and Is Semi-lethal via LBD1-Mediated Dimerisation | Mutations in the LIM-homeodomain transcription factor LMX1B cause nail-patella syndrome, an autosomal dominant pleiotrophic human disorder in which nail, patella and elbow dysplasia is associated with other skeletal abnormalities and variably nephropathy and glaucoma. It is thought to be a haploinsufficient disorder. Studies in the mouse have shown that during development Lmx1b controls limb dorsal-ventral patterning and is also required for kidney and eye development, midbrain-hindbrain boundary establishment and the specification of specific neuronal subtypes. Mice completely deficient for Lmx1b die at birth. In contrast to the situation in humans, heterozygous null mice do not have a mutant phenotype. Here we report a novel mouse mutant Icst, an N-ethyl-N-nitrosourea-induced missense substitution, V265D, in the homeodomain of LMX1B that abolishes DNA binding and thereby the ability to transactivate other genes. Although the homozygous phenotypic consequences of Icst and the null allele of Lmx1b are the same, heterozygous Icst elicits a phenotype whilst the null allele does not. Heterozygous Icst causes glaucomatous eye defects and is semi-lethal, probably due to kidney failure. We show that the null phenotype is rescued more effectively by an Lmx1b transgene than is Icst. Co-immunoprecipitation experiments show that both wild-type and Icst LMX1B are found in complexes with LIM domain binding protein 1 (LDB1), resulting in lower levels of functional LMX1B in Icst heterozygotes than null heterozygotes. We conclude that Icst is a dominant-negative allele of Lmx1b. These findings indicate a reassessment of whether nail-patella syndrome is always haploinsufficient. Furthermore, Icst is a rare example of a model of human glaucoma caused by mutation of the same gene in humans and mice.
| Nail-patella syndrome is a human genetic disease caused by an inactivating mutation in one copy of a gene called LMX1B, with the amount of protein produced from the remaining copy of the gene not being enough for normal function. Patients with this disease have malformations of their nails, elbows and kneecaps. Some patients also develop kidney disease and glaucoma. LMX1B controls where and when other genes are expressed and it is important during development. Studies in mice have shown that complete absence of Lmx1b is lethal at birth. In contrast to humans, mice with only one copy of the gene are normal. Here we describe a new mutant mouse, Icst, which has a mutation in Lmx1b that abolishes the ability of the protein to bind near genes that it controls. Mice with one normal and one copy of Lmx1b with the Icst mutation have eye defects and some die shortly after birth probably due to kidney failure. Therefore having one functional and one mutant copy of Lmx1b is more detrimental than having a half dose of functional protein. The Icst mouse is a model of human glaucoma where mutation of the same gene causes glaucoma in humans and mice.
| Nail-patella syndrome (NPS) (OMIM 161200) is an autosomal dominant human disease, cardinal features of which are nail dysplasia, absent or hypoplastic patellae and abnormal elbows along with iliac horns. In addition, about 30–40% of patients develop nephropathy, which can progress to renal disease [1]. Open angle glaucoma is another feature of the disease that occurs in about 30–40% of patients [2]. Mutations of the transcription factor LMX1B have been found to be the underlying cause of NPS [3]–[6]. LMX1B is a member of the LIM-homeodomain (LIM-HD) family of transcription factors. The protein has two N-terminal LIM domains that are involved in protein-protein interactions followed by a homeodomain that binds to target DNA binding sites. Disease-causing mutations range from complete gene deletion to various frameshift, nonsense, splice and missense mutations. The majority of missense mutations are found in the homeodomain and the LIM domains. There is great variation in the severity and range of phenotypes both within families that carry the same mutation and between families that carry different mutations in LMX1B. Several missense NPS homeodomain mutations tested in vitro have shown no dominant-negative effect on the transcriptional activity of wild-type protein [7], [8] and consequently NPS is thought to be a haploinsufficient disorder. Nevertheless, in a comprehensive study of 106 NPS patients from 32 families, patients with mutations in the homeodomain had more severe proteinurea than those with mutations of the LIM domains, although other aspects of NPS showed no phenotype-genotype correlation [9]. Association of haplotype with severity of nail dysplasia has also been reported [10].
Knockout studies in mice have shown that Lmx1b is required during development for dorsal patterning of the limb, the establishment of the midbrain-hindbrain boundary, the development of the cerebellum, for kidney development and for the specification of certain neuronal subtypes (reviewed in [11]). Mice that lack Lmx1b have ventralised limbs, kidney abnormalities, calvarial bone defects and an absent cerebellum [12]–[14]. There are also anterior segment eye defects [15]. Postnatal conditional deletion experiments have shown that Lmx1b is required for formation of the trabecular meshwork, the maintenance of corneal integrity and transparency and loss results in corneal neovascularisation [16]. Heterozygous knockout mice are apparently normal indicating that in the mouse haploinsufficiency for Lmx1b does not lead to mutant phenotypes [12], [15]. However, heterozygous knockout mice recover less well from unilateral nephrectomy than wild-type mice, suggesting some role in adult kidney regeneration and maintenance [17].
Here we report the identification of an Lmx1b missense mutation, Icst, which has a dominant-negative mode of action. In contrast to heterozygous knockout mice, heterozygous Lmx1bIcst mice have buphthalmic (enlarged or bulging) eyes and develop a glaucoma phenotype. In addition there is some post-natal lethality associated with kidney defects. We demonstrate that the difference in phenotypic consequence of the null and Icst alleles is due to LMX1BIcst protein acting in a dominant-negative manner. These findings have implications for the interpretation of the mode of action of mutant LMX1B in NPS.
We identified the N-ethyl-N-nitrosourea-induced mouse mutation, iris-corneal strands (Icst), in a screen for dominant eye mutations. Mice that carry Icst have buphthalmic eyes suggestive of high intra-ocular pressure [18]. We had previously mapped Icst to proximal Chr 2 between the markers D2Mit365 and D2Mit372 [18]. Within this interval we considered Lmx1b to be a strong candidate for harbouring the Icst mutation because glaucoma occurs in about 30–40% of NPS patients [2]. We amplified and sequenced the exons and flanking regions of Lmx1b from Icst mutant mice and found a single nucleotide change, a T to A transversion in exon 5 in the Icst allele (position 2:33,566,910 in Ensembl Release 74 mouse assembly GRCm38 (http://www.ensembl.org)) which was absent from the reference mouse sequence and an additional 17 mouse strains [19] (Figure 1A). This mutation causes substitution of hydrophobic valine with hydrophilic aspartic acid in the homeodomain (V265D). The affected valine is in the recognition helix, and is very highly conserved across species and paralogues. In the crystal structure of the related paired-type homeodomain, the equivalent valine directly contacts the DNA recognition sequence by making hydrophobic contacts with the second thymine in the TAAT core of the recognition sequence [20]. The nature of the mutation in Lmx1b, coupled with the complementation tests described below, indicate that Icst is the causative mutant allele; Lmx1bIcst. Examination of the sequence traces of RT-PCR products spanning exon 5 of Lmx1b from embryonic samples indicated that equal amounts of mutant and wild-type transcript are produced in heterozygotes (data not shown).
We asked whether the mutant LMX1BIcst protein was able to bind to its recognition sequence. We produced recombinant His-tagged full-length LMX1B protein (S form, see below) and the homeodomain alone, in both wild-type and Icst mutant versions, and used these in bandshift experiments to determine if they could bind to a known LMX1B recognition sequence from intron 1 of the Col4a4 gene [21]. Both wild-type full-length and homeodomain proteins bound but the LMX1BIcst proteins did not (Figure 1B). We then investigated the ability of wild-type and mutant LMX1B protein to transactivate transcription in a reporter assay. We tested two isoforms of the mouse LMX1B protein, a 372 amino acid protein (S) and a longer form (L) that includes an additional 29 N-terminal amino acids initiating from an upstream AUG (see Materials and Methods) [22]. Much of this additional sequence is conserved between humans and mouse but also includes a direct 18 bp repeat encoding a further six amino acids. We introduced the Icst mutation into the two isoforms and tested the ability of the wild-type and mutant versions to transactivate a luciferase reporter, previously shown to be activated by LMX1B, which has six copies of the LMX1B recognition sequence upstream of a minimal promoter [21]. Both L and S wild-type proteins transactivated the luciferase reporter construct on co-transfection, but neither of the LMX1BIcst proteins showed any transactivation activity (Figure 1C). These results showed that the Icst mutation disrupts the ability of the LMX1B protein to recognise and bind to its target sequences and activate transcription.
Mice that are heterozygous for the Lmx1b knockout allele do not have an eye phenotype (Figure 2A) [15]. In contrast, we observed that the eyes of Lmx1bIcst/+ mice had mild corneal opacity at a young age and by six to eight weeks Lmx1bIcst/+ eyes are buphthalmic and display a variety of abnormalities, as previously described [18]. Examples are shown in Figure 2B and 2C. Strands of tissue extend between the iris and cornea in some mice (Figure 2B) and corneal neovascularisation is seen, often associated with corneal ulcers (Figure 2C). Other abnormalities in the cornea included scarring and inflammation, with wrinkling of Descemet's membrane and flattening of the endothelial cells with migration onto the anterior iris surface (Figure 2E). It is likely that endothelial abnormalities contributed to the development of corneal oedema (Figure 2F) and ocular surface compromise which in some cases led to ulceration. Many Lmx1bIcst/+ eyes develop severe corneal ulcers with age. The bulging and distended appearance of the eyes suggested that there might be dysfunction of the drainage system at the iridocorneal angle resulting in raised intraocular pressure. We examined the angle by histology and found abnormalities that varied between eyes. In some Lmx1bIcst/+ eyes there was an open angle (Figure 2H) but in other cases the angle was closed with the iris adhering to the cornea (Figure 2I and 2J). Although the angle is still open in some eyes, it is often narrow and sometimes Schlemm's canal appears abnormally short (Figure 2L). The trabecular meshwork, a structure dependent on Lmx1b for its development [16], typically appears abnormal, often being compressed and hypomorphic (Figure 2L). In line with these histological observations, we found that intraocular pressure (IOP) was elevated in Lmx1bIcst heterozygotes (Figure 3). Between the ages of two and six months IOP was elevated in Lmx1bIcst heterozygotes compared to wild-types (Figure 3A). The mean IOP measurements in older mice at 6–12 months are not statistically different between the two groups, in part due to an increased incidence of low IOP values (<10 mmHg) in older Lmx1bIcst heterozygotes, which is caused by significant corneal damage (ulceration, sometimes with inflammation and perforation). Overall, about 35% of Lmx1bIcst heterozygotes were found to have high IOP as compared to about 5% of wild-types (Figure 3B). Higher IOP is most commonly associated with poorly understood abnormalities of the iridocorneal angle [23] and it is a common and important risk factor for developing glaucoma in humans and can lead to the neurodegenerative hallmarks of the disease which are retinal ganglion cell loss and optic disc cupping. We observed optic nerve cupping in some, but not all, Lmx1bIcst heterozygous mice at various ages (Figure 4A). There was also a profound reduction in the number of retinal ganglion cells in Lmx1bIcst heterozygous mice that was not seen in Lmx1bKO heterozygous mice (Figure 4B). In addition, in Lmx1bIcst heterozygous mice optic nerve damage occurs and the loss of axons appear to be progressive (Figure 4C).
Homozygous knockout mice that lack LMX1B die on the day of birth [12]. They have ventralised limbs, absent cerebellum and kidney and eye defects [12], [14], [15]. We intercrossed heterozygous Lmx1bIcst mice and collected offspring for genotyping at E17.5 and at weaning. At E17.5 the three expected genotypes were present at Mendelian ratios indicating that Lmx1bIcst does not cause lethality in utero (Table 1). However, no Lmx1bIcst homozygous mice were present at weaning and four pups that were dead at birth were found to be homozygotes (Table 1). We examined homozygous mutant embryos and observed a phenotype very similar to that described for the knockout. Lmx1bIcst homozygotes lack a cerebellum, have abnormal kidneys and ventralised limbs (Figure S1 and Figure 5D). We expected this cross to produce twice as many Lmx1bIcst heterozygotes as wild-type. However, at weaning we found equal numbers of these genotypes, indicating a deficit of heterozygotes and suggesting that the Icst allele is semi-lethal (Table 1). We also crossed heterozygous Lmx1bIcst mice with mice heterozygous for the knockout allele Lmx1btm1Rjo (hereafter, Lmx1bKO) and genotyped their offspring at weaning (Table 2). No double mutant mice were seen among 104 offspring, indicating that the two alleles do not complement each other and are thus allelic. There was a deficiency of about a third of Lmx1bIcst heterozygotes compared to Lmx1bKO heterozygotes, but this fell short of statistical significance (χ2 = 3.358, d.f. = 1, P = 0.067). To investigate this apparent dominant lethality further, we backcrossed both alleles of Lmx1b to C57BL/6J (Table 2) and genotyped the offspring. We found the expected 1∶1 ratio of heterozygotes to wild-type for the Lmx1bKO allele backcross whereas for the Lmx1bIcst allele a third fewer than expected heterozygotes were seen (χ2 = 26.3, d.f. = 1, P<0.0001). All of the mice carrying the Icst mutation had abnormal eyes showing that the mutant eye phenotype, as described in the previous section, is completely penetrant on the studied genetic backgrounds. The observed deficiency in the number of Lmx1bIcst heterozygotes indicates that the Lmx1bIcst allele is semi-lethal when heterozygous, in contrast the knockout allele which does not elicit a heterozygous phenotype. Therefore the mutant LMX1BIcst protein must exert a dominant effect during development and/or early postnatal life that always results in ocular abnormalities and can be lethal.
When and why do 30% of Lmx1bIcst heterozygous mice die? The deficiency of Lmx1bIcst/+ mice at weaning (Table 2) was present at P1 (χ2 = 3.922, d.f. = 1, P = 0.048) (Table 3) suggesting the lethality is perinatal and any Lmx1bIcst/+ mice that reach P1 survive as well as their wild-type littermates. We considered a possible cause of death to be kidney failure because of the importance of LMX1B in podocyte and slit diaphragm development [21], [24], [25]. We examined by electron microscopy the glomerular morphology of wild-type, Lmx1bIcst heterozygous and homozygous mice and, for comparison, glomeruli from Lmx1bKO heterozygous and homozygous mice at E17.5 and E18.5 (Figure 5). We took care to examine only the most mature glomeruli so that our analysis would not be confounded by comparing immature and mature podocytes. Both wild-type and Lmx1bKO heterozygous morphology was normal and no abnormalities were found (Figure 5A and 5B) whereas in the homozygous glomeruli the podocytes were effaced on the glomerular basement membrane (GBM) as previously reported (Figure 5C) [24], [25]. In Lmx1bIcst homozygotes, podocytes appeared undeveloped; the morphology resembling that seen in the knockout (Figure 5D). We also found glomerular abnormalities in all Lmx1bIcst/+ embryos examined although the degree varied between individuals (Figure 5 E–L). In glomeruli from two Lmx1bIcst/+ E17.5 embryos there was normal development in many areas with foot processes forming normally (Figure 5E). However, we found various abnormalities in both, representative examples are shown in Figure 5 F–I. In Figure 5F and 5H the GBM is split. In Figure 5F the podocyte is positioned flush against the GBM and foot processes have failed to develop and the GBM itself is fragmented, suggesting it is not adhering properly. In some glomeruli podocytes were immature and cuboidal in shape (Figure 5I). These two individuals might have been Icst heterozygotes which would have survived as the ultrastructural changes are not extensive. In another Lmx1bIcst heterozygote examined at E18.5 there was extremely abnormal morphology (Figure 5 J–L). The GBM is split (Figure 5J and 5K) and only some rudimentary foot process formation was observed (Figure 5J). Areas where podocytes were effaced onto a split GBM were also found (Figure 5L). As we found no evidence of normal glomerular development in this individual it is likely to be an example of one of the Lmx1bIcst heterozygotes which would die. This variability in the extent of the mutant kidney phenotype found in Lmx1bIcst heterozygotes is consistent with the finding that only one third of the Lmx1bIcst heterozygotes fail to survive.
We have shown above that Lmx1bIcst induces gain-of-function heterozygous phenotypes of the eye and the kidney that are not found in heterozygotes for the knockout allele. Both alleles are homozygous lethal at birth. To investigate if wild-type Lmx1b could rescue these phenotypes we made mice that were transgenic for a bacterial artificial chromosome, RP23-305A12, which contains the Lmx1b gene centrally located in a 225 kb insert (Tg(RP23-305A12), hereafter BAC). This transgene was introduced into both the Lmx1bIcst and Lmx1bKO lines and mice crossed to assess if it could rescue the mutant phenotypes. First we examined the effect of the transgene on the Icst heterozygous and homozygous phenotypes. When hemizygous, the transgene rescued the Lmx1bIcst heterozygous gross eye phenotype (Figure S2). However, the hemizygous transgene did not rescue the perinatal lethality affecting Lmx1bIcst homozygotes (Table 4). In contrast, the hemizygous transgene could rescue the perinatal lethality of the homozygous knockout allele, although the number of Lmx1bKO homozygous rescued pups was low and they were small in size with ventralised limbs (see below) (Table 4). Interestingly, variable rescue of the eye phenotype of transgenic Lmx1bKO homozygous mice was observed; although the majority of eyes were normal, some eyes were very abnormal, with damaged corneas and optic nerve heads (Figure S3).
We next asked if increasing the dose of transgenic Lmx1b could better rescue the mutant phenotypes (Table 5). We found that two copies of the transgene could rescue the Lmx1bIcst heterozygous semi-lethality; transgenic Lmx1bIcst heterozygotes survived in the normal Mendelian ratio. Two BAC copies could also rescue the homozygous perinatal lethality, albeit inefficiently (Table 5). The Lmx1bIcst/Icst rescued mice were smaller than their littermates and had the Icst mutant eye phenotype (Figure S4). For the knockout allele, a greater number of homozygous Lmx1bKO mice survived when the transgene was homozygous than when it was hemizygous (Table 5) and in most cases the eyes are grossly normal. Nevertheless, a few did have a mutant eye phenotype and in these cases expression of wild-type Lmx1b from the transgenic BAC was reduced (Figure S4).
All the rescued mice, whether Lmx1bKO/KO or Lmx1bIcst/Icst, had skeletal and limb defects (Figures 6 and 7). Lmx1bIcst/Icst mice homozygous for the transgene had paws that were clearly ventralised; the dorsal surfaces were largely devoid of hair and had thickened skin pads superficially, much like the ventral surface, although the skin was pigmented (Figure 6). Lmx1bKO/KO mice rescued by a hemizygous BAC transgene had the same ventralisation phenotype which was not substantially improved when the transgene was homozygous (Figure 6). However, when we examined the skeletons of these homozygous transgenic rescue mice by X-ray computed microtomography (µCT), we found that aspects of the skeletal phenotype had been rescued (Figure 7). The paw skeleton of Lmx1bIcst/Icst mice homozygous for the BAC transgene was ventralised, as was that of Lmx1bKO/KO mice hemizygous for the BAC transgene (Figure 7). When homozygous the BAC transgene largely rescued the paw skeleton to near wild-type in Lmx1bKO/KO mice, despite the ventralised surface (Figure 7). Other skeletal abnormalities were also differentially rescued. Amongst other skeletal defects in homozygous Lmx1bKO mice the patella is absent and the scapula is very small [12]. We find the same defects in homozygous Lmx1bIcst embryos (Figure S1 and data not shown), and absence of patella is, of course, a cardinal feature of NPS in human patients [1]. Lmx1bIcst heterozygotes (and Lmx1bKO heterozygotes) do have patellae (data not shown). When homozygous, the BAC transgene does not rescue the patellar and scapular defects in Lmx1bIcst/Icst mice (Figure 8B and 8E) but does in Lmx1bKO/KO mice (Figure 8C and 8F), again demonstrating that mutant phenotypes can be rescued more readily from the null background than when LMX1BIcst protein is present.
These BAC transgenic experiments demonstrate that a higher level of wild-type Lmx1b expression (i.e. two BAC copies) is required to elicit rescue of the Lmx1bIcst mutant phenotype than the Lmx1bKO mutant phenotype. This is consistent with an Lmx1bIcst pathogenic mechanism with the LMX1BIcst protein exerting a dominant-negative effect on co-expressed wild-type protein.
How does the LMXIBIcst mutant protein exert this dominant-negative effect on the wild-type protein? LMX1B does not homodimerise [26], [27]. Along with other LIM-HD proteins, LMX1B binds to co-factors via the two LIM domains [28]. One such co-factor is LDB1 which can itself homodimerise thus enabling the formation of homomeric or heteromeric LIM-HD complexes [27]. This raises the possibility that complexes containing both wild-type and mutant LMX1B could be formed in Lmx1bIcst heterozygous mice thus decreasing the level of functional complexes below that found in Lmx1bKO heterozygous mice. To test if such complexes can be formed we transfected cells with Myc-tagged wild-type LMX1B and FLAG-tagged LMX1BIcst either alone or together and immunoprecipitated using anti-Myc antibody. As expected the Myc-tagged wild-type LMX1B was present in the immunoprecipitated fraction but the FLAG-tagged LMX1BIcst was not, confirming that LMX1B indeed does not homodimerise (Figure 9A). When LDB1 was included in the transfections, LDB1 was co-immunoprecipitated with the wild-type protein showing that LMX1B binds to LDB1 as expected (Figure 9B). When all three proteins were present, FLAG-tagged LMX1BIcst protein was co-immunoprecipitated with the wild-type LMX1B (Figure 9B) showing that complexes containing both wild-type and mutant LMX1B are formed where the interaction is mediated by LDB1. Consistent with this, less LDB1 appears to be in the bound fraction when FLAG-tagged LMX1BIcst was included along with the Myc-tagged wild-type LMX1B, probably due to competition between the wild-type and mutant protein for binding to LDB1 (Figure 9B).
The Icst mutation, V265D, in the homeodomain of LMX1B abolishes DNA binding and transcriptional transactivation (Figure 1). Whilst heterozygous null (knockout allele) Lmx1b mice are phenotypically normal [12], [15] and present in Mendelian numbers (Table 2), a fraction of Lmx1bIcst heterozygous mice die with associated kidney GBM defects (Table 2 and Figure 5). No morphological abnormalities have been found in glomeruli of Lmx1bKO heterozygotes up to one year of age [24]. Those Lmx1bIcst heterozygotes that do survive have a highly penetrant eye phenotype which is not seen in Lmx1bKO heterozygotes (Figures 2–4). Depletion of Lmx1b expression in adult mice causes corneal opacity and neovascularisation indicating that a threshold level of Lmx1b expression in the cornea is necessary for the maintenance of corneal integrity [16]. These differences between the Icst and knockout heterozygous phenotypes indicate that the LMX1BIcst protein exerts a dominant-negative effect on the wild-type protein. Dominant-negative mutant activity is typically mediated via protein complexes, usually dimers or higher-order structures, in which participation of one non-functional subunit inactivates the complex [29]–[33]. As previously reported [26], [27] and confirmed by our results (Figure 9) LMX1B does not homodimerise. We have shown that both wild-type LMX1B and LMX1BIcst proteins are found in protein complexes mediated by LDB1 (Figure 9). Complexes containing both wild-type and Icst mutant protein are likely to be non-functional and in Icst heterozygotes the level of functional complexes containing only wild-type protein would be 25%, compared to 50% in the null heterozygotes. This provides an explanation for the difference in the heterozygous phenotype of the two alleles and leads to the prediction that missense mutation in the LIM domains abolishing protein-protein interactions would be equivalent to null alleles.
LMX1B has been shown to interact with LDB1 by yeast two-hybrid experiments [34] and complexes containing both proteins have been detected in rat glomeruli protein lysates [35]. The two genes have overlapping expression patterns [35], [36]. In support of the role of LDB1 in LMX1B function, in mice specific inactivation of Ldb1 in podocytes leads to gradual loss of foot processes and GBM defects are found which lead to renal failure [35]. However, other binding partners for LMX1B are known, for example TCF3 [37], [38] raising the possibility that proteins other than LDB1 may be responsible for mediating the dominant-negative effect in some cell types.
There is a broad spectrum of disease severity both within and between NPS families and no clear genotype–phenotype correlation between the nature of mutations and severity of disease although in all cases skeletal abnormalities are found [5]. It is widely believed to be a haploinsufficient disorder and indeed, patients with a complete deletion of LMX1B have been found [6]. Variability in disease manifestation in patients with the same mutation is often observed [39] indicating genetic background modification. Furthermore, mutation of LMX1B does not always result in NPS. Two groups have reported the detection of novel missense mutations affecting R246 in the homeodomain of LMX1B in patients with isolated renal disease. In one report a patient with nail-patella-like renal disease was found to have an R246Q mutation that has residual transcription activity [40]. In the other report patients with autosomal dominant focal and segmental glomerular sclerosis either with the same R246Q mutation or with a different mutation affecting the same amino acid, R246P, were described [41]. None of the patients had any of the other classic symptoms of NPS. The reason why in these patients malformations are confined to the kidney is not clear. It may be that the residual activity of the R246Q protein is sufficient for normal development outwith the kidney or that these mutations of R246 only compromise binding to a subset of target sequences.
The lack of phenotype in Lmx1b null heterozygous mice [12], [15] contrasts with heterozygous Lmx1bIcst mice which have a strongly penetrant eye phenotype and have glomerular abnormalities that resemble defects found in human NPS patients (Figure 5). They do not, however, have any skeletal abnormalities, which is the most prevalent aspect of NPS. Valine 265 in LMX1B, the equivalent residue to that mutated in the mouse Lmx1bIcst allele, has been found mutated in NPS patients to phenylalanine and to leucine [8], [22]. V265L (originally reported as V242L) along with other patient missense mutations have been tested for dominant-negative effects on wild-type protein in in vitro transcription reporter assays but none have been found [7], [8]. Likewise, we have been unable to find, in in vitro experiments, a dominant-negative effect of the LMX1BIcst protein on transcriptional reporter assays (data not shown), although it is clearly demonstrated by the phenotype in mice. It is apparent that mutant proteins can show dominant-negative activity in mice that is not seen in vitro and it is possible that some of the human LMX1B mutations may be dominant-negative. Indeed, it has been reported that patients with homeodomain mutations exhibit more severe proteinuria, and hence kidney defects, than patients with mutations in the LIM domains suggesting dominant-negative activity [9].
Glaucoma is usually associated with high IOP caused by dysfunction of the ocular drainage structures in the iridocorneal angle of the eye [23]. LMX1B mutations are well established to cause open-angle glaucoma in NPS patients, but due to the influence of modifier genes may also cause glaucoma without NPS. Although confirmation is required, LMX1B haplotype has been suggested to influence open-angle glaucoma in the general population (without other aspects of NPS) [42]. Similar to some of the Lmx1bIcst heterozygous mice, a narrow but open-angle phenotype with high IOP is present in some individuals with an LMX1B mutation [4]. The aetiology of IOP elevation in glaucoma remains poorly understood and it is likely to be mechanistically heterogeneous at the molecular level. Various studies have reported open-angle glaucoma phenotypes in mice [43]–[47]. However, in most cases and due to undefined multifactorial influences, these phenotypes have typically been mild or have not yet proven reproducible between laboratories.
Lmx1bIcst/+ mice have variably open or closed angles, and may be a model of anterior segment dysgenesis leading to high IOP, rather than primary open angle glaucoma. Nevertheless they should provide a valuable mouse model of glaucoma caused by dominant point mutation in a gene that also causes a form of human open-angle glaucoma. The phenotype is highly penetrant and reproducible in our colonies at different institutions. Thus, the Icst mutation provides a new tool for dissecting the molecular and pathologic features of IOP elevation and glaucoma, and for testing new therapies.
The animal studies described in this paper at the MRC Human Genetics Unit were carried out under the guidance issued by the Medical Research Council in Responsibility in the Use of Animals for Medical Research (July 1993) and Home Office Project Licence nos. PPL60/3124, PPL60/3785 and PPL60/4424. All experiments conducted at The Jackson Laboratory were approved by the institutional Animal Care and Use Committee. All animals were treated in accordance with the protocols established by the Association for Research in Vision and Ophthalmology. The Jackson Laboratory's pathogen surveillance program regularly screened for pathogens. Mice were housed in a 14 hour light to 10 hour dark cycle. Mutant and littermate control mice were housed together to control for cage-dependent differences. The Lmx1bIcst strain has been submitted to the European Mouse Mutant Archive (http://www. infrafrontier.eu) strain number EM:00114. Both the Lmx1bIcst and knockout lines were maintained on the C57BL/6J background. Clinical examinations were carried out as previously described [18], [48]. The BAC transgenic strain was made as described [49].
The exons and the immediate flanking sequences of Lmx1b were amplified from Icst, BALB/c, C3H and C57BL/6J genomic DNA using intronic primers that were also used for subsequent sequence analysis. PCR products were purified using Millipore Multi-screen PCR 96-well filtration system on a Biomek 2000 robotic platform and sequenced directly using Big Dye terminator cycle sequencing. Sequences were analysed using the Sequencher program.
To produce N-terminally histidine-tagged fusion proteins of full-length 372 amino acid LMX1B and the homeodomain alone, wild-type and Icst cDNAs were amplified by PCR introducing Pci I and Nco I sites at the 5′ of the full-length and homeodomain respectively and a Bam HI site at the 3′ of both and cloned into pGEM-T Easy (Promega). After digestion with Bam HI and Pci I or Nco I, as appropriate, the cDNAs were cloned into the Nco I and Bam HI sites of pET6H [50]. Recombinant histidine-tagged proteins were expressed in the Escherichia coli strain BL21 (DE3) pLysS essentially as described [51]. Proteins were analysed by sodium dodecyl sulphate–polyacrylamide gel electrophoresis to assess yield and purity and equal amounts of wild-type and mutant proteins were used in bandshift experiments. For expression in mammalian cells wild-type Lmx1b and Lmx1bIcst cDNAs were cloned into the expression vector pcDNA3.1 (Invitrogen). The original reported size of LMX1B protein is 372 amino acids [12]. An upstream in-frame ATG in the human sequence that would encode an additional 23 amino acids has been reported [22]. This sequence is conserved between mouse and human and the mouse LMX1B protein sequence in the database has been revised to include these extra 23 amino acids (entry O88609 in http://www.uniprot.org). In addition, there is a direct duplication of 18 bp encoding the first 6 amino acids of the 372 amino acid protein in the mouse genome sequence that is not conserved in human. This 18 bp sequence has been found to be absent from some Lmx1b cDNAs (e.g. BC125469) but it is found in the EST database (BY741174.1). By RT-PCR we found it to be present in Lmx1b transcripts (data not shown). We therefore made two versions of LMX1B, one 372 amino acids long (-S) and one including the additional N-terminal 29 amino acids (-L). Myc and FLAG-tagged mammalian expression vectors were made using the Gateway cloning system (Life Technologies). In brief, Lmx1bWT and Lmx1bIcst were amplified from the -L constructs described above and cloned into the donor vector pDONR™221 (Life Technologies) and then into the destination vectors pcDNA3.1Myc-HisDEST [52] and pDEST/C-SF-TAP [53] to give WT-Myc and Icst-FLAG respectively following the manufacturer's instructions. Full-length Ldb1 was amplified from mouse embryonic cDNA using primers that introduced a Hind III site at the 5′ end and a Bam HI site at the 3′ end and cloned using these sites into pcDNA3.1 (+) to give pcDNA-LDB1. All plasmids were verified by sequencing.
The FLAT probe from the Col4a4 gene intron 1 [21] was made by annealing the two oligonucleotides 5′-GGTTCATGAAAGTAATTATTTTCA-3′ and 5′-GGTTTGAAAATAATTACTTTCATG-3′ and end-labelled by filling in the four base 5′ single-stranded extensions with 32P dATP and 32P dCTP using Klenow polymerase. Bandshift analysis was carried out essentially as described [54]. 20,000 cpm FLAT probe (∼1×107 cpm/µg radiolabelled DNA probe) was used in each bandshift reaction. Bacterial extract protein concentrations were ∼5 µg/µl and we used 1, 2 or 3 µl per reaction.
Transfections were carried out using a MicroPorator MP-100 following the manufacturer's protocol (Microporator). To correct for transfection efficiency and viability, 2.5 ng of renilla reporter vector was also transfected. The day following transfection luciferase assays were carried out using the Dual-luciferase reporter assay system (Promega E1910) and readings were normalised using the renilla reporter. Each experiment was carried out in triplicate.
For anterior segment examination and photography, a Nikon FS-3V zoom slit-lamp biomicroscope was used with an attached Nikon D300S digital still camera and digital images were saved using Adobe Photoshop CS5 (Adobe, Inc.). Mouse paws were photographed using an imaging system comprising a Nikon AZ100 macroscope (Nikon UK Ltd, Kingston-on-Thames, UK) and a Qimaging Micropublisher 5 cooled colour camera (Qimaging, Burnaby, BC). Image capture was performed using in-house scripts written for IVision (BioVision Technologies, Exton, PA).
Eye histology was carried out as previously described [18], [48]. After dissection embryos were photographed and fixed overnight in 4% paraformaldehyde in PBS at 4°C. A small part of the tail was used for genotyping. After washing in PBS they were dehydrated by immersion in a series of increasing concentrations of alcohol, embedded in paraffin wax, sectioned and stained with haematoxylin and eosin. Slides were viewed on a Leica MZFLIII fluorescence stereo microscope fitted with a Coolsnap colour camera (Roper Scientific, Tucson, Arizona, USA). Image capture was controlled by in-house scripting of IPLab Spectrum (Scanalytics, Fairfax, VA, USA). For plastic-based processing, enucleated eyes were fixed (0.8% paraformaldehyde and 1.2% glutaraldehyde in 0.08 M phosphate buffer (pH 7.4)) and processed for plastic sectioning as previously described [48]. Serial sagittal sections passing through the optic nerve were collected, stained with hematoxylin and eosin, and analysed for pathologic alterations.
For all mice, IOP was measured as previously described in detail [55], [56]. Briefly, mice were acclimatised to the procedure room and anesthetized via an intraperitoneal injection of a mixture of ketamine (99 mg/kg; Ketalar, Parke-Davis, Paramus, NJ) and xylazine (9 mg/kg; Rompun, Phoenix Pharmaceutical, St. Joseph, MO) prior to IOP assessment.
Eyes were collected into 2×PBS and fixed in 2% paraformaldehyde in PBS for two minutes. After rinsing in 2×PBS for five minutes the retina was dissected and laid flat by making radial incisions and fixed in methanol at −20°C for one hour. The retinas were then fixed again in 4% paraformaldehyde in PBS for five minutes, rinsed in 2×PBS and blocked in wholemount buffer (2×PBS, 1% Cohn fraction BSA, 3% Triton X-100) for one hour and then incubated with anti-BRN3 antibody (SC6026, Santa Cruz) overnight. After three ten minutes washes in wholemount buffer the retinas were incubated in Alexa Fluor 594 donkey anti-goat secondary antibody (1/500) (Molecular Probes) in wholemount buffer for four hours. After three washes in wholemount buffer, retinas were rinsed with 2×PBS and post-fixed in 4% paraformaldehyde in PBS for five minutes, mounted in Vectashield hard set (Vector Labs). All washes and incubations were carried out at room temperature on an orbital shaker. Four images were taken at 90° angles to each other around the optic disc using the Nikon TiE-C1Si confocal microscope. The number of BRN3-positive cells in each of the areas was counted using the Velocity Image acquisition and analysis software (PerkinElmer, Waltham, MA, USA). Mice from two litters were analysed. In the first litter (age four months) two Lmx1bicst/+ mice were used and one each of the other two genotypes. In the second litter (age one month) one mouse of each genotype was used.
Optic nerves were processed and analysed as previously reported [48], [57]. Briefly, nerves were stained with paraphenylene-diamine which differentially stains single damaged axons allowing sensitive detection of axon injury. Nerves were determined to have one of three damage levels that are readily distinguishable by axon counting.
E17.5 embryonic kidneys were fixed overnight in 3% glutaraldehyde in cacodylate buffer at 4°C then post-fixed in 1% osmium tetroxide for two hours at 4°C. After dehydration through ascending grades of alcohol and propylene oxide they were impregnated with TAAB Embedding Resin (medium grade premix) and cured for 24 hours. Ultrathin sections were stained with uranyl acetate and lead citrate and viewed on a JEOL JEM 100CXII fitted with an AMT Digital Camera using the AMTv600 image capture software.
Live animals and whole mount embryos and pups were scanned using a Skyscan 1076 in vivo µCT system (Skyscan B.V., Aartselaar, Belgium). Live animals were scanned under fluothane anaesthesia. Animals were scanned at an isotropic resolution of 18.6 µm. Scans were performed at 50 kV, 200 µA using a 0.5° rotation step and a 0.5 mm aluminium filter. Higher resolution scans of limbs were performed using a Skyscan 1172 system at a resolution of 8.8 µm (60 kV, 167 mA, 0.6° rotation step, 0.5 mm aluminium filter). Scans were reconstructed using Skyscan NRecon software and analysed using Skyscan CTAn software. Three dimensional models were visualised using Skyscan CTVol software.
HEK 293T cells cultured in 100 mm dishes were transfected with WT-Myc (5 µg), ICST-FLAG (2 µg), and pcDNA-LDB1 (5 µg) as indicated in a total of 12 µg DNA adjusted with pcDNA3.1(-) as necessary using Lipofectamine LTX PLUS (Invitrogen). After 48 hours the cells were lysed using Cell Lysis Buffer (Cell Signaling Technology) and Myc-tagged complexes immunoprecipitated using the Profound c-Myc Tag IP/Co-IP Kit (Thermo Scientific) according to the manufacturer's instructions. Samples were separated on 4–12% NUPAGE gels (Life Technologies), transferred to Hybond-P (GE Healthcare) and probed with anti-Myc (Cell Signaling Technology, #2276), anti-FLAG (Cell Signaling Technology, #2368) and anti-LDB1(kind gift of Sam Pfaff [58]) antibodies using standard protocols and visualised with horseradish-peroxidase secondary antibody (GE Healthcare) and SuperSignal West Pico Chemiluminescent Substrate (Thermo Scientific) following the manufacturer's instructions.
Values were expressed as mean+standard error and P<0.05 was considered significant. For the data shown in Figure 1 a two-tailed unpaired Student's t-test was used for statistical analysis. For the data shown in Figure 3 a two-tailed unequal variance t-test was performed using JMP (http://www.jmp.com). For the ganglion cell counts data shown in Figure 4 we used log transformation of raw data to allow for a mean-variance relationship. We then tested for a mean difference between replicate mice within genotype using ANOVA. As replicate mice within genotype differed significantly, we compared variation in mean cell count between pairs of genotypes against variation between means of replicate mice within genotype using a two-tailed Student's t-test. For the data shown in Tables 1–5 chi square tests were performed using http://graphpad.com/quickcalcs.
|
10.1371/journal.ppat.1003335 | The Production of Monokaryotic Hyphae by Cryptococcus neoformans Can Be Induced by High Temperature Arrest of the Cell Cycle and Is Independent of Same-Sex Mating | Cryptococcus neoformans is a heterothallic fungal pathogen of humans and animals. Although the fungus grows primarily as a yeast, hyphae are produced during the sexual phase and during a process called monokaryotic fruiting, which is also believed to involve sexual reproduction, but between cells of the same mating type. Here we report a novel monokaryotic fruiting mechanism that is dependent on the cell cycle and occurs in haploid cells in the absence of sexual reproduction. Cells grown at 37°C were found to rapidly produce hyphae (∼4 hrs) and at high frequency (∼40% of the population) after inoculation onto hyphae-inducing agar. Microscopic examination of the 37°C seed culture revealed a mixture of normal-sized and enlarged cells. Micromanipulation of single cells demonstrated that only enlarged cells were able to produce hyphae and genetic analysis confirmed that hyphae did not arise from α-α mating or endoduplication. Cell cycle analysis revealed that cells grown at 37°C had an increased population of cells in G2 arrest, with the proportion correlated with the frequency of monokaryotic fruiting. Cell sorting experiments demonstrated that enlarged cells were only found in the G2-arrested population and only this population contained cells able to produce hyphae. Treatment of cells at low temperature with the G2 cell cycle arrest agent, nocodazole, induced hyphal growth, confirming the role of the cell cycle in this process. Taken together, these results reveal a mating-independent mechanism for monokaryotic fruiting, which is dependent on the cell cycle for induction of hyphal competency.
| Fungi typically grow vegetatively as either yeast or hyphae. Many of the major human fungal pathogens can generate both morphologies and are referred to as the dimorphic fungi. Cryptococcus neoformans is a yeast-like fungus that has not been traditionally thought to be dimorphic since hyphae production typically occurs during the mating process between cells of opposite mating types. However, C. neoformans also can generate the hyphal state from haploid cells (called monokaryotic or haploid fruiting) in the absence of the opposite mating type. Recent studies have shown that the mechanism behind this process also involves mating, however, the mating reaction occurs between cells of the same mating type. Here we describe a unique mechanism responsible for monokaryotic fruiting that is independent of mating and does not proceed through a diploid intermediate. Instead, the key requirement for hyphal induction appears to be cell cycle arrest. Importantly, arrested cells display an enlarged cell phenotype, which has been observed in vivo in recent reports and has been hypothesized to be a novel protection strategy against host defenses. C. neoformans appears to have an extensive morphological repertoire, which likely contributes to its success as both a pathogen and a saprophyte.
| Cryptococcus neoformans is a basidiomycetous fungal pathogen of humans and animals that typically causes opportunistic infections in patients with cellular immune defects [1]. Infection initiates in the lungs and frequently disseminates to the brain where it manifests as a fatal meningoencephalitis if untreated. AIDS patients are at increased risk for infection, though infection rates have decreased significantly with better AIDS management [2]. However, in spite of the reduction in AIDS-related cases, cryptococcosis remains a frequent life-threatening opportunistic mycosis for these patients in underdeveloped countries, and is a recently emergent disease in the United States Pacific Northwest [3] for as yet, unexplained reasons.
Naturally occurring strains of C. neoformans are heterothallic with two mating types, MATa and MATα, with both mating types being pathogenic, although most clinical isolates are MATα [4]. Although the taxonomy has been changing, four serotypes have been described (A, B, C, D) with serotypes A and D often referred to as C. neoformans variety grubii and variety neoformans respectively, and serotypes B and C being collectively referred to as C. neoformans variety gattii, or more recently, C. gattii [5]. For all serotypes, throughout the course of infection and under normal culture conditions, the fungus grows as an encapsulated yeast. Under appropriate in vitro conditions, however, the fungus can produce two kinds of hyphae; dikaryotic hyphae during MATa×MATα sexual reproduction and monokaryotic hyphae (from individual MATa or MATα strains) during monokaryotic fruiting [6]. Basidiospores can be produced from both hyphal types. Environmental factors required for sexual reproduction and monokaryotic fruiting are similar and include culture under low temperature (25°C), low moisture, and nutrient limitation [6]. Many genes, including homologs of the Saccharomyces cerevisiae pheromone response pathway, are required to produce both types of hyphae [7], [8], [9], [10], [11]. There are, however, distinct differences between the two hyphal types. Structurally, dikaryotic hyphae have fused clamp connections and a pair of nuclei (one MATa and one MATα) per hyphal compartment while monokaryotic hyphae have unfused clamp connections and a single nucleus per hyphal compartment. Sexual reproduction in C. neoformans has been characterized in detail and largely follows the pheromone response paradigm that has been developed from decades of S. cerevisiae research [12]. Less clear is the mechanism by which monokaryotic fruiting occurs.
Recent studies have concluded that monokaryotic fruiting in C. neoformans variety neoformans can result from α-α mating [13] and may be an important part of the natural life cycle of this fungus [14], with possible implications for human disease [15]. We recently reported that high temperature seed culture conditions could induce very robust monokaryotic fruiting in C. neoformans variety neoformans [11] and assumed these growth conditions enhanced α-α cell fusion. However, this assumption proved to be false. Instead, we found that cells arrested in the G2 stage of the cell cycle were competent to undergo monokaryotic fruiting at high frequency, in the absence of α-α cell fusion. Importantly, this mechanism proceeds through enlarged cells, a morphological phenotype that has been increasingly observed in vivo and is hypothesized to serve as a strategy for avoiding host defenses [16], [17], [18]. These results demonstrate that C. neoformans has evolved a number of different mechanisms for modifying cellular morphology to suit its specific environment, with some of these mechanisms contributing to the success of this fungus as a pathogen.
Previous studies of monokaryotic fruiting typically were performed by patching cells from a seed culture onto filament agar and then screening for a hyphal fringe around the periphery [6], [11], [19]. Our recent observation of the role of temperature in this process [11] led us to test whether or not high temperature increases the intensity of hyphae production after inoculation onto filament agar, or increases the number of cells that produce hyphae. A suspension of cells from a 24 h, 37°C seed culture was spread onto filament agar to observe monokaryotic fruiting in individual cells, which would reveal whether or not all cells from the seed culture were capable of undergoing monokaryotic fruiting. Figure 1A demonstrates that only part of the population grown at 37°C was able to undergo monokaryotic fruiting and that hyphae began to appear as early as 4 hours after plating onto filament agar (Fig. 1B). This phenomenon was not a strain artifact nor was it restricted to a single serotype as all four serotypes were found to be able to produce hyphae under the above inducing conditions (Fig. 1C–F). These results suggest that only a specific type of cell is capable of undergoing monokaryotic fruiting, and that this capability requires seed culture conditions that enable these cells to become competent for hyphal production.
Lin et al., have demonstrated that same-sex mating is one way in which monokaryotic fruiting can occur [13]. However, because the previous experiment showed that individual cells were still capable of monokaryotic fruiting in spite of being well separated from potential mating partners on spread plates, we hypothesized that either same-sex mating occurs during the seed culture growth period, or that another mechanism, which does not involve same-sex mating, could also lead to monokaryotic fruiting.
To investigate these two possibilities, we first screened for evidence of α-α mating during monokaryotic fruiting using the α-α cell fusion assay to test different combinations of complementing MATα auxotrophs that would be predicted to yield prototrophic colonies on unsupplemented MIN agar. No fusants were observed from MATα×MATα crosses plated onto MIN agar after growth as a seed culture on YPD at 37°C for 24 h. However, assisted α-α matings showed that each of these strains was capable of undergoing α-α fusion, ruling out an α-α cell fusion defect (Fig. 2A). The most likely explanation for these results was that because α-α fusion is a rare event [13], our conditions, although not detecting an α-α fusion, were still not excluding this possible mechanism. To exclude a cryptic fusion event as an explanation for our observations, we utilized a cpk1Δ mutant to further test whether or not α-α fusion was required for temperature-induced monokaryotic fruiting. Cpk1p, the MAP kinase in the C. neoformans pheromone response pathway, is required for α-a fusion during sexual reproduction as well as α-α fusion during monokaryotic fruiting [9], [13]. Therefore, if α-α fusion was the mechanism of monokaryotic fruiting in our system, the cpk1Δ mutant should not produce hyphae after plating onto filament agar because it could not fuse with the complementing strain. Our results showed that neither assisted nor unassisted α-α mating reactions (WSA-2126×WSA-591×WSA-65 or WSA-2126×WSA-591) with the cpk1Δ mutant (WSA-2126) showed evidence of α-α cell fusion (Fig. 2B). However, when WSA-2126 was tested for monokaryotic fruiting ability after high temperature seed culture, this strain fruited normally (Fig. 2B). These results confirmed that monokaryotic fruiting could occur independently of α-α fusion.
The production of enlarged cells in C. neoformans has been reported to occur when cells are exposed to opposite mating type cells in standard mating type mixes [11], and in confrontation assays, which are performed by streaking cells of opposite mating type in close proximity to each other [20]. Clinical studies have also found this cell type in vivo [16], [18], [21], [22]. Because only certain cells produced hyphae during the quantitative monokaryotic fruiting assay, we decided to determine if there were developmental differences among cells after high temperature seed culture. Microscopic observation of wet mounts prepared from cells scraped off of filament agar after growing as a seed culture at 37°C showed that filaments always originated from enlarged cells (Fig. 3A). When cells from 30°C and 37°C seed cultures were screened for enlarged cells, we only observed enlarged cells from the 37°C seed culture, although this incubation temperature produced both enlarged and smaller, normal-sized cells (Fig. 3B and 3C). To confirm that the enlarged cells were responsible for the production of hyphae during monokaryotic fruiting, large and small cells from a 37°C seed culture were micromanipulated onto filament agar and then monitored for hyphae production. Analysis of hyphae production from each single cell revealed that small cells only grew into yeast cells while the large cells grew as both yeast and hyphae (Fig. 3D). These results demonstrated that only a subset of the population grown at high temperature, which can be distinguished by the larger cell size, becomes competent to produce hyphae.
The need for enlarged cells prior to the initiation of monokaryotic fruiting suggested that the mechanism for production of this phenotype possibly involved changes in cell ploidy since yeast ploidy has been noted to be associated with cell size [23]. In C. neoformans, monokaryotic fruiting has been shown to result in ploidy changes of yeast cells produced specifically from the hyphal filaments [13]. These observations suggested to us that monokaryotic fruiting may occur through ploidy changes, which are manifested as enlarged cells that arise from an endoduplication event, as has been previously suggested [24].
Cell cycle analysis of seed cultures grown at different temperatures (25°C, 30°C, 35°C, 37°C and 40°C) all revealed only 1n and 2n DNA content peaks, with no evidence of a 4n peak (Fig. 4A). This result showed that there was no ploidy change after high temperature (35°C, 37°C, 40°C) seed culture growth, demonstrating that the enlarged cells, which were responsible for monokaryotic fruiting, were haploid. Additionally, we found that as seed culture temperature increased from 25°C to 40°C, the percentage of cells in G1 decreased, the percentage of cells in G2 increased, and the percentage of cells in S phase was similar until dropping almost to 0 at 37°C (Fig. 4B). DAPI staining confirmed the relationship between cell size and G2 arrest as the staining patterns of large and small cells differed (Fig. 4C). The smaller cells displayed a compact nuclear staining pattern while the larger cells displayed a larger, diffuse staining pattern, which has been observed in other G2-arrested fungi [25], [26]. These results demonstrated that the effect of increasing temperature on monokaryotic fruiting involved the cell cycle, specifically G2 arrest.
In spite of the FACS results showing that hyphal-competent cells were haploid, we could not rule out a change in ploidy just before, or during growth in the hyphal phase. Unfortunately ploidy determination by FACS analysis of hyphae is physically restricted by the filamentous characteristics of the cells. However, hyphal ploidy can be determined using a blastospore assay [13]. The results of this assay revealed that 76 out of 78 blastospores (97%) were haploid, demonstrating that the hyphae produced during monokaryotic fruiting were haploid, which again excluded α-α mating or endoduplication as monokaryotic fruiting mechanisms in our system (Fig. 5A). As a further control, the two diploid blastospores were sub cultured repeatedly for an additional two weeks and then retested by FACS, which revealed that they remained diploid, ruling out the possibility that haploid blastospores could be segregation products of diploid blastospores. While a late endoduplication event could explain the two diploid spores, this possibility is unlikely since the 78 spores were picked from 78 independent hyphae. However based on the recovery of two diploid blastospores in our assay, and the rarity of basidiospore production during monokaryotic fruiting, we hypothesized that there could be two hyphal types produced during monokaryotic fruiting, which could be distinguished by ploidy. One type could be vegetative in nature and not undergo basidiosporogenesis (haploid) and a second type could be generated that ultimately produced basidiospores (diploid). To address these two possibilities, 37°C seed culture cells were spread onto filament agar and the resultant colonies screened for basidiospore chains. The blastospore assay was performed on blastospores recovered from hyphae with and without basidiospores. Cell cycle analysis again showed that all of the blastospores were haploid, regardless of whether or not the hypha produced basidiospores (Fig. 5B), which was consistent with our previous observations that excluded α-α mating or endoduplication as the mechanisms of monokaryotic fruiting. Together, these results demonstrate that endoduplication is not required for monokaryotic fruiting as these hyphae are produced from haploid cells and remain haploid, in spite of being able to occasionally generate diploid yeast cells.
To determine whether only G2-arrested cells undergo monokaryotic fruiting, we sorted G1 and G2 phase, 37°C seed culture cells according to DNA content. Microscopic observation of sorted G1 and G2 cells revealed that G2 phase cells were much larger than G1 phase cells (Fig. 6). We next sorted live 37°C seed culture cells according to cell size and performed cell cycle analysis on the two populations. The results indicated that the smaller cells had a single DNA content peak at the 1n position while the enlarged cells had a single DNA content peak at the 2n position, which lead us to conclude that the small cells were G1 phase cells and the large cells were cells in G2 arrest (Fig. 7). The two populations were then assayed for monokaryotic fruiting ability, which revealed that monokaryotic fruiting was a property only of the G2 fraction (Fig. 7).
As a final confirmation that monokaryotic fruiting requires G2 arrest, cells were treated with nocodazole, a G2/M arrest agent that inhibits and disassembles microtubules [27]. Cells were grown as seed cultures at 30°C (the non-permissive seed culture temperature), then assayed for monokaryotic fruiting ability. The experiment showed that cells treated with nocodazole became competent to produce hyphae on filament agar in a dose-dependent manner even though they were grown as a seed culture under conditions that did not normally lead to monokaryotic fruiting, whereas untreated cells only grew as yeasts (Fig. 8). Taken together, these results demonstrate that G2-arrested cells can serve as a starting point for cells that undergo monokaryotic fruiting.
In this study we have identified a novel mechanism in C. neoformans that leads to the production of hyphae, with or without basidiospores, by haploid cells (monokaryotic fruiting). This mechanism appears to be dependent on the cell cycle and initiates from cells in G2 arrest. It occurs in the absence of α-α mating and/or endoduplication, thereby demonstrating that monokaryotic fruiting can occur asexually. Previous studies have shown that sexual reproduction can occur between cells of the same mating type, resulting in monokaryotic fruiting [13], and that this phenomenon occurs in nature [24]. Under the specific conditions of this study, notably a 37°C seed culture temperature, we saw no evidence of an α-α cell fusion event, nor did we find evidence of endoduplication within the hyphae even though we screened hyphae that had produced basidiospores. During C. neoformans basidiosporogenesis, sexual reproduction results in meiosis in the basidium followed by successive mitotic divisions that yield the nuclei, which ultimately are inserted into spores as they form on the basidial surface [20]. Lin et al. observed that when fruiting was derived from an α-α fusant, sporogenesis was robust with spore chains that were long and phenotypically similar to α-a mating during sexual reproduction [13]. This process was found to be impaired in dmc1 mutants, which are meiotic mutants that still produce spores, but at a much lower frequency than sexually produced spores, and with truncated spore chains that sometimes occur as dyads (two rather than four chains) [13]. The phenotype of basidiospores produced in the dmc1 strains was strikingly similar to what we observed in this study and what was previously reported [6]. These observations may suggest that basidiosporogenesis can occur mitotically without meiosis, although we cannot exclude a duplication event in the basidium immediately followed by meiosis and sporogenesis. We did, however, test a dmc1 mutant and found that it was able to undergo monokaryotic fruiting under our conditions (data not shown).
Because our study was done in the same strain background as the study by Lin et al., we reviewed the conditions of both experiments and found some differences that may explain the contrasting differences in ploidy. Our study used a high temperature seed culture condition, which results in rapid hyphae production upon filament agar plating. The seed culture conditions in Lin's study were not clear, however, their plating medium was V8 agar, which is normally used for mating C. neoformans, and their incubation period was for a period of weeks, whereas we screened at 24 hrs and observed hyphae in as little as four hours, although cells also produced hyphae on V8 agar under our conditions. Both filament agar and V8 agar have high agar contents; however, V8 agar is an undefined medium with V8 juice as the basal ingredient. Filament agar, on the other hand, is a low-glucose, defined medium with Yeast Nitrogen Base without amino acids and without ammonium sulfate as the source of vitamins and cofactors. Although both media are starvation media, they are substantially different in composition, which may be one explanation for the differences in hyphal types that we observed. The seed culture conditions, or more precisely, the cell cycle stage may be another explanation.
Our initial investigation of the relationship between the cell cycle and monokaryotic fruiting focused on detecting what we presumed would be a transition to diploidy in yeast cells at some point during seed culture growth, which we believed would coincide with the appearance of enlarged cells in the seed culture and the association of this cell type with the hyphal progenitor. The high frequency of fruiting and enlarged cells in the seed culture suggested that detection of the diploidization event would be unambiguous. However, the data showed that instead of an α-α fusion or endoduplication event, which would result in a diploid cell, the actual mechanism that resulted in hyphae production was G2 arrest. The reason for the requirement of G2 arrest to induce hyphae is not clear, and while G2 arrest is required for monokaryotic fruiting, not all arrested cells produced hyphae. We suspect that the subpopulation of non-hyphal, G2-arrested cells consisted of cells that escaped G2, and then proceeded to bud rather than differentiate into a hypha. These two outcomes resemble the decision point that a pheromone-exposed, G2-arrested Ustilago maydis cell faces with regard to which of the two developmental paths it will follow (conjugation tube formation or budding) [26]. How the generation of occasional diploid blastospores occurs is also not clear. Given that C. neoformans hyphae produce typical basidiomycete-like clamp connections, the incomplete fusion of these structures in monokaryotic hyphae combined with an aberrant segregation event during the budding of blastospores off of the hyphal compartment may yield the diploid cells that we observed at low frequency.
Previous studies of the C. neoformans cell cycle have identified a number of stressors that cause G2 arrest, including oxygen depletion [28], stationary growth phase [29], and temperature [30]. Under our experimental conditions, hyphal competency occurs prior to transfer to hyphal inducing conditions (starvation on filament agar) and not during growth on filament agar. Therefore, stationary growth phase is not a factor nor is oxygen depletion since cultures were grown on the agar surface, and only for 24 hrs. Consequently, growth temperature seems to be responsible for inducing competency. Under our conditions, the 37°C incubation temperature differs from the original incubation temperature (30°C) for monokaryotic fruiting [6], which suggests that temperature is an important variable. In fact, while we did not see the temperature effect on all strains of C. neoformans, we were able to induce hyphae in all four serotypes. Interestingly, the reports of enlarged cells in vivo [16], [18], [21], [22] reflect growth at elevated temperature in the mammalian body. Other pathways have also been shown to influence C. neoformans cell size in vitro, including cAMP, RAS, and PKA [16], [31], [32], suggesting the possibility of conserved stimuli that may regulate these pathways. Presently, the STE12α signal transduction pathway seems to be the major or sole regulator of monokaryotic fruiting as this gene is required for monokaryotic fruiting regardless of inducing conditions. Fruiting still occurred normally in cpkΔ1 (pheromone response pathway), and cacΔ1 (cAMP pathway) mutants, ruling out these pathways as regulators of temperature-induced monokaryotic fruiting. Other pathways, such as the calcineurin signal transduction pathway cannot be ruled out, but are more complicated to test since some mutants in this pathway do not grow at high temperature [33].
A key characteristic of enlarged cells in vivo appears to be polyploidy, which has been hypothesized to arise when the M phase of the cell cycle is skipped [16], [18]. This enlarged cell phenotype appears to be a potential virulence factor as they are poorly phagocytized, if at all [18]. Perhaps increasing cell size evolved as a physical defense mechanism against predatory grazers, which in turn, protects cells from being phagocytized in vivo via the same mechanisms. It appears that the enlarged cell phenotype can be produced by multiple mechanisms: cell cycle arrest, a-α or α-α cell fusion, and endoduplication, each of which may have a different purpose. The developmental options available after cell cycle arrest may have been selected for in C. neoformans to enhance survival in its specific environmental niche while inadvertently creating an important human fungal pathogen. What remains to be determined is how high temperature generation of the large cell and monokaryotic fruiting phenotypes was incorporated into the evolution of C. neoformans. With the exception of C. neoformans, virtually all members of this genus grow poorly or not at all at mammalian ambient temperature. In contrast, all of the major human fungal pathogens grow at 37°C and virtually all of them have a hyphal phenotype. Perhaps the association of many basidiomycetes with rotting wood or decaying vegetation in general led to high temperature exposure and subsequent genetic selection during self-heating, compost-like conditions caused by microbial metabolism of organic matter. Once nutrients were consumed, hyphal extension towards additional nutrients and/or sporulation in the absence of nutrients could have completed the evolution of C. neoformans into a pathogen via development of a mechanism of infectious particle (basidiospores) dispersion combined with the ability to grow at elevated temperatures. Selection for the molecular linkage of pathways controlling cell cycle, nutrient sensing, and ultimately, differentiation, could have been the outcome of this lifestyle and allowed the fungus to coordinately regulate these pathways, thus enabling it to effectively exist as a saprophyte or pathogen.
YPD agar, MIN agar, V8 agar, and filament agar were prepared as described previously [6], [7], [11] with or without amino acids or nucleic acid supplements as required. JEC-21 is a wild type, MATα isolate that was used in the initial characterization of monokaryotic fruiting in C. neoformans [34]. WSA-79 is a serotype D clinical isolate from Maryland, WSA-522 is a serotype A clinical isolate from Thailand, WSA-533 is a serotype B environmental isolate from Australia, and WSA-2507 is a serotype C clinical isolate from Maryland. Additional strains, all of which were derived from the original JEC-21 - JEC-20 congenic pair [34], consisted of the following genotypes: WSA-1 (MATα lys2), WSA-70 (MATα ade2 lys2), WSA-591 (MATα ade2), WSA-1226 (MATα ura5), WSA-2126 (MATα ura5 cpk1Δ::ura5), WSA-3002 (MATα/MATα fusant from WSA-1226×WSA-70×WSA-68), WSA-3019 (MATα haploid blastospore), WSA-3070 (MATα/MATα diploid blastospore recovered from JEC-21 hypha without basidiospore chains), WSA-3098 (MATα haploid blastospore recovered from JEC-21 hypha with basidiospore chains), WSA-3112 (MATα haploid blastospore recovered from hypha without basidiospore chains), WSA-3145 (MATα ura5 cpk1Δ::ura5 CPK1::NEOr). MATa strains included WSA-65 (MATa ura5 lys1 ade2) and WSA-68 (MATa ura5 ade2 lys2).
Nocodazole (Sigma-Aldrich, St. Louis, MO) stock was prepared in DMSO at 1.5 mM and then used to prepare different dilutions in YPD broth (YPD-0.075 µm nocodazole, YPD-0.15 µm nocodazole, YPD-0.30 µm nocodazole, YPD-0.60 µm nocodazole, YPD-1.2 µm nocodazole). JEC-21 cells were added to 1.5 ml nocodazole-YPD broth in 15 ml snap cap tubes (BD Biosciences, Franklin Lakes, NJ), at a final concentration of 1×106 cells/ml. Tubes were shaken at 200× RPM at 30°C for 24 hrs. Five µl of overnight culture were then dropped onto filament agar and incubated at 25°C for 5 days. The quantitative assay was performed by growing cells as above, and then plating cells onto filament agar as described in the quantitative monokaryotic fruiting assay. Pictures were taken with an Olympus SZX12 stereo microscope (Olympus, Center Valley, PA) at 2× magnification.
To perform α-α cell fusions, 1×106 cells of complementary, auxotrophic, MATα strains were mixed, cultured on YPD agar at 37°C for 24 hrs, and then transferred onto filament agar plates. The plates were incubated at 25°C for 24 hrs. The mixture was scraped from the plate, suspended in 1.0 ml sterile H2O, and then 200 µl of cells from this suspension were spread onto MIN agar plates, which were then incubated at 30°C for 4 days to screen for fusants. Assisted mating reactions, in which two MATα strains were induced to fuse by including a MATa helper strain, were performed according to previously described methods [19]. Plates were photographed at 0.5× magnification.
Cells were harvested, fixed, and stained with propidium iodide (Sigma-Aldrich, St. Louis, MO), and then sorted as described by Sia et al. [35]. Cell cycle analysis was performed using a BD FACSCalibur flow cytometer (Becton Dickinson Biosciences, Sparks, MD). CellQuest Pro software was used for cell collection, and data analysis was performed using ModFit and FlowJo. G1, S, and G2 phases were identified using the Dean-Jett-Fox mathematical model. G2-arrested cells were identified as described [25], [26], [36], [37], [38]. This population of cells typically shows a large cell phenotype, an increased proportion of cells with 2C DNA content when compared to controls, and a DAPI staining pattern that shows larger nuclei vs. smaller condensed nuclei of G2 phase cells (see below). Cell sortings were performed on a BD FACSAria III (Beckton Dickinson Biosciences) cell sorter and analyzed using BD FACSDiva 6.1 software. Aliquots of sorted live cells were also used to perform the quantitative monokaryotic fruiting assay. All FACS analysis was performed at the Flow Cytometry Core Laboratory at The University of Texas Health Science Center at San Antonio.
Cells were stained with DAPI according to the method described by Fuchs et al. [39]. Briefly, 1×107 cells were harvested from 24 h YPD agar plates, which were grown at either 30°C or 37°C, washed twice with phosphate buffer (0.1 M KH2PO4, 1.25 mM EGTA, 1.25 mM MgCl2 at pH 6.9), then resuspended in 400 µl fixative (5% paraformaldehyde in wash buffer) followed by incubation at room temperature for 90 minutes. The cells were then washed three times in wash buffer and incubated with Lysing Enzymes (Trichoderma harzianum, Sigma-Aldrich, 1 mg/ml in sterile distilled water) for 20 min at 37°C. The cells were then washed once with sterile water, gently resuspended in 400 µl 0.3% Triton X-100 (Sigma-Aldrich), and permeabilized by incubation at room temperature for 15 minutes. The suspension was then pelleted and washed three times with PBS. Cells were stained in DAPI (Sigma-Aldrich) (0.1, 0.2, or 0.5 µg/ml) for 15 minutes, washed twice with PBS, and then resuspended in 200 µl prior to visualization.
Images were captured on a Zeiss AxioImager Z1 microscope (Carl Zeiss Microscopy, LLC, Thornwood, NY) equipped with an AxioCam MR3_2 CCD camera, using the filters for DAPI (365 nm excitation, 395 nm beam splitter, 420–470 nm emission filter) and differential interference contrast (DIC, Nomarski contrast). Image analysis and adjustments were performed using Axiovision software (Zeiss, Version 4.8). Images were adjusted only for frame alignment (to overlay DAPI over DIC), brightness, and contrast (adjustment of +0.01 contrast units over default), and all images received the same treatment.
|
10.1371/journal.pcbi.1007321 | The FACTS model of speech motor control: Fusing state estimation and task-based control | We present a new computational model of speech motor control: the Feedback-Aware Control of Tasks in Speech or FACTS model. FACTS employs a hierarchical state feedback control architecture to control simulated vocal tract and produce intelligible speech. The model includes higher-level control of speech tasks and lower-level control of speech articulators. The task controller is modeled as a dynamical system governing the creation of desired constrictions in the vocal tract, after Task Dynamics. Both the task and articulatory controllers rely on an internal estimate of the current state of the vocal tract to generate motor commands. This estimate is derived, based on efference copy of applied controls, from a forward model that predicts both the next vocal tract state as well as expected auditory and somatosensory feedback. A comparison between predicted feedback and actual feedback is then used to update the internal state prediction. FACTS is able to qualitatively replicate many characteristics of the human speech system: the model is robust to noise in both the sensory and motor pathways, is relatively unaffected by a loss of auditory feedback but is more significantly impacted by the loss of somatosensory feedback, and responds appropriately to externally-imposed alterations of auditory and somatosensory feedback. The model also replicates previously hypothesized trade-offs between reliance on auditory and somatosensory feedback and shows for the first time how this relationship may be mediated by acuity in each sensory domain. These results have important implications for our understanding of the speech motor control system in humans.
| Speaking is one of the most complex motor tasks humans perform, but it’s neural and computational bases are not well understood. We present a new computational model that generates speech movements by comparing high-level language production goals with an internal estimate of the current state of the vocal tract. This model reproduces many key human behaviors, including making appropriate responses to multiple types of external perturbations to sensory feedback, and makes a number of novel predictions about the speech motor system. These results have implications for our understanding of healthy speech as well as speech impairments caused by neurological disorders. They also suggest that the mechanisms of control are shared between speech and other motor domains.
| Producing speech is one of the most complex motor activities humans perform. To produce even a single word, the activity of over 100 muscles must be precisely coordinated in space and time. This precise spatiotemporal control is difficult to master, and is not fully adult-like until the late teenage years [1]. How the brain and central nervous system (CNS) controls this complex system remains an outstanding question in speech motor neuroscience.
Early models of speech relied on servo control [2]. In this type of feedback control schema, the current feedback from the plant (thing to be controlled–for speech, this would be the articulators of the vocal tract, as well as perhaps the phonatory and respiratory systems) is compared against desired feedback and any discrepancy between the current and desired feedback drives the generation of motor commands to move the plant towards the current production goal. A challenge for any feedback control model of speech is the short, rapid movements that characterize speech motor behavior, with durations in the range of 50-300 ms. This is potentially shorter than the delays in the sensory systems. For speech, measured latencies to respond to external perturbations of the system range from 20-50 ms for unexpected mechanical loads [3, 4] to around 150 ms for auditory perturbations [5, 6]. Therefore, the information about the state of the vocal tract conveyed by sensory feedback to the CNS is delayed in time. Such delays can cause serious problems for feedback control, leading to unstable movements and oscillations around goal states. Furthermore, speech production is possible even in the absence of auditory feedback, as seen in the ability of healthy speakers to produce speech when auditory feedback is masked by loud noise [7, 8]. All of the above factors strongly indicate that speech cannot be controlled purely based on feedback control.
Several alternative approaches have been suggested to address these problems with feedback control in speech production and other motor domains. One approach, the equilibrium point hypothesis [9–11], relegates feedback control to short-latency spinal or brainstem circuits operating on proprioceptive feedback, with high-level control based on pre-planned feedforward motor commands. Speech models of this type, such as the GEPPETO model, are able to reproduce many biomechanical aspects of speech but are not sensitive to auditory feedback [12–16]. Another approach is to combine feedback and feedforward controllers operating in parallel [17, 18]. This is the approach taken by the DIVA model [19–22], which combines feedforward control based on desired articulatory positions with auditory and somatosensory feedback controllers. In this way, DIVA is sensitive to sensory feedback (via the feedback controllers) but capable of producing fast movements despite delayed or absent sensory feedback (via the feedforward controller).
A third approach, widely used in motor control models outside of speech, relies on the concept of state feedback control [23–26]. In this approach, the plant is assumed to have a state that is sufficiently detailed to predict the future behavior of the plant, and a controller drives the state of the plant towards a goal state, thereby accomplishing a desired behavior. A key concept in state feedback control is that the true state of the plant is not known to the controller; instead, it is only possible to estimate this state from efference copy of applied controls and sensory feedback. The internal state estimate is computed by first predicting the next plant state based on the applied controls. This state prediction is then used to generate predictions of expected feedback from the plant, and a comparison between predicted feedback and actual feedback is then used to correct the state prediction. Thus, in this process, the actual feedback from the plant only plays an indirect role in that it is only one of the inputs used to estimate the current state, making the system robust to feedback delays and noise.
We have earlier proposed a speech-specific instantiation of a state feedback control system [27]. The primary purpose of this earlier work was to establish the plausibility of the state feedback control architecture for speech production and suggest how such an architecture may be implemented in the human central nervous system. Computationally, our previous work built on models that have been developed in non-speech motor domains [24, 25]. Following this work, we implemented the state estimation process as a prototypical Kalman filter [28], which provides an optimal posterior estimate of the plant state given a prior (the efference-copy-based state prediction) and a set of observations (the sensory reafference), assuming certain conditions such as a linear system. We subsequently implemented a one-dimensional model of vocal pitch control based on this framework [29].
However, the speech production system is substantially more complex than our one-dimensional model of pitch. First, speech production requires the multi-dimensional control of redundant and interacting articulators (e.g., lips, tongue tip, tongue body, jaw, etc.). Second, speech production relies on the control of high-level task goals rather than direct control of the articulatory configuration of the plant (e.g., for speech, positions of the vocal tract articulators). For example, speakers are able to compensate immediately for a bite block which fixes the jaw in place, producing essentially normal vowels [30]. Additionally, speakers react to displacement of a speech articulator by making compensatory movements of other articulators: speakers lower the upper lip when the jaw is pulled downward during production of a bilabial [b] [3, 4], and raise the lower lip when the upper lip is displaced upwards during production of [p] [31]. Importantly, these actions are not reflexes, but are specific to the ongoing speech task. No upper lip movement is seen when the jaw is displaced during production of [z] (which does not require the lips to be close), nor is the lower lip movement increased if the upper lip is raised during production of [f] (where the upper lip is not involved). Together, these results strongly indicate that the goal of speech is not to achieve desired positions of each individual speech articulator, but must rather be to achieve some higher-level goal. While most models of speech motor production thus implement control at a higher speech-relevant level, the precise nature of these goals (vocal tract constrictions [32–34], auditory patterns [2, 12, 21], or both [14, 22]) remains an ongoing debate.
One prominent model that employs control of high-level speech tasks rather than direct control of articulatory variables is the Task Dynamic model [32, 35]. In Task Dynamics, the state of the plant (current positions and velocities of the speech articulators) is assumed to be available through proprioception. Importantly, this information is not used to directly generate an error or motor command. Rather, the current state of the plant is used to calculate values for various constrictions in the vocal tract (e.g., the distance between the upper and lower lip, the distance between the tongue tip and palate, etc.). It is these constrictions, rather than the positions of the individual articulators, that constitute the goals of speech production in Task Dynamics.
The model proposed here (Feedback Aware Control of Tasks in Speech, or FACTS) extends the idea of articulatory state estimation from the simple linear pitch control mechanism of our previous SFC model to the highly non-linear speech articulatory system. This presents three primary challenges: first, moving from pitch control to articulatory control requires the implementation of control at a higher level of speech-relevant tasks, rather than at the simpler level of articulator positions. To address this issue, FACTS is built upon the Task Dynamics model, as described above. However, unlike the Task Dynamics model, which assumes the state of the vocal tract is directly available through proprioception, here we model the more realistic situation in which the vocal tract state must be estimated from an efference copy of applied motor commands as well as somatosensory and auditory feedback. The second challenge is that this estimation process is highly non-linear. This required that the implementation of the observer as a Kalman filter in SFC be altered, as this estimation process is only applicable to linear systems. Here, we implement state estimation as an Unscented Kalman Filter [36], which is able to account for the nonlinearities in the speech production system and incorporates internal state prediction, auditory feedback, and somatosensory feedback. Lastly, the highly non-linear mapping between articulatory positions and speech acoustics must be approximated in order to predict the auditory feedback during speech. Here, we learn the articulatory-to-acoustic mapping using Locally Weighted Projection Regression or LWPR [37]. Thus, in the proposed FACTS model, we combine the hierarchical architecture and task-based control of Task Dynamics with a non-linear state-estimation procedure to develop a model that is capable of rapid movements with or without sensory feedback, yet is still sensitive to external perturbations of both auditory and somatosensory feedback.
In the following sections, we describe the architecture of the model and the computational implementation of each model process. We then describe the results of a number of model simulations designed to test the ability of the model to simulate human speech behavior. First, we probe the behavior of the model when sensory feedback is limited to a single modality (somatosensation or audition), removed entirely, or corrupted by varying amounts of noise. Second, we test the ability of the model to respond to external auditory or mechanical perturbations. In both of these sections, we show that the model performs similarly to human speech behavior and makes new, testable predictions about the effects of sensory deprivation. Lastly, we explore how sensory acuity in both the auditory and somatosensory pathways affects the model’s response to external auditory perturbations. Here, we show that the response to auditory perturbations depends not only on auditory acuity but on somatosensory acuity as well, demonstrating for the first time a potential computational principle that may underlie the demonstrated trade-off in response magnitude to auditory and somatosensory perturbations across human speakers.
A schematic control diagram of the FACTS model is shown in Fig 1. Modules that build on Task Dynamics are shown in blue, and the articulatory state estimation process (or observer) is shown in red. Following Task Dynamics, speech tasks in the FACTS model are hypothesized to be desired constrictions in the vocal tract (e.g., close the lips for a [b]). Each of these speech tasks, or gestures, can be specified in terms of it’s constriction location (where in the vocal tract the constriction is formed) and it’s constriction degree (how narrow the constriction is). We model each gesture as a separate critically-damped second-order system [32]. Interestingly, similar dynamical behavior has been seen at a neural population level during the planning and execution of reaching movements in non-human primates [38, 39] and recently in human speech movements [40], suggesting that a dynamical systems model of task-level control may be an appropriate first approximation to the neural activity that controls movement production. However, the architecture of the model would also allow for tasks in other control spaces, such as auditory targets (c.f. [13, 19]), though an appropriate task feedback control law for such targets would need to be developed (consistent with engineering control theory, we refer to the term “controller” as a “control law”). To what extent, if any, incorporation of auditory targets would impact or alter the results presented here is not immediately clear. This is a promising avenue for future research, as it may provide a way to bridge existing models which posit either constriction- or sensory-based targets. We leave such explorations for future work, but note here that the results presented here may apply only to the current formulation of FACTS with constriction-based targets.
FACTS uses as the Haskins Configurable Articulatory Synthesizer (or CASY) [41–43] as the model of the vocal tract plant being controlled. The relevant parameters of the CASY model required to move the tongue body to produce a vowel (and which fully describe the articulatory space for the majority of the simulations in this paper) are the Jaw Angle (JA, angle of the jaw relative to the temporomandibular joint), Condyle Angle (CA, the angle of the center of the tongue relative to the jaw, measured at the temporomandibular joint), and the Condyle Length (CL, distance of the center of the tongue from the temporomandibular joint along the Condyle Angle). The CASY model is shown in Fig 2.
The model begins by receiving the output from a linguistic planning module. Currently, this is implemented as a gestural score in the framework of Articulatory Phonology [33, 34]. These gestural scores list the control parameters (e.g., target constriction degree, constriction location, damping, etc.) for each gesture in a desired utterance as well as each gesture’s onset and offset times. For example, the word “mod” ([mad]) has four gestures: simultaneous activation of a gesture driving closure at the lips for [m], a gesture driving an opening of the velum for nasalization of [m], and a gesture driving a wide opening between the tongue and pharyngeal wall for the vowel [a]. These are followed by a gesture driving closure between the tongue tip and hard palate for [d] (Fig 3).
The task state feedback control law takes these gestural scores as input and generates a task-level command based on the current state of the ongoing constriction tasks. In this way, the task-level commands are dependent on the current task-level state. For example, if the lips are already closed during production of a /b/, a very different command needs to be generated than if the lips are far apart. These task-level commands are converted into motor commands that can drive changes in the positions of the speech articulators by the articulatory state feedback control law, using information about the current articulatory state of the vocal tract. The motor commands generate changes in the model vocal tract articulators (or plant), which are then used to generate an acoustic signal.
The articulatory state estimator (sometimes called an observer in other control models) combines a copy of the outgoing motor command (or efference copy) with auditory and somatosensory feedback to generate an internal estimate of the articulatory state of the plant. First, the efference copy of the motor command is used (in combination with the previous aritculatory state estimate) to generate a prediction of the articulatory state. This is then used by a forward model (learned here via LWPR) to generate auditory and somatosensory predictions, which are compared to incoming sensory signals to generate sensory errors. Subsequently, these sensory errors are used to correct the state prediction to generate the final state estimate.
The final articulatory state estimate is used by the articulatory state feedback control law to generate the next motor command, as well as being passed to the task state estimator to estimate the current task state, or values (positions) and first derivatives (velocities) of the speech tasks (note the Task State was called the Vocal Tract State in earlier presentations of the model [44, 45]). Finally, this estimated task-level state is passed to the task state feedback control law to generate the next task-level command.
A more detailed mathematical description of the model can be found in the methods.
Here we present results showing the accuracy of the learned forward model and of various modeling experiments designed to test the ability of the model to qualitatively replicate human speech motor behavior under various conditions, including both normal speech as well as externally perturbed speech.
Fig 4 visualizes a three dimensional subspace of the learned mapping from the 10-dimensional articulatory state space to the 3-dimensional space of formant frequencies (F1—F3). Specifically, we look at the mapping from the tongue condyle length and condyle angle to the first (see Fig 4A–4C) and second formants (see Fig 4D–4F), projected onto each two-dimensional plane. We also plot normalized histograms of the number of receptive fields that cover each region of the space (represented as a heatmap in Fig 4A and 4D and with a thick blue line in the other subplots). In each figure, the size of the circles is proportional to the absolute value of the error between the actual and predicted formant values. Overall, the fit of the model results in an average error magnitude of 4.2 Hz (std., 9.1 Hz) for F1 and 6.6 Hz (std., 19.1 Hz) for F2. For comparison, the range of the data was 310–681 Hz for F1 and 930–2197 Hz for F2. Fit error increases in regions of the space that are relatively sparsely covered by receptive fields. In addition, the higher frequency of smaller circles at the margins of the distribution (and therefore the edges of the articulatory space) suggest that we may need fewer receptive fields to cover these regions. Of course, this means that we do see some bigger circles in these regions where the functional mapping is not adequately represented by a small number of fields. Also note that we are only plotting the number of receptive fields that are employed to cover a given region of articulatory space, and this is not indicative of how much weight they carry in representing that region of space.
How does the presence or absence of sensory feedback affect the speech motor control system? While there is no direct evidence to date on the effects of total loss of sensory information in human speech, some evidence comes from when sensory feedback from a single modality is attenuated or eliminated. Notably, the effects of removing auditory and somatosensory feedback differ. In terms of auditory feedback, speech production is relatively unaffected by it’s absence: speech is generally unaffected when auditory feedback is masked by loud masking noise [7, 8]. However, alterations to somatosensory feedback have larger effects: blocking oral tactile sensation through afferent nerve injections or oral anesthesia leads to substantial imprecision in speech articulation [46, 47].
Fig 5 presents simulations from the FACTS model testing the ability of the model to replicate the effects of removing sensory feedback seen in human speech. All simulations modeled the vowel sequence [ǝ a i]. 100 simulations were run for each of four conditions: normal feedback (5B), somatosensory feedback only (5C), auditory feedback only (5D), and no sensory feedback (5E). For clarity, only the trajectory of the tongue body in the CASY articulatory model is shown for each simulation.
In the normal feedback condition (Fig 5B), the tongue lowers from [ǝ] to [a], then raises and fronts from [a] to [i]. Note that there is some variability across simulation runs due to the noise in the motor and sensory systems. This variability is also found in human behavior and the ability of the state feedback control architecture to replicate this variability is a strength of this approach [25].
The effect of removing auditory feedback (Fig 5C) leads to a significant, though small, increase in the variability of the tongue body movement as measured by the tongue location at the movement endpoint (Fig 5I), though this effect was not seen in measures of Condyle Angle or Condyle Length variability (Fig 5G and 5H). Interestingly, while variablity increased, prediction error slightly decreased in this condition (Fig 5F). Overall, these results are consistent with experimental results that demonstrate that speech is essentially unaffected, in the short term, by the loss of auditory information (though auditory feedback is important for pitch and amplitude regulation [48] as well as to maintain articulatory accuracy in the long term [48–50]).
Removing somatosensory feedback while maintaining auditory feedback (Fig 5B) leads to an increase in both variability across simulation runs as well as an increase in prediction error (Fig 5F–5I). This result is broadly consistent with the fact that reduction of tactile sensation via oral anaesthetic or nerve block leads to imprecise articulation for both consonants and vowels [47, 51] (though the acoustic effects of this imprecision may be less perceptible for vowels [47]). However, a caveat must be made that our current model does not include tactile sensation, only proprioceptive information. Unfortunately, it is impossible to block proprioceptive information from the tongue, as that afferent information is likely carried along the same nerve (hypoglossal nerve) as the efferent motor commands [52]. It is difficult to prove, then, exactly how a complete loss of proprioception would affect speech. Nonetheless, the current model results are consistent with studies that how shown severe dyskinesia in reaching movements after elimination of proprioception in non-human primates (see [53] for a review) and in human patients with severe proprioceptive deficits [54]. In summary, although the FACTS model currently includes only proprioceptive sensory information rather than both proprioceptive and tactile signals, these simulation results are consistent with a critical role for the somatosensory system in maintaining the fine accuracy of the speech motor control system.
While removal of only auditory feedback lead to only small increases in variability (in both FACTS simulations and human speech), our simulations show speech in the complete absence of sensory feedback (Fig 5E) shows much larger variability than the absence of either auditory or somatosensory feedback alone. This is consistent with human behvaior [51], and occurs because without sensory feedback there is no way to detect and correct for the noise inherent in the motor system (shown by the large prediction errors and increased articulatory variability in Fig 5F).
The effects of changing the noise levels in the system can be see in Fig 5J and 5K. For these simulations, only one type of feedback was used at a time: somatosensory (cyan) or auditory (purple). Noise levels (shown on the x axis) reflect both the sensory system noise and the internal estimate of that noise, which were set to be equal. Each data point reflects 100 stable simulations. Data for the acoustic-only simulations are not shown for noise levels below 1e-5 as the model became highly unstable in these conditions to due inaccurate articulatory state estimates (the number of unstable or divergent simulations is shown in Fig 5J). For the somatosensory system, the prediction error and articulatory variability (shown here for the Condyle Angle) decrease as the noise decreases. However, for the auditory system, both prediction error and articulatory variability increase as the noise decreases. Because of the Kalman gain, decreased noise in a sensory or predictive signal leads not only to a more accurate signal, but also to a greater reliance on that signal compared to the internal state prediction. When the system relies more on the somatosensory signal, this results in a more accurate state estimate as the somatosensory signal directly reflects the state of the plant. When the system relies more on the auditory signal, however, this results in a less accurate state estimate as the auditory signal only indirectly reflects the state of the plant as a result of the nonlinear, many-to-one articulatory-to-acoustic mapping of the vocal tract. Thus, relying principally on the auditory signal to estimate the state of the speech articulators leads to inaccuracies in the final estimate and, subsequently, high trial-to-trial variability in movements generated from these estimates.
In sum, FACTS is able to replicate the variability seen in human speech, as well as qualitatively match the effects of both auditory and somatosensory masking on speech accuracy. While the variability of human speech in the absence of proprioceptive feedback remains untested, the FACTS simulation results make a strong prediction that could be empirically tested in future work if some manner of blocking or altering proprioceptive signals could be devised.
When a downward mechanical load is applied to the jaw during the production of a consonant, speakers respond by increasing the movements of the other speech articulators in a task-specific manner to achieve full closure of the vocal tract [3, 4, 31]. For example, when the jaw is perturbed during production of a bilabial stops /b/ or /p/, the upper lip moves downward to a greater extent than normal to compensate for the lower jaw position. This upper lip lowering is not found for jaw perturbations during /f/ or /z/, indicating it is specific to sounds produced using the upper lip. Conversely, tongue muscle activity is larger following jaw perturbation for /z/, which involves a constriction made with the tongue tip, but not for /b/, for which the tongue is not actively involved.
The ability to sense and compensate for mechanical perturbations relies on the somatosensory system. We tested the ability of FACTS to reproduce the task-specific compensatory responses to jaw load application seen in human speakers by applying a small downward acceleration to the jaw (Jaw Angle parameter in CASY) starting midway through a consonant closure for stops produced with the lips (/b/) and tongue tip (/d/). The perturbation continued to the end of the word. As shown in Fig 6A, the model produces greater lowering of the upper lip (as well as greater raising of the lower lip) when the jaw is fixed during production of /b/, but not during /d/, mirroring the observed response in human speech.
In addition to the task-specific response to mechanical perturbations, speakers will also adjust their speech in response to auditory perturbations [55]. For example, when the first vowel formant (F1) is artificially shifted upwards, speakers produce within-utterance compensatory responses by lowering their produced F1. The magnitude of these responses only partially compensates for the perturbation, unlike the complete responses produced for mechanical perturbations. While the exact reason for this partial compensation is not known, it has been hypothesized to relate to small feedback gains [19] or conflict with the somatosensory feedback system [14]. We explore the cause of this partial compensation below, but focus here on the ability of the model to replicate the observed behavior.
To test the ability of the FACTS model to reproduce the observed partial responses to auditory feedback perturbations, we simulated production of a steady-state [ǝ] vowel. After a brief stabilization period, we abruptly introduced a +100 Hz perturbation of F1 by adding 100 Hz to the perceived F1 signal in the auditory processing stage. This introduced a discrepancy between the produced F1 (shown in black in Fig 6B) and the perceived F1 (shown in blue in Fig 6B). Upon introduction of the perturbation, the model starts to produce a compensatory lowering of F1, eventually reaching a steady value below the unperturbed production. This compensation, like the response in human speakers, is only partial (roughly 20 Hz or 15% of the total perturbation).
Importantly, FACTS produces compensation for auditory perturbations despite having no auditory targets in the current model. Previously, such compensation has been seen as evidence in favor of the existence of auditory targets for speech [14]. In FACTS, auditory perturbations cause a change in the estimated state of the vocal tract on which the task-level and articulatory-level feedback controllers operate. This causes a change in motor behavior compared to the unperturbed condition, resulting in what appears to be compensation for the auditory perturbation. Our model results thus show that this compensation is possible without explicit auditory goals. Of course, these results do not argue that auditory goals do not exist. Rather, we show that they are not necessary for this particular behavior.
The amount of compensation to an auditory perturbation has been found to vary substantially between individuals [55]. One explanation for the inter-individual variability is that the degree of compensation is related to the acuity of the auditory system. Indeed, some studies have found a relationship between magnitude of the compensatory response to auditory perturbation of vowel formants and auditory acuity for vowel formants [56] or other auditory parameters [57]. This point is not uncontroversial, however, as this relationship is not always present and the potential link between somatosensory acuity and response magnitude has not been established [58]. If we assume that acuity is inversely related to the amount of noise in the sensory system, this explanation fits with the UKF implementation of the state estimation procedure in FACTS, where the weight assigned to the auditory error is ultimately related to the estimate of the noise in the auditory system. In Fig 7B, we show that by varying the amount of noise in the auditory system (along with the internal estimate of that noise), we can drive differences in the amount of compensation the model produces to a +100 Hz perturbation of F1. When we double the auditory noise compared to baseline (top), the compensatory response is reduced. When we halve the auditory noise (bottom), the response increases.
Interestingly, the math underlying the UKF suggests that the magnitude of the response to an auditory error should be tied not only to the acuity of the auditory system, but to the acuity of the somatosensory system as well. This is because the weights assigned by the Kalman filter take the full noise covariance of all sensory systems into account. We verified this prediction empirically by running a second set of simulated responses to the same +100 Hz perturbation of F1, this time maintaining the level of auditory noise constant while varying only the level of somatosensory noise. The results can be seen in Fig 7C and 7D. When the level of somatosensory noise is increased, the response to the auditory perturbation increases. Conversely, when the level of somatosensory noise is reduced, the compensatory response is reduced as well. These results suggest that the compensatory response in human speakers should be related to the acuity of the somatosensory system as well as the auditory system, a hypothesis which we are currently testing experimentally. Broadly, however, these results agree with, and provide a testable hypothesis about the cause of, empirical findings that show a trading relationship across speakers in their response to auditory and somatosensory perturbations [59].
The proposed FACTS model provides a novel way to understand the speech motor system. The model is an implementation of state feedback control that combines high-level control of speech tasks with a non-linear method for estimating the current articulatory state to drive speech motor behavior. We have shown that the model replicates many important characteristics of human speech motor behavior: the model produces stable articulatory behavior, but with some trial-to-trial variability. This variability increases when somatosensory information is unavailable, but is largely unaffected by the loss of auditory feedback.
The model is also able to reproduce task-specific responses to external perturbations. For somatosensory perturbations, when a downward force is applied to the jaw during production of an oral consonant, there is an immediate task-specific compensatory response only in those articulators needed to produce the current task. This is seen in the increased movement of the upper and lower lips to compensate for the jaw perturbation during production of a bilabial /b/ but no alterations in lip movements when the jaw was perturbed during production of a tongue-tip consonant /d/. The ability of the model to respond to perturbations in a task-specific manner replicates a critical aspect of human speech behavior and is due to the inclusion of the task state feedback control law in the model [35].
For auditory perturbations, we showed that FACTS is able to produce compensatory responses to external perturbations of F1, even though there is no explicit auditory goal in the model. Rather, the auditory signal is used to inform the observer about the current state of the vocal tract articulators. We additionally showed that FACTS is able to produce the inter-individual variability in the magnitude of this compensatory response as well as the previously observed relationship between the magnitude of this response and auditory acuity.
We have also shown that FACTS makes some predictions about the speech motor system that go beyond what has been demonstrated experimentally to date. FACTS predicts that a complete loss of sensory feedback would lead to large increases in articulatory variability beyond those seen in the absence of auditory or somatosensory feedback alone. Additionally, FACTS predicts that the magnitude of compensation for auditory perturbations should be related not only to auditory acuity, but to somatosensory acuity as well. These concrete predictions can be experimentally tested to assess the validity of the FACTS model, testing which is ongoing in our labs.
It is important to note here that, while the FACTS model qualitatively replicates the patterns of variability seen in human movements when feedback is selectively blocked, alternative formulations of the model could potentially lead to a different pattern of results. In the current model, somatosensory and auditory feedback are combined to estimate the state of the speech articulators for low-level articulatory control. Given that somatosensory feedback is more directly informative about this state, it is perhaps unsurprising that removing auditory feedback results in smaller changes in production variability than removing somatosensory feedback. However, auditory feedback may be more directly informative about the task-level state, including cases where the task-level goals are articulatory [60] (as in the current version of the model) or, more obviously, where the task-level goals are themselves defined in the auditory dimension [19]. Indeed, a recent model for limb control has suggested that task-level sensory feedback (vision) is incorporated into a task-level state estimator, rather than being directly integrated with somatosensory feedback in the articulatory controller [61]. A similar use of auditory feedback in the task-level state estimator in FACTS, rather than in the articulatory-level estimator in the current version, may produce different patterns of variability when sensory feedback is blocked. We are currently working on developing such an alternative model to address this issue.
The current version of FACTS uses constriction targets as the goals for task-level control. There are a few considerations regarding this modelling choice that warrant some discussion. First, the ultimate goal of speech production in any theory, at an abstract level, must be to communicate a linguistic message through acoustics. Additionally, all speech movements will necessarily have deterministic acoustic consequences. However, this does not imply that auditory goals must be used at the level of control, which is implemented in FACTS based only on constriction targets. Second, the current results should not be taken as arguing against the existence of auditory goals. Indeed, we believe that auditory goals may play an important role in speech production. While we have shown that auditory targets are not necessary for compensation to acoustic perturbations, they may well be necessary to explain other behaviors [14, 62]. Future work can test the ability of FACTS to explain these behaviors. Lastly, while the architecture of FACTS is compatible with auditory goals at the task level, the results of the current model may depend on the choice of task-level targets. Again, future work is planned to explore this issue.
One of the major drawbacks of the current implementation of FACTS is that the model of the plant only requires kinematic control of articulatory positions. While a kinematic approach is relatively widespread in the speech motor control field–including both DIVA and Task Dynamics–there is experimental evidence that the dynamic properties of the articulators, such as gravity and tissue elasticity, need to be accounted for [63–66]. Moreover, speakers will learn to compensate for perturbations of jaw protrusion that are dependent on jaw velocity [59, 67–69], indicating that speakers are able to generate motor commands that anticipate and cancel out the effects of those altered articulatory dynamics. While the FACTS model in its current implementation does not replicate this dynamical control of the speech articulators, the overall architecture of the model is compatible with control of dynamics rather than just kinematics [23]. Control of articulatory dynamics would require a dynamic model of the plant and the implementation of a new articulatory-level feedback control law that would output motor commands as forces, rather than (or potentially in addition to) articulatory accelerations. Coupled with parallel changes to the articulatory state prediction process, this would allow for FACTS to control a dynamical plant without any changes to the overall architecture of the model.
Another limitation of the current FACTS model is that it does not incorporate sensory delays in any way. Sensory delays are non-neglibile in speech (roughly 30-50 ms for somatosensation and 80-120 ms for audition [19]). We are currently exploring methods to incorporate these delays into the model. One potential avenue is to use an extended state representation, where the state (articulatory and/or task) is represented as a matrix where each column represents a time sample [70]. Interestingly, this approach has shown that shorter-latency signals are assigned higher weights in the Kalman gain, even when they are inherently more noisy. This suggests another potential reason for why the speech system may rely more on somatosensation for online control than audition, since its latency is much shorter.
While a detailed discussion of the neural basis of the computations in FACTS is beyond the scope of the current paper, in order to demonstrate the plausibility of FACTS as a neural model of speech motor control, we briefly touch on potential neural substrates that may underlie a state-feedback control architecture in speech [23, 24, 27]. The cerebellum is widely considered to play a critical role as an internal forward model to predict future articulatory and sensory states [26, 71]. The process of state estimation may occur in the parietal cortex [24], and indeed inhibitory stimulation of the inferior parietal cortex with transcranial magnetic stimulation impairs sensorimotor learning in speech [72], consistent with a role in this process. However, state estimation for speech may also (or alternatively) reside in the ventral premotor cortex (vPMC) for speech, where the premotor cortices are well situated for integrating sensory information (received from sensory corteces via the arcuate fasiculus and the superior longitudinal fasiculus) with motor efference copy from primary motor cortex and cerebellum [27]. Another possible role for the vPMC might be in implementing the task state feedback control law [61]. Primary motor cortex (M1), with its descending control of the vocal tract musculature and bidirectional monosynaptic connections to primary sensory cortex, is the likely location of the articulatory feedback control law, converting task-level commands from vPMC to articulatory motor commands. Importantly, the differential contributions of vPMC and M1 observed in the movement control literature is consistent with the hierarchical division of task and articulatory control into two distinct levels as specified in FACTS. Interestingly, recent work using electrocorticography has shown that areas in M1 code activation of task-specific muscle synergies similar to those proposed in Task Dynamics and FACTS [73]. This suggests that articulatory control may rely on muscle synergies or motor primitives, rather than the control of individual articulators or muscles [74].
We have currently implemented the state estimation process in FACTS as an Unscented Kalman Filter. We intend this to be purely a mathematically tractable approximation of the actual neural computational process. Interestingly, recent work suggests that a related approach to nonlinear Bayesian estimation, the Neural Particle Filter, may provide a more neurobiologically plausible basis for the state estimation process [75]. Our future extensions of FACTS will involve exploring implementing this type of filter.
In conclusion, the FACTS model uses a widely accepted domain-general approach to motor control, is able to replicate many important speech behaviors, and makes new predictions that can be experimentally tested. This model pushes forward our knowledge of the human speech motor control system, and we plan to further develop the model to incorporate other aspects of speech motor behavior, such as pitch control and sensorimotor learning, in future work.
We use the following mathematical notation to present the analyses described in this paper. Matrices are represented by bold uppercase letters (e.g., X), while vectors are represented in italics without any bold case (either upper or lower case). We use the notation XT to denote the matrix transpose of X. Concatenations of vectors are represented using bold lowercase letters (e.g., x = [x ẋ]T). Scalar quantities are represented without bold and italics. Derivatives and estimates of vectors are represented with dot and tilde superscripts, respectively (i.e., ẋ and x ˜, respectively).
In FACTS, we represent the state of the vocal tract tasks xt = [xt ẋt]t at time t by a set of constriction task variables xt (given the current gestural implementation of speech tasks, this is a set of constriction degrees such as lip aperture, tongue tip constriction degree, velic aperture, etc. and constriction locations, such as tongue tip constriction location) and their velocities ẋt. Given a gestural score generated using a linguistic gestural model [76, 77], the task state feedback control law (equivalent to the Forward Task Dynamics model in [32]) allows us to generate the dynamical evolution of xt using the following simple second-order critically-damped differential equation:
x ¨ t = M - 1 ( - B x ˙ ˜ t - C ( x ˜ t - x 0 ) ) (1)
where x0 is the active task (or gestural) goal, M, B, and C are respectively the mass matrix, damping coefficient matrix, and stiffness coefficient matrix of the second-order dynamical system model. Essentially, the output of the task feedback controller, ẍt, can be seen as a desired change (or command) in task space. This is passed to the articulatory state feedback control law to generate appropriate motor commands that will move the plant to achieve the desired task-level change.
Although the model does include a dynamical formulation of the evolution of speech tasks (following [32, 35]), this is not intended to model the dynamics of the vocal tract plant itself. Rather, the individual speech tasks are modelled as (abstract) dynamical systems.
The desired task-level state change generated by the task feedback control law, ẍt, is passed to an articulatory feedback control law. In our implementation of this control law, we use Eq 2 (after [32]) to perform an inverse kinematics mapping from the task accelerations ẍt to the model articulator accelerations ät, a process which is also dependent on the current estimate of the articulator positions ãt and velocities a ˙ ˜ t. J(ã) is the Jacobian matrix of the forward kinematics model relating changes in articulatory states to changes in task states, J ˙ ( a ˜ , a ˙ ˜ ) is the result of differentiating the elements of J(ã) with respect to time, and J(ã)* is a weighted Jacobian pseudoinverse of J(ã).
a ¨ t = J ( a ˜ t ) * x ¨ t - J ( a ˜ t ) * J ˙ ( a ˜ t , a ˙ ˜ t ) a ˙ ˜ t (2)
In order to generate articulatory movements in CASY, we use Runge-Kutta integration to combine the previous articulatory state of the plant ([at−1 ȧt−1]T) with the output of the inverse kinematics computation (ät−1, the input to the plant, which we refer to as the motor command). This allows us to compute the model articulator positions and velocities for the next time-step ([at ȧt]T), which effectively “moves” the articulatory vocal tract model. Then, a tube-based synthesis model converts the model articulator and constriction task values into the output acoustics (parameterized by the vector y t a u d). In order to model noise in the neural system, zero-mean Gaussian white noise ε is added to the motor command (ät−1) received by the plant as well as to the somatosensory (y t s o m a t) and auditory (y t a u d) signals passed from the plant to the articulatory state estimator. Currently, noise levels (standard deviation of Gaussian noise) are tuned by hand for each of these signals (see below for details). Together, the CASY model and the acoustic synthesis process constitute the plant. The model vocal tract in the current implementation of the FACTS model is the Haskins Configurable Articulatory Synthesizer (or CASY) [41–43].
The articulatory state estimator generates an estimate of the articulatory state of the plant needed to generate state-dependent motor commands. The final state estimate (ât) generated by the observer is a combination of an articulatory state prediction (ãt) generated from an efference copy of outgoing motor commands, combined with information about the state of the plant derived from the somatosensory and auditory systems (yt). This combination of internal prediction and sensory information is accomplished through the use of an Unscented Kalman Filter (UKF) [36], which extends the linear Kalman Filter [28] used in most non-speech motor control models [23, 25] to nonlinear systems like the speech production system.
First, the state prediction is generated using a forward model (F) that predicts the evolution of the plant based on an estimate of the previous state of the plant (ãt−1) and an efference copy of the previously issued motor command (ät−1). Based on this predicted state, another forward model (H) generates the predicted sensory output y ^ t = [ y ^ t s o m a t y ^ t a u d ] T (comprising somatosensory and auditory signals y ^ t s o m a t and y ^ t a u d, respectively) that would be generated by the plant in the predicted state. Currently, auditory signals are modelled as the first three formant values (F1-F3; 3 dimensions), and somatosensory signals are modelled as the position and velocities of the speech articulators in the CASY model (20 dimensions).
a ^ t= F ( a ˜ t - 1 , a ¨ t - 1 ) (3)
y ^ t= [ H s o m a t ( a ^ t s o m a t ) H a u d ( a ^ t a u d ) ] (4)
These predicted sensory signals are then compared with the incoming signals from the somatosensory (y t s o m a t) and auditory (y t a u d) systems, generating the sensory prediction error (comprising both somatosensory and auditory components) Δ y t = [ Δ y t s o m a t Δ y t a u d ] T:
Δ y t a u d= y ^ t a u d - y t a u d (5)
Δ y t s o m a t= y ^ t s o m a t - y t s o m a t (6)
These sensory prediction errors are used to correct the initial articulatory state prediction, giving a final articulatory state estimate ãt:
a ˜ t = a ^ t + K t Δ y t (7)
where K t is the Kalman Gain, which effectively specifies the weights given to the sensory signals in informing the final state estimate. Details of how we generate F, H, and K are given in the following sections.
Finally, we estimate the vocal tract state estimate at the next time step by passing the articulatory state estimate into a task state estimator, which in our current implementation is a forward kinematics model (see Eq 2) [32]. J(ã), the Jacobian matrix relating changes in articulatory states to changes in task states, is the same as in Eq 2.
x ˜ t= f ( a ˜ t ) (17)
x ˙ ˜ t= J ( a ˜ t ) a ˙ ˜ t (18)
This task state estimate is then passed to the task feedback controller to generate the next task-level command ẍt using Eq 1.
There are a number of tunable parameters in the FACTS model. These include: 1) the noise added to ä in the plant, yaud in the auditory system, and ysomat in the somatosensory system; 2) the internal estimates of the process (ä) and observation y noise; and 3) initial values for the process, observation, and state covariance matrices used in the Unscented Kalman Filter. Internal estimates of the process and observation noise were set to be equal to the true noise levels. Noise levels were selected from a range from 1e-1 to 1e-8, scaled by the norm of each signal (equivalent to a SNR range of 10 to 1e8), to achieve the following goals: 1) stable system behavior in the absence of external perturbations, 2) the ability of the model to react to external auditory and somatosensory perturbations, 3) and a partial compensation for external auditory perturbations in line with observed human behavior. The final noise values used were 1e-4 for the plant/process noise, 1e-2 for the auditory noise, and 1e-6 for the somatosensory noise. The discrepancy in the values for the noise between the two sensory domains is proportional to the difference in magnitude between the two signals (300-3100 Hz for the auditory signal, 0-1.2 mm or mm/s for the articulatory position and velocity signals). Process and observation covariance matrices were initialized as identity matrices scaled by the process and observation noise, respectively. The state covariance matrix was initialized as an identity matrix scaled by 1e-2. A relatively wide range of noise values produced similar behavior: the effects of changing the auditory and somatosensory noise levels are discussed in the results section.
|
10.1371/journal.pgen.1000082 | The Inexorable Spread of a Newly Arisen Neo-Y Chromosome | A newly arisen Y-chromosome can become established in one part of a species range by genetic drift or through the effects of selection on sexually antagonistic alleles. However, it is difficult to explain why it should then spread throughout the species range after this initial episode. As it spreads into new populations, it will actually enter females. It would then be expected to perform poorly since it will have been shaped by the selective regime of the male-only environment from which it came. We address this problem using computer models of hybrid zone dynamics where a neo-XY chromosomal race meets the ancestral karyotype. Our models consider that the neo-Y was established by the fusion of an autosome with the ancestral X-chromosome (thereby creating the Y and the ‘fused X’). Our principal finding is that sexually antagonistic effects of the Y induce indirect selection in favour of the fused X-chromosomes, causing their spread. The Y-chromosome can then spread, protected behind the advancing shield of the fused X distribution. This mode of spread provides a robust explanation of how newly arisen Y-chromosomes can spread. A Y-chromosome would be expected to accumulate mutations that would cause it to be selected against when it is a rare newly arrived migrant. The Y can spread, nevertheless, because of the indirect selection induced by gene flow (which can only be observed in models comprising multiple populations). These results suggest a fundamental re-evaluation of sex-chromosome hybrid zones. The well-understood evolutionary events that initiate the Y-chromosome's degeneration will actually fuel its range expansion.
| Comparisons between related species have shown that, over evolutionary time scales, Y-chromosomes tend to degenerate and can be completely lost. How then can we explain the persistence of Y-chromosomes to the present? One possibility is that losses are counter-balanced by the origin of new Y chromosomes, which then spread throughout the species in which they have arisen. The first of these two processes, the generation of new Y chromsomes, is more readily understood: it can occur if an autosome (a non sex chromosome) fuses with an X chromosome. This form might become established in one locality. However, its subsequent geographic spread has been more challenging to explain. Problems arise if gene flow carries them to another part of the species range. Crosses can then occur which introduce the new Y chromosome into females, who are expected to suffer reduced fitness. The new sex chromosomes are therefore selected against when they are in the minority. We use simulations to show that they can nevertheless spread, if they meet the ancestral forms at a front so the chromosomes intermingle in a hybrid zone. Paradoxically, the degeneration of the Y will actually intensify selection, thereby speeding its spread.
| Our understanding of sex chromosome evolution has increased immensely in the past decade. Theoretical expectations [1],[2] have been experimentally verified in a wide variety of organisms, including fish, fruitflies, mammals and plants [3]–[8]. For example, it is widely documented that over a series of generations, a Y-chromosome will eventually stop recombining with the X over most of its length. As a consequence there are increased rates of transposition, degeneration, heterochromatinization and loss of function of genes on the Y, amongst other changes [5], [8]–[13].
It appears then, that the inexorable fate of Y-chromosomes is degeneration and perhaps loss. It is even possible that all sexually dimorphic species lacking a Y have previously passed through a Y-possessing stage [1],[14], as is the case for Caernohabditis elegans [15] (the logic would also apply to equivalent W chromosomes in species with heterogametic females). The persistence of Y-chromosomes to the present day therefore suggests that they can repeatedly arise de novo. One straightforward way in which new Ys can be created is by the fusion between an autosome and X-chromosome followed by its fixation. This paper models the evolution of such neo-XY sex chromosome systems and, in particular, asks why they should become established throughout a species' range. The analysis suggests that the spread of neo-Ys is much more likely than suggested by current models, and that this new proposal could be tested by analysis of sex chromosome hybrid zones.
A concrete example can be a useful guide for explaining and constructing evolutionary models, so we make use of the well-studied example of the neo-XY race of the grasshopper Podisma pedestris. Phylogenetic comparison [16] suggests that the ancestral P. pedestris karyotype had females with two X-chromosomes and males with one (but no Y): this is known as an XX/XO sex-determining system [17]. The system changed following the centric fusion of the X with an autosome (Au) to create a larger metacentric neo-X. The fused karyotype has become fixed in populations in the southern part of the species' distribution in the French Alps. In these fixed populations, the females contain two neo-X chromosomes, and hence no unfused Au. In males, however, the unfused Au chromosomes have continued to pair with the homologous section of the neo-X. These unfused Au are now restricted to males and are consequently designated neo-Y chromosomes. The karyotypes of the original unfused, and derived fused race are illustrated in Figure 1 (karyotypes A, B, F, G).
Sex chromosomes are often involved in fusions. Indeed, human sex chromosomes are believed to be the products of at least three chromosomal fusions [6],[18],[19], as is the Drosophila Y-chromosome [20]. The occurrence of a fusion is insufficient to explain the genesis of a neo-XY system however. Following the fusion event, the new karyotypes must also become fixed throughout the species range (or part of it). The establishment of neo-XY systems does appear to occur repeatedly in evolution. Good evidence comes from the Orthoptera, which have conveniently large chromosomes for surveys of karyotype. White calculates that there have probably been six independent fixations of the XY system from an ancestral XO condition in the Australian subfamily Morabinae alone (based on karyotypes from about 80 species). More generally, the fixation of the XY system has been reported in at least 21 genera of Acrididae [16](and references therein).
Attempts to explain this establishment fall broadly into two traditions. Firstly, cytogeneticists have noticed that chromosomal rearrangements often confer reduced fertility in heterozygotes. The cause may be a direct effect of meiotic aberrations [21],[22]; in other cases selection against changed recombination patterns is suspected (see [23] for an example). These forms of selection actually act against the fusion when it is rare; but it could nevertheless become established in small isolated populations if genetic drift elevated its frequency until it became the commoner type and hence favoured by selection [24]. The difficulty with this explanation is to account for how the fusion would subsequently spread from a single isolated population to other populations. Lande [24] proposed that the fused race could colonize sites left vacant by local extinctions, whereas Hewitt [25] argues that spreading would be more effective if the initial population was located on the expanding margin of the species range, as it spreads into new territory–most likely during an episode of rapid climatic change.
A second perspective comes from consideration of the alleles that were segregating on autosomes before the fusion occurred. Fusion with a sex chromosome might bring alleles into linkage with the newly created sex chromosome and confer a fitness benefit; the key alleles might be sexually antagonistic (benefiting one sex at the expense of the other) [26] or deleterious recessives [27] (especially in strongly inbreeding populations). In both cases the selection is expected to be much more effective in promoting Y-autosome fusions, and such events might repeatedly add new genetic material to existing Y-chromosomes. Linkage with sexually antagonistic alleles could also produce selection for the fixation of new X-autosome fusions, and hence the creation of neo-Ys.
There is some evidence that sexually antagonistic alleles may indeed have appreciable effects. Rice conducted an imaginative breeding design in which a haploid Drosophila genome was restricted to one sex for several generations and then returned to the other sex [28]–[30]. The results were striking. In less than 30 generations, sex specific fitness differences had become established in the sex-restricted genome. The rapidity of the response was interpreted as showing that sexually antagonistic alleles had been segregating in the founder population.
Even with strong selection on sexually antagonistic alleles, the advantage provided to the fused chromosome would be weak [26]. Nonetheless, this selective process, or the action of drift, might establish the neo-XY system locally in part of the species distribution. The spread throughout the whole species range is more difficult to explain. Any advantage to the fusion when rare is expected to be transient, because of the well-understood evolutionary events affecting new sex chromosomes. Alleles reducing female fitness, can accumulate readily on the Y [1],[2], particularly if they also had beneficial effects in males.
Our analysis, has uncovered a paradoxical effect that nevertheless favours the geographic spread of the neo-XY system. If sexually antagonistic alleles have become established on the Y, the genetic interactions at the boundary between neo-XY and ancestral populations can favour the spread of the neo-X. Surprisingly, the results hold even if the net effect of selection against the neo-Y in females outweighs the benefits in males.
The pattern of chromosome segregation in crosses involving individuals with different sex chromosome combinations is illustrated in Figure 2. The letters correspond to the karyotypes shown in Figure 1, the area of each cell is proportional to the number of each karyotype in the offspring. This scheme was translated into a set of equations for the frequency of each karyotype as a function of the frequencies in the previous generation, assuming random mating and after weighting each karyotype by its fitness. A program to iterate the equations was written in the statistical language R [31], and is listed in the supporting information (Dataset S1). The initial analysis revisited, and then extended the results of [26]. Consider two sexually antagonistic alleles that might be segregating on the autosome Au during the period before the chromosomal fusion. The two alleles (a and b) have different fitness in the two sexes (specified by w♀aa, w♀ab & w♀bb for females, and w♂aa, w♂ab & w♂bb for males). The a allele was assumed to be favored in males, and the b in females so w♀bb = w♂aa = 1. The program calculated the outcome of selection for all possible combinations of w♂bb and w♀aa in the range 0–1 at intervals of 1/40 (with a specified dominance). If there is polymorphism at this locus before the fusion takes place, the fusion will (in some cases) generate linkage disequilibrium leading to selection for the fixation of the fusion. The fitness combinations leading to such polymorphism can be illustrated by initiating a simulation with only the ancestral XX∶XO karyotypes, and including a chromosome carrying the a allele (equivalent to the green chromosome in Figure 1) at low frequency. Figure 3A illustrates a range of fitness combinations producing polymorphism (achieved from an initial a frequency of 0.1% with all genotypes in Hardy-Weinberg proportions).
The combinations of fitness that can lead to selection for the fixation of the fusion have been explored in some detail [26]. In Figure 3, they correspond to the b allele becoming linked to the X by the fusion. This selection for the fusion could be demonstrated in our simulations by initiating the fusion at a low frequency (0.4%, corresponding to the frequency of the fusion, if migration had introduced the Y at 0.1%) and then iterating the equations for 1000 generations. The analysis was extended to investigate the effect of the slightly reduced fertility that is expected in females heterozygous for the fusion. This additional selection was set at s = 0.01, the value estimated for P. pedestris. The recombination rate between the b allele and the X centromere was set at zero to maximize the selection for the fusion [26]. As the Y-chromosome evolves, recombination is expected to be reduced over a greater proportion of the Y [1], hence the modelled effects are increasingly likely to occur. Indeed, in the case of P. pedestris, which is assumed to have a young neo-XY system, the recombination is already displaced way from the Y centromere [32], perhaps simply as effect of the fusion itself, and very strong linkage disequilibrium has been found even in the middle of the zone for an X-marker [33].
These initial calculations involved a single panmictic population. The outcome can be different when the population is subdivided. The next step was therefore to consider a situation in which the neo-XY system had become established in an isolated area, and come into contact with the ancestral (XX:XO) karyotype. Gene flow between the two chromosomal races would then produce a hybrid zone. A computer simulation of a linear array of 40 populations was used to model this situation. Initially the left hand 20 populations were fixed for the ancestral karyotype and the remainder for the neo-XY. There was gene flow of 8% between adjacent populations (total gene flow of 16%). Population size was uniform across the simulated populations. In other words there was no density trap to pin the zone down to a particular location (as described by [34]). The two ends of the array could either be set to receive gene flow from populations fixed for the ancestral karyotypes, or to only receive gene flow from their more central neighbour. Both options were used to check for any effect on the outcome of the simulations. For each generation, after gene flow, the expected frequencies of genotypes in the next generation were calculated as before.
In this extension of the model, the green chromosome in Figure 1 is considered to be a neo-Y-chromosome. The a allele would have been fixed on the neo-Y which could also have accumulated additional sexually antagonistic alleles tightly linked to the X centromere (due to the extension of the non-recombining region). We have argued that we would expect some alleles to be selected against in males and favoured in females, and for there to be selection against the chromosomal heterozygotes. However, it is helpful to understand the combination of these effects by first examining the behaviour of these three forms of selection individually. We therefore summarize the results by defining three selection regimes. The first two correspond to points on the X and Y axes of Figure 3: Firstly, male-beneficial variants could have become established on the Y (w♂(AU·)<1), were ‘Au·’ represents genotypes containing Au– the autosomal homologue of the neo-Y; secondly, female-deleterious variants could occur on the Y (w♀(Y·)<1). The third simple case is selection against females heterozygous for the fused X (w♀(FU)<1). We then simulated all possible combinations of the fitness regimes. Table 1 sets out the karyotypes with reduced fitness in each regime.
Having determined the basic patterns produced by the different fitness regimes, we assessed the spread of the neo-XY system throughout the possible parameter range shown in Figure 3. The simulations were run until fixation or until 10 000 generations. We explored the full range of values for sf and sm in the presence of selection against chromosomal heterozygotes, which was set to that estimated in P. pedestris of sh = 0.01 (fitnesses specified in the first and second rows of Table 1, combined multiplicatively). The dominance of the sexually antagonistic selection (dm and df in Table 1) was varied: from dm = 1 or 0.5 for the male effect (not zero since recessive male-beneficial alleles would have had no advantage on the Y), and df = 1, 0.5 or 0 for the female effect. In addition to these forms of selection, we also considered the possibility that there had also been evolution of coadaptation or dosage compensation between the sex chromosomes in the established neo-XY and XO populations, giving rise to the fitnesses in the last two rows of Table 1.
The model zone width was converted to values that could be observed in the field using the relationship σ2 = mD2, were σ2 is the variance in parent-offspring dispersal and is a measure of migration. For P. pedestris, it has been estimated by mark-release-recapture experiments to be 400 m2 per generation [35]. The value D represents the distance in the field equivalent to that between adjacent simulated populations. Since the simulated migration rate, m, was 0.16, the real hybrid zone width of 800 m [34] is equivalent to the distance between 16 simulated populations.
When all fitnesses were set to one, the simulated width of the zone increased with time–matching the neutral expectation w = 2.51σ√t, where w is the width, σ is the parent-offspring dispersal per generation and t is time in generations [36] (results not shown). Similarly, in the case w♀(FU)<1, the simulated width of the fusion cline fitted analytical expectations, as long as the female-specific nature of selection was taken into account (see Results). The program to simulate the structured populations was written in Java and the full source is available from the authors upon request.
Figure 3B shows the neo-XY chromosomes can invade an ancestral XO population when they are introduced at low frequency, if there is strong sexually antagonistic selection (the central area of the Figure). The spread of the fused X (the neo-X) was accompanied by fixation of the a allele. The model included weak selection against females heterozygous for the fusion, hence the fusion was selected against when rare. Consequently there were combinations of low to moderate selection (i.e. around (1,1)) for which the fusion did not spread. This failure to spread occurred irrespective of the dominance of the a allele in males, or the b allele in females (results not shown).
Figure 4 summarizes the various outcomes of the simulated meeting between the two chromosomal races to form a hybrid zone. The relative frequency of the fusion has been plotted as against distance along the array of populations. This relative frequency was calculated as ff/(ff + fu), where fi specifies the frequency of the chromosome of type i∈{f, u, Y, Au}, representing fused X, unfused X, Y and autosome respectively. Similarly, the frequency of the Y was calculated as fY/(fY + fAu). Note that the denominator increases with fu, since unfused individuals carry more of these chromosome (i.e. more Au and/or Ys, see Figure 1). Once the two races meet, gene flow produces a sigmoidal transition in the frequency of the fused X and the Y-chromosomes. For some parameter values the Y cline or the fusion cline spread as a wave of advance, indicated by arrows in Figure 2, in other cases the clines were stable or decayed (see Discussion).
Existing analytical models do not describe much of this behaviour, but there are well known results for the case of heterozygote disadvantage w♀(FU)<1 under which the fusion cline would assume a fixed position and width. The relationship between the strength of selection (in the range 0.05–0.9) and width (the inverse of the maximum slope [37]) closely fitted the expected relationship w = √8σ/√s [38] as long as the selection coefficient, s, was multiplied by 2/3 to compensate for selection acting on females only (r2>99.99%, regression coefficient = 1.016).
The examples in Figure 4, in which the neo-XY system spreads (upper two panels), involve strong sexually antagonistic selection–chosen to clearly illustrate the qualitative difference in outcome from single population models (in which it did not for these values). The neo-XY system spreads even when there is strong selection against the Y in females.
We investigated the rate of spread of the neo-XY system under less severe selection (Table 2). Selection against the X chromosome heterozygotes of 1% (sh = 0.01) slowed down the rate of spread slightly (up to 50%), but did not prevent it even when the sexually antagonistic selection was weak (e.g. sm = sf = 0.005). Other forms of selection against introgression could both accelerate or retard the rate of spread. We investigated the fitnesses that might be generated by coadaptation and dosage compensation of the sex chromosomes (Table 1). Both forms of selection accelerated the spread when the selection for the Y in males was greater than the disadvantage in females (sm>sf) and retarded or even slightly reversed the direction of spread under the converse (sm<sf) (Table 2). The effect of dominance was minor over most fitness combinations and the outcomes were qualitatively unaffected. We show the effects of male dominance in Table 2, but omit the female for brevity.
In the Introduction we outlined how the fixation of an X-autosome fusion could be explained by selection in favour of sexually antagonistic alleles linked to the fused centromere. Charlesworth and Charlesworth [26] have shown that there is net selection in favor of the fusion for fitness combinations that lead to polymorphism at the sexually antagonistic locus: which fall in the central shown in Figure 3A. However, Figure 3B suggests that this form of selection might be readily counteracted, even by a very minor (1%) reduction in the fertility of female fusion heterozygotes, w♀(FU). In particular, even relatively strong sexually antagonistic selection is overwhelmed: notice that when this additional selection is applied (Figure 3B) the fusion does not spread for fitness values within 0.15 of the point (1,1) (i.e. selection coefficients of up to 15%) even if they fall within the polymorphic area in Figure 3A. Selection against female fusion heterozygotes is considered likely because of non-disjunction at meiosis [21], and is indeed suspected to occur in P. pedestris [37]. In these circumstances, it is easier to envisage the fusion becoming established by genetic drift, than deterministically under the action of selection.
Whatever the reason for the fusion initially becoming established in one locality, once it is fixed, the Au autosome will then be restricted to males, and would consequently have become a neo-Y. The subsequent evolution of the sex chromosomes would therefore take a course that would at first sight seem to make the spread of the neo-XY system even more unlikely. In particular the neo-Y is expected to accumulate further sexually antagonistic effects, which would in turn select for the loss of recombination, and its eventual degeneration [1]. The selection acting would therefore be stronger (further away from the point (1,1) in Figure 3B) and for the most part this would lead to even stronger selection to eliminate the neo-XY system: only if selection is strong and of similar order in males and females (the triangular area) would selection favour its spread.
The model illustrated in Figure 3B assumes a small starting number of neo-XY individuals. Remarkably, the outcome is completely overturned if the XO and neo-XY populations are assumed to meet in a hybrid zone. The spread of the neo-XY system would actually be driven by the selection regimes that lead to its elimination in Figure 3B.
It may be simplest to start interpreting the results using the biologically unrealistic case of selection only against females containing Y-chromosomes, w♀(Y·)<1. Since only females with an unfused X chromosome can contain a Y, this regime leads to selection against them, causing the fusion cline to advance (Figure 4, w♀(Y·)<1). However, in the absence of other selection, this effect is transient since the direct selection on the Y removes it from populations containing unfused chromosomes. In other words, the autosome (Au) advances because it is favoured by selection. The Y persists only in the heartlands of the fused chromosome range, because there it experiences no disadvantage because it cannot enter females.
There is a comparable indirect effect on the fusion in the case of selection only in favour of Y-chromosomes in males, w♂(Au·)<1. In populations that are polymorphic (for Y/Au), unfused males are more likely to contain at least one advantageous Y because they have double the number of these chromosomes (in fused males, the X replaces one of them). Hence w♂(Au·)<1 results in selection against the fusion (Figure 4, w♂(Au·)).
We can extend these explanations to the most interesting and biologically relevant result–the wave of advance for both the Y and the fusion clines under sexually antagonistic selection (w♂(Au·)<1 & w♀(Y·)<1). When the selection against the Y in females is stronger, the Y-chromosome tends to be removed from the fusion cline as under w♀(Y·)<1. However, as the fusion advances (for the same reason as under w♀(Y·)<1) the Y follows behind, up to the margins of the fusion cline, thereby indefinitely maintaining the selection for the advance of the fusion. Note that this neo-XY success depends on the gene flow continually bringing the Y-chromosomes into the zone, which is why it did not occur in the single partially isolated population of Figure 3B.
The neo-XY system also spread when there is a net advantage to the Y (under sexually antagonistic selection). However, in this case, the Y spreads as a traveling wave ahead of the fusion cline. The fused X will also spread because, as the Y becomes common, the selection against unfused females gets stronger whereas the benefit to unfused males is reduced (since both fused and unfused males tend to carry the favourable Ys once they become common).
The observed speed of spread of the fusion was relatively small compared to dispersal: the fastest being equivalent to 450 generations to move 1 km in P. pedestris, or 1/10th of the dispersal distance per generation. It would be difficult to observe by repeated sampling, but would cause consistent movement over evolutionary time. In addition, the sexually antagonistic effects could increase in magnitude with time, because newly evolved sexually antagonistic alleles arising in within the neo-XY range would be expected to spread until they met the hybrid zone.
We consider the forms of dominance in Table 2 to be most likely, although we have explored other combinations and found no qualitatively different outcomes. Dominant male beneficial effects would fix faster initially, making them more likely to evolve, however such alleles could spread in XO or neo-XY populations, so they need not be restricted to the neo-XY race; hence there could be a range of dominance for male beneficial effects. The first female deleterious effects on the Y would probably be due to loss of function mutations, and would therefore be less than fully dominant because the functional ancestral allele would still be present.
In addition to heterozygote disadvantage we also considered alternative forms of selection that might act against recombinant karyotypes in the hybrid zone. As long as these additional effects are of the same order as the sexually antagonistic selection, the results are essentially unchanged. It is, of course, possible to find some fitness combinations that slow down, or even slight reverse the spread, and it could be counter-acted by assymetrical gene flow (e.g. due to a density gradient [34]) or other selection (e.g. due to change in the environment). However, Rice's experiments [28]–[30] suggest that the alleles with large sexually antagonistic effects are segregating in natural populations, so we would expect their effects to predominate in the zone as soon as it was formed. If this interpretation is correct, more detailed analysis of hybrid zones should provide additional evidence of this sexually antagonistic selection (see below).
Interestingly, strong selection on the Y chromosome resulted in a broad fusion cline (Figure 4). The result emphasizes that the width of the cline in the character for which the hybrid zone was originally discovered, need not indicate the strength of selection. In fact the term “hybrid zone” can be misleading in these cases; it is preferable to refer to different clines. The broad fusion cline in the presence of strong selection is particularly relevant to the P. pedestris hybrid zone, where strong selection is detected in the F1 in lab crosses [37],[39] and in the field [40], yet the fusion cline is much wider than expected from the observed selection [41]. Previously, this discrepancy has been explained by a model in which selection is spread over many loci [41] only some of which need be linked to the fusion, but our results offer an alternative possibility: that the action of the selection is indirect and due to the well understood initial events in sex chromosome evolution.
One implication of the results is that sex chromosome hybrid zones are a valuable, yet unexploited, source of information on early sex chromosome evolution. We suggest that it will be rewarding to obtain markers that distinguish the Y-chromosome from its homologous autosome (Au in our notation) and to survey their geographic distribution across known sex chromosome hybrid zones. Often the clines of different characters coincide (have the same centre) [42],[43], however we would expect them to be displaced in the case of sexually antagonistic selection. Comparing the Y and the fusion cline as in Figure 4 using real hybrid zone data is a robust way to identify the selection regime operating.
The conventional explanation for a narrow sex chromosome hybrid zone is that there is selection against the chromosomal heterozygotes [34],[37]. In that case there would be a narrow transition for the chromosomal fusion, but the distribution of the Y would be very similar to the neutral case (compare Figure 4, w♀(FU)<1 with the neutral case). However, if sexually antagonistic selection is operating, then these two clines will be displaced and the position on the Y cline relative to the fusion cline will indicate the relative strength of male beneficial and female deleterious effects on the Y. For example faster male evolution [44] would be supported if the Y cline were ahead of the fusion. This novel information on the forms of selection affecting young Y-chromosomes in natural populations has not previously been tapped.
A second indication of sexually antagonistic selection would be cline movement. In some cases it has proved possible to detect the actual movement of hybrid zones by repeated surveys or reviewing museum collections e.g. [45],[46]. In other cases the movement would be too slow, or held back by barriers to gene flow or gradients in population density [47]. It should still prove possible to identify slow or historical movement by surveys of other loci throughout the nuclear and cytoplasmic genome (for a review see [48]).
The realisation that surveys of sex chromosome hybrid zones can answer questions relating to the early evolution of sex chromosomes is exciting because such hybrid zones are already known and waiting to be analyzed. Examples include Drosophila americana [49], the morabine grasshopper Vandiemenella (Warramaba) viatica [25] and the grasshopper Podisma pedestris [34]. A great advantage of hybrid zone studies is that they involve wild populations [25] and some may represent snapshots of the actual establishment of a neo-XY system in nature, thus allowing the testing of theoretical predictions in biologically realistic conditions. |
10.1371/journal.pgen.1000652 | Fast Growth Increases the Selective Advantage of a Mutation Arising Recurrently during Evolution under Metal Limitation | Understanding the evolution of biological systems requires untangling the molecular mechanisms that connect genetic and environmental variations to their physiological consequences. Metal limitation across many environments, ranging from pathogens in the human body to phytoplankton in the oceans, imposes strong selection for improved metal acquisition systems. In this study, we uncovered the genetic and physiological basis of adaptation to metal limitation using experimental populations of Methylobacterium extorquens AM1 evolved in metal-deficient growth media. We identified a transposition mutation arising recurrently in 30 of 32 independent populations that utilized methanol as a carbon source, but not in any of the 8 that utilized only succinate. These parallel insertion events increased expression of a novel transporter system that enhanced cobalt uptake. Such ability ensured the production of vitamin B12, a cobalt-containing cofactor, to sustain two vitamin B12–dependent enzymatic reactions essential to methanol, but not succinate, metabolism. Interestingly, this mutation provided higher selective advantages under genetic backgrounds or incubation temperatures that permit faster growth, indicating growth-rate–dependent epistatic and genotype-by-environment interactions. Our results link beneficial mutations emerging in a metal-limiting environment to their physiological basis in carbon metabolism, suggest that certain molecular features may promote the emergence of parallel mutations, and indicate that the selective advantages of some mutations depend generically upon changes in growth rate that can stem from either genetic or environmental influences.
| Effects of mutations can change under different genetic backgrounds or environmental factors, also known as epistasis and genotype-by-environment interactions (G×E), respectively. Though epistasis and G×E are traditionally treated as distinct phenomena, our study of a beneficial mutation highlights their commonality. This mutation resulted from insertion of the same transposable element upstream of a novel cobalt transport system in 30 of 32 independent populations during evolution in metal-limited media. The resulting increased cobalt uptake provided a selective benefit that depended upon two environmental factors: cobalt limitation and growth substrates whose metabolism requires a particular vitamin B12 (which contains cobalt) -dependent biochemical pathway. Furthermore, this mutation exhibited epistatic and G×E interactions with other cellular processes in a generic way, such that its selective advantage increased as cells were able to grow faster. This growth-rate dependence accords with a simple model: the slowest of multiple physiological processes needed for growth exerts the greatest control over an organism's growth rate. It suggests that as growth results from the performance of the entire physiological system, genes or environmental factors that affect distinct physiological processes may thus interact through their convergent effects on growth phenotypes.
| Adaptation is a product of genetic modification and natural selection imposed by environmental challenges. A complete understanding of adaptation of biological systems thus requires identification of how selection acts upon organismal traits and mapping adaptive phenotypes to underlying genotypic changes. Experimentally testing the genotype-phenotype association and phenotypic effects of mutations is an ongoing research direction in many fields of biology [1]–[3]. Studies on mutations have shown that genetic interactions (epistasis) are common in biological systems [4]–[7] and fitness effects of beneficial mutations can vary greatly depending on environmental conditions (genotype-by-environment interactions, G×E) [8]–[10]. Many studies of beneficial mutations, however, stop short of elucidating the exact molecular mechanisms connecting genotypic changes to phenotypic adaptation [11]–[13]. The lack of this level of information has rendered prediction of fitness effects, epistasis, and G×E interactions elusive. On the other hand, much of our current knowledge of biological systems has come from studying phenotypes of deleterious gene knockouts. Such approaches have uncovered many gene functions and genetic interactions but provided little information about the quantitative response of biological networks to environmental or genetic perturbations as well as the functional significance of a gene in the context of adaptation. A complementary approach to studying the function and evolution of biological systems, therefore, is to characterize molecular mechanisms through which beneficial mutations alter physiology, and reciprocally, how physiological differences due to genetic backgrounds or environments influence the effects of beneficial mutations.
In recent years, evolution experiments using microorganisms have offered a powerful means to investigate the genetic basis of adaptation [14]. Evolution of experimental populations is often conducted using resource-limiting conditions, a challenge many organisms encounter in nature. One competitive strategy to survive under such a scenario is to enhance resource uptake through transport systems. If physiological acclimation is insufficient to alleviate resource limitation, natural selection can favor mutations that further increase uptake capacity. Phenotypes competent to import resources at low concentrations emerge frequently in microbial populations subjected to evolution under resource limitation [10],[15],[16]. Interestingly, beneficial mutations emerging from evolution experiments often occur repeatedly at particular loci [17]. Phylogenetic and quantitative genetic studies of natural populations have also identified many cases of parallel genetic evolution in both micro- and macroorganisms. Frequent observation of genetic parallelism underlying adaptation suggests that, in addition to environmental factors that confine the direction of phenotypic evolution, certain features of the genetic architecture, such as DNA sequence space, genome structure, and the organization of physiological networks may further constrain the breadth of evolutionary trajectories [18].
Metals are essential but often growth-limiting in nature. They are involved in a wide range of physiological processes, such as stabilizing protein structure, relaying cellular signals, and facilitating catalysis in nearly one-third of enzymes [19],[20]. Their most biologically active forms, free cations, however, are limiting in many ecosystems due to oxidation or complexation with organic or inorganic matter [21],[22]. Metal deficiency has been shown to limit the bioproductivity in marine ecosystems and tropical agricultural systems worldwide [23],[24]. For host-pathogen arms races, both animals and plants secrete ligands to sequester metal cations in body fluids to suppress pathogen proliferation, while pathogens have evolved counter-strategies to snatch metals from these ligand-metal complexes [25]–[27]. Clearly, sophisticated metal transport systems and metal-dependent gene regulation mechanisms represent biological adaptation to maintaining metal homeostasis [28],[29] and emphasize the importance of metal acquisition as a prominent fitness component under metal limitation.
In this study, we examined the genetic and physiological basis of adaptation to metal limitation in experimental populations of Methylobacterium extorquens AM1 (hereafter Methylobacterium) grown in media that we present here as being metal-deficient. In addition to multi-carbon (multi-C) substrates like succinate, Methylobacterium can grow on single-carbon (C1) compounds like methanol and serves as a model to dissect and engineer metabolic systems of C1-utilizing bacteria [30]. Growth of Methylobacterium on methanol and on succinate, however, involves several distinct biochemical pathways, and dramatic differences in global gene expression and metabolic profiles have been observed between these two growth modes [31]–[34]. Experimental evolution of populations of Methylobacterium grown for 1500 generations on methanol or succinate revealed tradeoffs during adaptation to these two substrates [35]. These tradeoffs were found to be both asymmetric and variable: methanol-evolved populations consistently showed improvements on both substrates, whereas approximately half of the succinate-evolved populations completely lost the ability to grow on all C1 compounds. Unexpectedly, in these experiments examining tradeoffs, as well as those applying experimental evolution to select for improved growth of Methylobacterium bearing an engineered metabolic pathway, we later discovered that the growth media used for evolution was metal deficient due to over-chelation by ethylenediaminetetraacetic acid (EDTA) present in the media. Long-term evolution under such conditions led to the emergence of mutants with enhanced metal uptake in experimental populations founded by either genotype. Here we show that adaptation of Methylobacterium to metal limitation entailed remarkably parallel transpositions of an insertion sequence (IS) element that increased expression of a novel cobalt transporter system. The selective advantage of improving cobalt uptake was specific to methanol growth under metal limitation and seemed to result from sustaining biosynthesis of vitamin B12, a cobalt-containing cofactor, to support two vitamin B12-dependent reactions in C1 metabolism. Intriguingly, this mutation provided a higher selective advantage in genetic backgrounds or growth conditions that conferred faster growth rates, indicating growth-rate dependent epistatic and G×E interactions. This generic growth-rate dependence suggests that as growth results from the performance of the entire physiological system, genes or environmental factors that affect distinct physiological functions may thus interact through their convergent effects on growth phenotypes.
The IS transposition that occurred across multiple experimental populations was first identified in an evolved isolate, CM1145, from one of the eight methanol-evolving populations (termed F1 to F8) founded by an engineered Methylobacterium strain (hereafter termed the EM strain) (Table S1). In the EM strain, the endogenous formaldehyde oxidation pathway required for growth on C1 compounds was replaced with a phylogenetically-unrelated formaldehyde oxidation pathway from Paracoccus denitrificans [36] (see Plasmid and Strain Construction in the Materials and Methods section). To identify physiological changes that occurred during adaptation of F populations founded with the EM strain, we performed a preliminary microarray analysis to compare genome-wide mRNA pools between the EM strain and the evolved isolate CM1145 (GEO accession no. GSE14875). Further analysis of changes in the transcriptional profile during adaptation in this, and other replicate populations is underway (Chou and Marx, unpublished). Among the observed transcriptional changes from our initial experiment, a putative metal transport cassette increased expression by 50-fold in strain CM1145, relative to the EM strain. Real-time PCR analysis of the two uncharacterized genes in this cassette, icuA and icuB (improved cobalt uptake phenotype, GenBank accession no. EU679505), revealed 70.8±13.0-fold and 20.0±4.7-fold increased transcription, respectively (throughout we report the mean and 95% confidence intervals based on three replicates). Open reading frames (ORFs) of icuA and icuB overlap by 4 bp. The icuA gene encodes a 704-amino acid protein homologous to TonB-dependent outer membrane receptors. The icuB gene encodes a protein of 243 amino acids exhibiting no significant sequence similarity to any characterized gene in public databases. The CD-Search program [37] clustered IcuB with a group of uncharacterized ORFs (CDD accession no. COG5266) predicted to encode periplasmic components of the ABC-type cobalt transport system. PCR amplification of the icuAB locus of strain CM1145 detected a 1.6 kb size increase within its 5′ upstream region. Sequencing of the PCR product revealed transposition of an insertion sequence, ISMex4 (GenBank accession no. EU679504), into a site 113 bp upstream of the icuA start codon (icuAB1145 allele with a ‘Type I’ insertion, thus here icuABT1) (Figure 1A).
Previous studies have shown that transpositions of IS elements may activate transcription of downstream genes by introducing IS-associated outward-directed promoters or by creating hybrid promoters at the junction of insertion [38]. To investigate how ISMex4 insertions enhance transcription of the downstream icuAB genes, we measured the promoter activity of the 5′ upstream region of the WT icuAB allele (icuABWT) and fragments covering various parts of the icuABT1 5′ upstream regions with a promoter-probe plasmid using transcriptional fusions to GFPuv (Figure 2). The promoter activity of either a 113-bp or a 968-bp 5′ upstream region of the icuABWT allele were below the detection limit during growth on methanol. By contrast, ISMex4 alone exhibited significant promoter activity, and the highest activity was observed in the full-length icuABT1 5′ upstream region (ISMex4 plus the adjacent 113-bp 5′ upstream region). Interestingly, a 282-bp fragment spanning the icuABT1 insertion junction did not exhibit detectable promoter activity. These results suggested that insertion of ISMex4 raised transcription of icuAB genes through its outward promoter activity and a synergistic effect between ISMex4 and the adjoining 5′ upstream region, rather than through formation of a hybrid promoter at the insertion junction.
We used the aforementioned PCR-based screen to survey the icuAB locus of evolved isolates across all 8 F populations grown in methanol, as well as the 8 replicate populations each from 4 different evolution experiments founded by the WT strain. These populations, (Table 1, termed A, B, C & D) were grown for 1500 generations on methanol, succinate, both, or alternating between them, respectively [35]. Insertions of ISMex4 into the icuAB 5′ upstream region occurred in evolved isolates from 30 out of the 32 A, C, D, and F populations, all of which were evolved solely or partially on methanol. On the contrary, none of isolates from the 8 B populations evolved solely in succinate acquired such mutation. PCR amplification using the 8 B population samples did not detect ISMex4 insertion into the icuAB locus among these populations. The pattern of ISMex4 insertions present among A, C, D, F populations versus that of B populations is significantly different (Fisher's exact test, P<10−6). Sequencing the icuAB 5′ region revealed that isolates from 26 populations had an identical ISMex4 insertion as icuABT1. In addition, a second type of ISMex4 insertion was found 12 bp upstream of the icuA start codon in strain CM1059 from population C3 (icuAB1059 allele with a ‘Type II’ insertion, or icuABT2) and subsequently in four other populations (Figure 1A). This extreme parallelism cannot be accounted for by the presence of these mutations at low frequencies in the ancestral stocks because two types of ancestral genotypes were used in these experiments. In addition, each population was inoculated from a single colony of its respective ancestor. The icuABT2 allele increased transcription of icuA and icuB by 5.9±0.3 and 6.1±1.4 fold, respectively. For both icuABT1 and icuABT2, the transposase gene of ISMex4 was in inverse orientation to the icuAB genes. Sequencing of the icuAB 5′ upstream and coding regions of evolved isolates from B populations and the two F populations free of ISMex4 insertion did not identify mutations of any type. ISMex4 has 8 identical copies in the Methylobacterium genome [39]. Analysis of these 8 insertion sites along with new insertions identified in this study deduced a 4-bp consensus target sequence (5′-BTAR-3′) that duplicates upon transposition of ISMex4 (Figure 1B) [40]. Analysis by the Mfold program suggested that ISMex4 insertion sites tend to locate in regions prone to form single-strand DNA (ssDNA) secondary structure (Figure 1C and S1) [41].
To investigate the phenotypes of ISMex4 insertions and the corresponding selection pressure, the icuABT1 or icuABT2 alleles were introduced into WT Methylobacterium to replace icuABWT. Since transposition of ISMex4 dramatically elevated transcription of two putative metal-transport genes, we tested whether metal uptake was enhanced by measuring growth rate and fitness of the WT strain and icuABT1 mutant on methanol in media prepared with various doses of trace metal solution (TMS). Growth rate and fitness of the icuABT1 mutant were significantly higher than the WT strain in media prepared with 0.5-, 1- (regular dose), 2-, 3-, and 4-fold TMS, but differences between these two strains diminished with increasing dose, becoming indistinguishable with a 5-fold dose (Figure 3A). The selective advantage of the icuABT1 mutant and its growth rates relative to those of the WT strain were tightly correlated across tested conditions (Pearson's r = 0.990, P<0.001). These results indicated: (1) Growth media made with the regular dose of TMS were metal deficient and insufficient to sustain optimal growth of Methylobacterium on methanol; (2) Faster growth of the icuABT1 mutant under metal limitation offered a significant competitive advantage.
The observation of poor growth of the WT strain in media with the regular dose of TMS (k = 0.098±0.002) was surprising, given that the growth rate of the same strain at this dose was much higher (k = 0.186±0.003) during the early evolution of these populations. Two observations suggested that the chemical properties of TMS may change upon light exposure: (1) The color of TMS shifted from purple to orange after light exposure (Figure 3B); (2) Growth media made with light-exposed TMS tended to confer faster growth. One potential light-sensitive component in TMS is EDTA, a metal chelator widely applied in growth media to prevent metal precipitation. Previous studies have shown that over-chelation by EDTA can inhibit growth by depleting free metal cations [42],[43]. However, such growth inhibition can be alleviated by exposing media to light, which causes photo-dissociation and photo-degradation of metal-EDTA complexes [44]. We tested if suboptimal growth of Methylobacterium in our media resulted from a similar issue. Indeed, the growth rate difference seen above between the WT and icuABT1 mutant vanished in growth media made with light-exposed TMS, consistent with the EDTA over-chelation model (Figure 3C). To ensure the consistency throughout the experiments, TMS and growth media were stored in the dark. Growth media made with the regular dose of TMS were thus termed metal-poor (MP) media. In addition, a different TMS enriched for unchelated metal cations was developed for making metal-rich (MR) media to facilitate the characterization of the Icu phenotype (see Growth Media in the Materials and Methods section). MR media served as a negative control treatment as growth phenotypes of the WT strain and icuABT1 mutant in MR media were indistinguishable from each other (Figure 3C).
As faster growth of the icuABT1 mutant in MP media supported our hypothesis that increased icuAB expression enhanced uptake of certain metal species, we tested each of the 7 metals in TMS (Ca, Co, Cu, Fe, Mn, Mo, Zn) to see which one accounted for the beneficial effect. We first measured growth rates of the WT strain and icuABT1 mutant in MP media supplemented with a 3-fold extra dose of EDTA or each of the 7 metal species. While 3-fold extra EDTA completely inhibited growth of both strains, addition of any of the metal species improved growth rates of the icuABT1 mutant (data not shown). Growth rates of the WT strain increased to a smaller extent, and only in response to Co, Fe, Mn, or Zn. These results suggest two possibilities: (1) Growth of Methylobacterium in MP media is deficient in all 7 metal species, and overexpression of icuAB confers a fitness advantage by enhancing uptake of all of these metals; (2) Addition of any of these metals saturated the metal-chelation capacity of EDTA, resulting in an increase of free metal cations, one (or more) of which was responsible for poor growth and specifically transported by IcuAB. To circumvent the potentially confounding factor of EDTA chelation, we tested growth of the WT strain and icuABT1 mutant on methanol in EDTA-free growth media (see Growth Media in the Materials and Methods section) titrated for the availability of Co, Fe, Mn, or Zn. In the absence of EDTA, only cobalt limitation dramatically slowed growth of both strains. Critically, growth rates of the icuABT1 mutant were higher than the WT strain at 1.05 (0.062 ppm) and 2.1 (0.124 ppm) nM Co2+ (P<0.05) (Figure 3D). By contrast, growth responses of both strains were indistinguishable under Fe, Mn, or Zn titration (Figure S2), suggesting that the beneficial effect of IcuAB overexpression likely resulted from improving cobalt uptake.
As ISMex4 transpositions ahead of icuAB were nearly universal in populations grown solely or partially on methanol yet were never observed in populations grown solely on succinate, this dichotomy suggested that the advantage of enhancing cobalt uptake came from biochemical reactions specific to methanol (or C1) but not succinate (or multi-C) metabolism. Indeed, in MP media the icuABT1 and icuABT2 mutants received higher fitness gains (15.4±0.7% and 7.3±0.2%, respectively) during growth on methanol than on succinate (0.5±0.3% and 2.2±0.8%, respectively) (Figure 4). To identify the responsible biochemical pathway in C1,metabolism, we characterized growth phenotypes of the WT strain and icuABT1 mutant on C1 (methanol, formate), C2 (ethanol), 3C1+C2 (betaine), C3 (pyruvate), and C4 (succinate) compounds in MP and MR media. In MR media, growth rates of the WT stain and icuABT1 mutant were indistinguishable on all tested substrates (Figure 5A). In MP media, growth of the icuABT1 mutant was significantly faster than the WT strain only on methanol, formate, ethanol, and betaine (P<0.05). As consumption of these four substrates involves metabolism of C1 or C2 units, this pattern suggests that the demand for cobalt may reside in the overlap of C1 and C2 metabolism. One such candidate is the ethylmalonyl-CoA (EMC) pathway. This pathway is required to regenerate glyoxylate from acetyl coenzyme A (acetyl-CoA) during growth on C1 and C2, but not C3 and C4 compounds of Methylobacterium [45]–[47]. Notably, two enzymes in the EMC pathway, methylmalonyl-CoA mutase (MCM) and ethylmalonyl-CoA mutase (ECM), require adenosylcobalamin (AdoCbl, a type of vitamin B12) for catalytic function. We thus hypothesized that cobalt limitation may lower production of AdoCbl, which impedes growth of Methylobacterium on C1 and C2 compounds by slowing down regeneration of glyoxylate through the EMC pathway. This hypothesis predicted: (1) Supplementation with glyoxylate, which has been shown to complement mutants defective in the EMC pathway [48],[49], should enhance growth in MP media; (2) Addition of cobalamin (Cbl) should produce similar effects. Indeed, adding glyoxylate to MP media significantly increased the growth rate of the WT strain but to a lesser extent for the icuABT1 mutant (P<0.05) (Figure 5B), while adding glyoxylate to MR media did not elevate growth rates of either strain. Furthermore, adding Cbl to MP media increased the growth rate of the WT strain slightly (P<0.05) but had no effect on the icuABT1 mutant. Adding Cbl to MR media had no effect on growth of either strain. These results support our hypothesis that shortage of AdoCbl reduces production of glyoxylate via the EMC pathway and thus decelerates C1 metabolism of the WT strain under cobalt limitation.
Insertion of ISMex4 increases expression of the downstream icuA and icuB genes. To investigate the individual contribution of these two genes to fitness gain for methanol growth under metal limitation, we overexpressed icuA, icuB, or icuAB, at two expression levels using expression plasmids carrying the Plac and Ptac promoters, respectively. The promoter activity of the Ptac promoter is approximately 9-fold higher than the Plac promoter [50],[51]. In MP media, overexpression of icuA, icuB, and icuAB by the Plac promoter conferred 16%, 5%, and 16% fitness increases (Figure 6A). Overexpression of the icuA and icuB by the Ptac promoter provided 1% and 2% fitness increases, respectively. Notably, overexpression of icuAB by the Ptac promoter incurred a 13% fitness cost under the same growth condition. As overexpression of membrane proteins is often toxic to the organism [52], the negative impact of expressing icuAB genes at a higher level may result when its cost exceeds the benefit. In MR media, overexpression of icuA, icuB, and icuAB by the Plac promoter conferred no benefit and became deleterious when being expressed by the Ptac promoter. Collectively, these results suggest: (1) Overexpression of icuA is sufficient to produce a fitness gain similar to the icuABT1 allele; (2) An intermediate optimal expression level exists for the icuA, icuB, and icuAB genes; (3) Expression of icuA or icuB alone by the Ptac promoter provided positive selective advantages; however, when these two genes were co-expressed by the same promoter, the sum of fitness effect became negative, indicating a negative sign epistasis [53].
To investigate the functional essentiality of the icuAB gene cassette, we characterized the phenotypes of ΔicuA, ΔicuB, and ΔicuAB strains grown on methanol or succinate in MP or MR media. Deletion of the icuAB cassette was close to selectively neutral (1.006±0.005) during growth on methanol in MR media but resulted in 1.6±0.4%, 1.8±0.7%, and 1±0.1% fitness loss during growth on methanol in MP media, on succinate in MP media, and on succinate in MR media, respectively (Figure S3). Results suggest that Methylobacterium possesses alternative systems to uptake cobalt.
In the WT genetic background, acquiring the icuABT1 allele increased growth rate on methanol by 30% in MP media, but introducing this allele to replace icuABWT of the EM strain did not increase its growth rate under the same growth condition (k = 0.061±0.002 and 0.062±0.004, respectively). In addition, growth rates of the EM strain on methanol in MP and MR media were indistinguishable (k = 0.063±0.001). As the icuABT1 allele emerged in 6 of 8 F populations, these findings raised two questions: (1) Why did the icuABT1 allele exert no detectable effect on growth rate in MP media in the EM genetic background? (2) Why were growth rates of the EM strain in MP and MR media indistinguishable? Growth is a process of biomass assimilation whose rate depends on the rates of multiple resource inputs. A decrease in growth rate may thus weaken advantages conferred by beneficial mutations, like icuABT1, that enhance uptake rates under resource limitation. Since growth of the EM strain was ∼3-fold slower than that of the WT strain, this remarkable difference led us to hypothesize that the selective advantage of the icuABT1 allele may scale generically with the baseline growth rate of the strain. This hypothesis predicts: (1) the selective advantage of the icuABT1 allele should increase when introduced into genetic backgrounds with higher baseline growth rates and (2) the selective advantage should correlate with growth rates modulated by environmental factors independent of cobalt concentrations. First, we measured the fitness effect of the icuABT1 relative to icuABWT alleles in a panel of genetic backgrounds exhibiting different growth rates: the WT strain, the EM strain, strain CM1145, and three EM-derived strains each bearing individual beneficial mutations found in strain CM1145 (see Plasmid and Strain Construction in the Materials and Methods section). Second, we measured the fitness effect of the icuABT1 allele in the WT genetic background across a range of growth rates resulting from incubation at different temperatures. A potential limitation of this approach is that the genetic and environmental treatments applied undoubtedly modify various phenotypes besides just growth rate, such that each perturbation might display unique interactions with the icuABT1 allele. Intriguingly, the selective advantage of the icuABT1 allele in MP media showed a simple and generic trend: significant positive correlations with the with baseline growth rates across all genetic backgrounds (Pearson's r = 0.940, P<0.01) and incubation temperatures (Pearson's r = 0.989, P<0.02) (Figure 6B). On the contrary, fitness and growth rates with or without icuABT1 were indistinguishable across all genetic backgrounds and incubation temperatures in MR media where cobalt is not limiting (data not shown). The above results suggest that: (1) the physiological demand on cobalt uptake is higher under faster growth and (2) the selective advantage of the Icu phenotype in MP media may increase as populations adapt toward faster growth.
Despite having been discovered fortuitously, our work represents the first study to investigate the genetic basis of adaptation to metal limitation in an experimental evolution system. As a component of Cbl (or vitamin B12), cobalt is critical to biosynthesis of this important coenzyme [54]–[56]. Low concentrations of cobalt in the agricultural and marine ecosystems has been shown to impact human and animal health and reduce vitamin B12 production in the ocean, respectively [57],[58]. In this study, evolution under cobalt limitation resulted in emergence of mutants with enhanced cobalt uptake from independent microbial populations. The genetic basis of these independent adaptive events were unusually parallel: resulting from transpositions of ISMex4 into two sites in the icuAB 5′ upstream region in 30 of 32 populations grown partial or solely on methanol. On the contrary, such mutation events were never detected in the 8 populations grown solely on succinate. The highly parallel but distinct evolutionary consequences prompted us to investigate the physiological basis of adaptation and molecular features that might promote parallel genetic evolution. We showed that ISMex4 transposition resulted in overexpression of icuAB genes, which enhanced cobalt uptake and conferred a substantial fitness increase during growth on methanol in MP media but to a minimal extent on succinate. Our physiological assays further pinpointed the major selective advantage to the need for Cbl in the EMC pathway specifically required for methanol metabolism of Methylobacterium, likely resulting from its two AdoCbl-dependent reactions catalyzed by ECM and MCM, respectively. Though the genome sequence suggests two additional Cbl-dependent enzymatic reactions (methione synthase and two ribonucleotide reductases) in Methylobaterium [39], the specific growth defect of the WT strain on methanol in MP media and its complementation by glyoxylate support this notion. Like other bacteria, the cytosolic concentration of Cbl in Methylobacterium is quite low (∼590 nM) [59]. Cobalt deficiency may thus reduce biosynthesis of Cbl, further lowering its concentration in the cytosol, consequently preventing adenosylation of Cbl and its assembly into ECM and MCM.
Our findings from a laboratory system might have profound implications on how cobalt limitation impacts microbial ecology and evolution in nature. Methylobacterium spp. are plant-associated bacteria commonly found on leaves where they compete for nutrients secreted by plants [60]. The ability to utilize methanol, a by-product of plant cell wall synthesis, provides a substantial selective advantage to Methylobacterium during epi- and endophytic growth [61]. Nevertheless, the scarce concentration of cobalt (<8 ppb) in plant tissue may pose a difficulty to cobalt transport of Methylobacterium as well as other plant-associated bacteria [62]. The importance of cobalt to C1 metabolism of Methylobacterium makes it compelling to investigate the functional significance of icuAB during plant colonization. In fact, cobalt limitation in plants has been demonstrated to inhibit growth and root nodulation of nitrogen-fixing rhizobia [63],[64]. Cobalt may also play a role in the crown gall disease caused by Agrobacterium tumefaciens. This pathogen requires indole-3-acetic acid synthesized by a cobalt-containing enzyme to induce abnormal proliferation of plant cells [65]. It would be interesting to apply experimental evolution to study adaptation of plant-associated bacteria in plants grown in cobalt depleted soils. On the other hand, as cobalt is nonessential to plants but essential to many plant microflora [62], it is tempting to ask if the cobalt requirement from plant microflora causes indirect selection on regulation of plant cobalt concentration to welcome mutualistic symbionts or repel harmful pathogens.
On a broader scale, low cobalt concentrations in the environment can greatly impact the supply of vitamin B12 to ecosystems as vitamin B12 is essential to many organisms but only synthesized by prokaryotes [66]. Across the North Atlantic Ocean, the abundance of phytoplankton and dissolved vitamin B12 were found to correlate with cobalt concentrations (0.88−4.77 ppb) [58]. The same study also demonstrated the ability of cobalt to stimulate growth of phytoplankton and vitamin B12 production in seawater. Prochlorococcus and Synechococcus, two dominant photosynthetic bacteria in the open ocean, have an absolute cobalt requirement and appear to secret high-affinity ligands to facilitate cobalt uptake [67],[68]. Combined with genetic and genomic tools, experimental evolution with marine microorganisms represents a promising approach to unravel the genetic and physiological bases of adaptation to metal limitation in the ocean. In addition, the presence of an icuB homologue (72% amino acid similarity, Genbank accession no. ZP_00051365) in the genome of the marine magnetotactic bacterium Magnetospirillum magnetotacticum MS-1 makes it appealing to address its evolutionary origin and ecological significance.
While environmental and physiological constraints set the stage for the emergence of the Icu phenotype, parallel evolution at the genetic level appeared to be promoted by transposition preference of ISMex4, the chromosomal location of the icuAB locus, and clonal interference. In the present study, transposition of ISMex4 conferred a selective advantage by increasing icuAB expression, whereas in another study we found an ISMex4 transposition that increased fitness by reducing the transcript level of an overexpressed gene (Chou and Marx, unpublished), stressing the versatile role of IS elements in the evolution of gene expression. Transposition of most IS elements exhibit some degree of target site selectivity [69]. Analysis of ISMex4 insertion sites revealed a 4-bp conserved target sequence that tends to locate in regions prone to form a stem-loop structure. The presence of two ISMex4 copies 15 kb and 38 kb downstream of the icuAB genes, respectively, may have increased its chance of transposition into this nearby locus as several IS elements exhibit a high frequency of local-hopping. In addition to aforementioned features that may promote recurrent ISMex4 transpositions, the predominance of the high-fitness icuABT1 allele over the icuABT2 allele across populations suggests a potential role for clonal competition among asexual lineages (Table 1). Identification of two methanol-evolved populations (F5, F6) free of icuABT1 and icuABT2 alleles pointed to the possibility of alternative mutational targets. Compared to growth in MR media, growth rates of evolved isolates from these two populations were just 20% lower in MP media (data not shown), similar to the phenotype of the icuABT1 mutant. Collectively, these results suggest a model shaping genetic parallelism in our system: local-hopping and target selectivity of ISMex4 may lead to high frequency but limited types of transposition while the large fitness advantages gained by icuABT1 and icuABT2 alleles allow them to outcompete other weaker beneficial mutations conferring similar phenotypes in these asexual populations. As the proposed genetic features favoring ISMex4 transposition and its resulting selective advantage can be manipulated easily through mutagenesis and trace metal supplementation, respectively, our system offers the power to experimentally address how mutation rates and the strength of natural selection affect parallel evolution and the dynamics of adaptation.
The physiological effects of an allele depend on expression levels, genetic backgrounds, and environmental conditions. Predicting the behavior and evolution of biological systems requires a comprehensive understanding of how these parameters influence physiology and thus shape the fitness landscape. Experimental evolution offers a valuable alternative besides conventional genetic approaches to uncover biochemical functions and physiological links of genes as well as their contribution to fitness in the evolutionary context. In this study, growth phenotypes of icuAB knockout mutants are minute and unspecific to either carbon substrate or growth media, providing no clue to the functional significance of this gene cassette. Nevertheless, by characterizing the phenotypes of beneficial mutations and reconstructing their fitness effect through overexpression experiments, our results revealed the biochemical function of this gene cassette and demonstrated an intermediate optimal expression level that constrains the breadth of phenotypic evolution. Moreover, identification of the physiological processes icuABT1 and icuABT2 contribute to sets the stage to address whether they interact with other mutations or environments in a manner similar to those tested for deleterious mutations. Previous work has shown that growth defects of deleterious mutations tend to be reduced by either environmental stress or the presence of other deleterious mutations [7],[70],[71]. These results have supported a simple model where growth rate is mainly limited by the slowest physiological process [7]. It has remained unclear whether the same principle would apply to certain beneficial mutations, such that they become more advantageous when limitations imposed by other physiological processes are relieved. By modulating growth rate through either incubation temperatures or genetic backgrounds, we found a consistent increase in the selective advantage of a beneficial mutation with increasing growth rate. This growth-rate dependence is in accord with the model described above: By alleviating genetic or environmental constraints, increases in growth rate raised the fitness effect of increased cobalt uptake. The synthesis of previous work with deleterious mutations and current findings from a beneficial mutation suggest a physiology-mediated mechanism through which mutations and environments interact. This mechanism has two profound implications for the evolution and function of biological systems: (1) Some mutations will only be beneficial (or deleterious) when favorable mutations or environmental changes alleviate other physiological limitations, suggesting a general mechanism for historical contingency and environment-dependent evolutionary potential. (2) As higher-order phenotypes (e.g. growth, differentiation, development, locomotion) integrate across multiple physiological inputs, genes and environmental factors that affect seemingly distant physiological processes may thus interact through their convergent effects upon higher-order phenotypic outputs. We anticipate similar observations will continue to emerge from further exploration of the commonality of epistatic, G×E, and even environment-by-environment interactions as flavors of the same phenomenon: systems-wide physiology-mediated interactions.
Unmarked allelic exchange plasmids for introducing adaptive mutations or deleting genes were constructed based on pCM433, a sacB-based suicide plasmid [72]. Two 3,380-bp PCR fragments containing icuABT1 and icuABT2 alleles were cloned into pCM433 to generate pHC40 and pHC82, respectively (Table S1). The 606-bp and 579-bp PCR products containing pntA1145 (encoding membrane-bound transhydrogenase) and gshA1145 (encoding γ-glutamylcysteine synthetase) alleles from strain CM1145 were cloned into pCM433 to generate pHC36 and pHC38, respectively. Plasmids pHC65, pHC67, and pHC68 designed to delete icuA, icuB, and icuAB, respectively, were generated by consecutively cloning the 5′ upstream and 3′ downstream regions encompassing the designed deletions into pCM433. A constitutive expression plasmid and a fluorescent promoter-probe plasmid were constructed based on pCM132, a broad-host-range plasmid conferring kanamycin resistance [51]. The lacZ gene of pCM132 was deleted and replaced by a 33-bp multiple cloning site to generate pHC41. The promoter-probe plasmid, pHC42, was generated by cloning a 734-bp PCR fragment containing the ribosome binding site (RBS) of the fae gene (encoding formaldehyde-activating enzyme) of Methylobacterium and the reporter GFPuv gene from pKF133 into pHC41 [73],[74]. A 51-bp synthetic fragment containg the constitutive promoter Ptac was inserted upstream of this reporter to make pHC62. Fragments for testing promoter activity were PCR amplified and inserted into the same cloning site of pHC42 to make pHC44, pHC46, pHC47, pHC51, pHC55. Expression plasmids, pHC60 and pHC91, were generated by cloning the aforementioned 51-bp fragment containing Ptac and a 44-bp synthetic fragment containing the Plac promoter into pHC41, respectively. PCR fragments containing RBSfae-icuA, RBSfae-icuB, or RBSfae-icuAB were subsequently cloned into pHC60 and pHC91 to generate pHC69, pHC70, pHC71, and pHC92, pHC93, pHC94, respectively.
The EM strain is a variant (Chou and Marx, unpublished) of a previous strain (CM253K.1 with pCM106) shown to be capable of slow growth on methanol [36]. This strain lacks a functional tetrahydromethanopterin-dependent formaldehyde oxidation pathway due to deletion of mptG (encoding β-ribofuranosylaminobenzene 5′-phosphate synthase [75]) that eliminated tetrahydromethanopterin biosynthesis. Instead, two genes belonging to the foreign glutathione-dependent formaldehyde oxidation pathway of Paracoccus denitrificans (flhA, encoding hydroxymethyl-glutathione dehydrogenase and fghA, encoding formyl-glutathione hydrolase) were expressed from the strong PmxaF promoter in plasmid pCM160 [51]. This replacement resulted in restoration of growth on methanol at a one-third the rate of WT [36]. EM-derived strains carrying one of the three other adaptive mutations from strain CM1145 were generated as above. These beneficial mutations affected the expression of aforementioned fghA, pntAB, and gshA genes. Further analysis of the physiological effects of these beneficial mutations will be described subsequently (Chou and Marx, unpublished). The icuABWT allele was introduced into strain CM1145 by the same allelic exchange method using pHC39. The icuABT1 allele was introduced into the WT strain, the EM strain, and EM-derived strains bearing individual adaptive mutations using pHC40. The icuABT2 allele was introduced into the WT strain using pHC82. Gene knockouts of icuA, icuB, and icuAB were generated by deleting the whole open reading frames from the WT strain. The genotypes of resultant mutants were confirmed by PCR. Strains carrying promoter-probe plasmids or expression plasmids were made by introducing these plasmids from E. coli 10-beta (New England Biolabs) into WT Methylobacterium, or its isogenic strain CM1180 [35] expressing the yellow fluorescent protein Venus, through tri-parental mating [76].
The general formula for one liter of all growth media consists of 1 ml of TMS, 100 ml of phosphate buffer (25.3 g of K2HPO4 and 22.5 g of NaH2PO4 in 1 liter of deionized water), 100 ml of sulfate solution (5 g of (NH4)2SO4 and 0.98 g of MgSO4 in 1 liter of deionized water), 799 ml of deionized water, and the desired carbon sources. One liter of the TMS (pH 5) used in MP media and growth media for evolution experiments consists of 12.738 g of EDTA disodium salt dihydrate, 4.4 g of ZnSO4·7H2O, 1.466 g of CaCl2·2H2O, 1.012 g of MnCl2·4H2O, 0.22 g of (NH4)6Mo7O24·4H2O, 0.314 g of CuSO4·5H2O, 0.322 g of CoCl2·6H2O, and 0.998 g of FeSO4·7H2O in 1 liter of deionized water [35]. The growth media used for evolution experiments were prepared with this photoactive TMS stored under variable light exposure [35]. MP media were prepared with the same TMS but stored in dark to prevent photochemical reactions. For light exposure experiments, the same TMS was aliquotted into 15 ml plastic tubes (Falcon) covered or uncovered with aluminum foil, then subject to constant light source (broad spectrum, 81 µmol photons m−2 s−1) for 1 month at 25°C. TMS used in MR media was modified by adding a 4-fold extra dose of iron to displace chelated metals from EDTA-metal complexes [77]. This modified TMS consisted of 10 ml of 179.5 mM FeSO4, 80 ml of premixed metal mix (12.738 g of EDTA disodium salt dihydrate, 4.4 g of ZnSO4·7H2O, 1.466 g of CaCl2·2H2O, 1.012 g of MnCl2·4H2O, 0.22 g of (NH4)6Mo7O24·4H2O, 0.314 g of CuSO4·5H2O, and 0.322 g of CoCl2·6H2O in 1 liter of deionized water, pH 5), and 10 ml of deionized water. EDTA-free media were prepared without adding premixed TMS. Instead, each of the 7 trace metal species was supplemented as 0.1 ml of separate solutions (153.02 mM ZnSO4, 99.71 mM CaCl2, 51.13 mM MnCl2, 1.78 mM (NH4)6Mo7O24, 12.58 mM CuSO4, 13.53 mM CoCl2, and 35.9 mM FeSO4). Glassware used with EDTA-free media was pre-washed with 0.05 N HCl to eliminate trace metal remnants.
The A, B, C and D populations were founded by the WT strain [35] while the F populations were founded by the EM strain. All populations were grown in 9.6 ml of growth media contained in 50 ml Erlenmeyer flasks and incubated in a 30°C shaking incubator at 225 rpm. Growth media for evolution experiments consisted of identical minimal growth media supplemented with different carbon sources: A and F populations with 15 mM methanol, B populations with 3.5 mM succinate, C populations with 7.5 mM methanol and 1.75 mM succinate, and D populations alternating between 15 mM methanol and 3.5 mM succinate. The A, B, C and D populations were transferred into fresh growth media at 1/64 dilution rate every two days. Due to the slow growth of the EM strain, the F populations were transferred at the same dilution rate every four days in the first 300 generations of evolution. Transfers of the F populations after generation 300 were switched to two-day cycles. Populations were sampled periodically and preserved at −80°C for later analysis.
For each strain, three independent mid-exponential phase cultures (defined as half-maximal OD600) were harvested and processed by a method described previously [31]. Total RNA was extracted using the RNeasy Mini Kit (QIAGEN), followed by removing residual genomic DNA with the Turbo DNA-free Kit (Ambion). The absence of DNA contamination was verified by PCR amplification of an untranscribed region using primers CM-mxaEdf (5′CTAAGGAAGCCCTGCGATG-3′) and CM-mxaEdr (5′-CCCTCCCGTCTGTTTTTCC-3′). RNA was quantified by a Nanodrop ND-1000 Spectrophotometer (Thermo Scientific) and checked for degradation by an Agilent Bioanalyser 2100 or agarose gel electrophoresis. The preliminary microarray experiment used three independent RNA isolations from each strain that were pooled together with equal quantity. cDNA synthesis, labeling, hybridization to Agilent 60-mer oligo microarrays, and scanning of microarrays were performed by MOgene by following a previously described procedure [31]. cDNA synthesis for real-time PCR was performed using 1 µg total RNA with the qScript cDNA Synthesis Kit (Quanta Biosciences) according to the manufacturer's instructions. The primers used to amplify and detect transcripts of the icuA, icuB, and rpsB genes are HCAM105 (5′-GCTTGCCACCTTCAGCCAGATC-3′) and HCAM106 (5′-ATGGTGACCTTGTTGAAGGCGTTGTA-3′), HCAM107 (5′-TCATCCTCACCGCGCTGC-3′) and HCAM108 (5′-GCTTTGAGCGCGGGCATTG-3′), and HCAM111 (5′-TGACCAACTGGAAGACCATCTCC-3′) and HCAM113 (5′-TTGGTGTCGATCACGAACAGCAG-3′), respectively. Two-step real-time PCR experiments were performed in three replicates with the PerfeCTa SYBR Green SuperMix (Quanta Biosciences) according to the manufacturer's instructions on a DNA Engine Opticon2 (MJ Research). The rpsB gene (encoding 30S ribosomal protein S2) was chosen as the reference gene for data normalization. Data analysis was performed with the Opticon Monitor v. 2.02 (MJ Research). The average threshold cycle (Ct) value for each gene was calculated from triplicate reactions for RNA samples by following a previously described method [78]. The ΔCt value described the difference between Ct of the target gene and Ct of the reference rpsB gene. The ΔΔCt value described the difference between the ΔCt of the WT strain and that of the evolved or mutant strains. The difference in expression was calculated as 2ΔΔCt.
Genomic DNA of 3–6 isolates from each evolved populations was prepared using an alkaline lysis DNA extraction method [79]. The 5′ upstream region of the icuABWT allele was amplified by primer HCAMp7 (5′-CCGATGGTGAGATCTGGGTCTTCAG-3′) and HCAMp8 (5′-CGTCACCTCCTGACATCTCGATTTAC-3′). Insertions of ISMex4 upstream of the icuAB cassette were detected by PCR amplification using primer HCAMp7 and HCAMp38 (5′-ACCAGCACCCGTCCGAGC-3′). The sizes of PCR products were determined by electrophoresis in 1% (w/v) TAE agarose gel. In cases where no PCR product was obtained from sampled isolates, the genomic DNA of the corresponding populations was extracted and PCR amplified through the same means. PCR products of interest were purified with the QIAquick PCR Purification Kit (QIAGEN) and sequenced by MWG Biotech.
Prior to growth rate assays and competition experiments, all strains were acclimated in growth medium supplemented with carbon sources used in the ensuing assays. Three replicate cultures of each strain were sampled periodically and the change in OD600 was measured using a Bio-Rad microplate reader model 680. Competition experiments were performed by following a previously described procedure [35]. Briefly, after one round of acclimation, test strains and a reference strain expressing the yellow fluorescent protein Venus were mixed by a 1∶1 volume ratio, diluted 1∶64 into 9.6 ml of fresh media which were incubated under the conditions described above. The ratios of non-fluorescent cells in mixed populations were measured by passing population samples before (R0) and after (R1) competition growth through a BD LSR II flow cytometer (BD Biosciences) for at least 50000 cell counts per sample. Fitness values (W) relative to the reference strain were calculated by a previously described equation assuming an average of 64-fold size expansion of mixed populations during competitive growth [35]:
Prior to fluorescence measurements, strains harboring plasmids derived from pHC42 were acclimated in MP media plus 15 mM methanol and 25 µg/ml kanamycin sulfate. Cultures were then grown to exponential phase in the same media without antibiotic. Optical density values at 600 nm (OD600) and fluorescence intensities were measured by a Safire2 microplate reader (Tecan). The excitation and emission wavelengths for GFPuv were set as 397 nm and 506 nm, respectively [80]. The WT strain was used as control to determine cellular autofluorescence of Methylobacterium. To normalize the fluorescence intensity, the fluorescence value of a sample was first divided by its OD600. This ratio for the negative control was then subtracted from those of samples to obtain the fluorescence above background. Finally, these values were normalized by dividing the negative control ratio to give the GFPuv fluorescence relative to the background autofluorescence.
|
10.1371/journal.pbio.1000098 | Specificity and Plasticity of Thalamocortical Connections in Sema6A Mutant Mice | The establishment of connectivity between specific thalamic nuclei and cortical areas involves a dynamic interplay between the guidance of thalamocortical axons and the elaboration of cortical areas in response to appropriate innervation. We show here that Sema6A mutants provide a unique model to test current ideas on the interactions between subcortical and cortical guidance mechanisms and cortical regionalization. In these mutants, axons from the dorsal lateral geniculate nucleus (dLGN) are misrouted in the ventral telencephalon. This leads to invasion of presumptive visual cortex by somatosensory thalamic axons at embryonic stages. Remarkably, the misrouted dLGN axons are able to find their way to the visual cortex via alternate routes at postnatal stages and reestablish a normal pattern of thalamocortical connectivity. These findings emphasize the importance and specificity of cortical cues in establishing thalamocortical connectivity and the spectacular capacity of the early postnatal cortex for remapping initial sensory representations.
| During brain development, the emergence of distinct areas in the cerebral cortex involves an interplay between patterning of the cortical sheet in the early embryo and later influences of incoming connections made from other brain areas, namely the thalamus. Connectivity between the thalamus and the cortex is initially smooth and graded, and a prominent model for how thalamocortical connectivity is established proposes thalamic axons are topographically sorted as they course through subcortical regions and then passively delivered to appropriate areas of the cortical sheet. We have used mutant mice lacking the guidance molecule Semaphorin-6A to test this model. In these mutants, Semaphorin-6A axons from the visual part of the thalamus are subcortically misrouted and fail to innervate the presumptive visual cortex, which is instead invaded by somatosensory thalamic axons. Despite this major disruption in initial connectivity, many visual thalamic axons find their way specifically to visual cortex, arriving several days later than usual. These late-arriving axons often follow alternate routes, and upon arrival are able to out-compete earlier-arriving somatosensory axons to reestablish grossly normal thalamocortical connectivity. These results argue strongly against an essential role for early subcortical targeting in the establishment of thalamocortical connectivity patterns and suggest instead the existence of highly specific target-selection mechanisms that match thalamic axons with appropriate cortical areas.
| A dynamic interplay exists between the processes of cortical arealization and those controlling the guidance and targeting of thalamocortical projections [1–5]. Early in development, both the thalamic field and the cortical sheet appear homogeneous in cytoarchitecture, and connections between them form in a smoothly topographic fashion, with dorsolateral thalamus projecting to caudal cortex and ventromedial thalamus to rostral cortex [6–8]. The cytoarchitectonic resolution of these fields into discrete cortical areas and thalamic nuclei occurs later [8–13] with the elaboration of many aspects of the cortical areas dependent on appropriate thalamic innervation [1–5,14,15].
Several lines of evidence have led to the theory that subcortical sorting of thalamic axons within the ventral telencephalon largely determines their final targeting within the cortex [16–20]. For example, in mutants in the transcription factor Ebf1 or in the Dlx1/Dlx2 double mutants, a subset of thalamic axons is misrouted ventrally, resulting in a caudal shift of the remaining axons within the ventral telencephalon [16]. This shift is projected onto the cortex so that at birth, caudal cortical areas are contacted by axons that would normally project to more rostral areas. The ultimate effect of this derangement on thalamocortical connectivity could not be assessed in these mutants, however, as they die perinatally.
On the other hand, many experiments have revealed the existence of cortex-specific cues that control thalamocortical targeting [21–27]. For example, changes in patterning of the cortical sheet in Emx2 [21,24], Fgf8 [23], or COUP-TF1 [27] mutants lead to parallel alterations in the patterns of thalamocortical connectivity. In each of these cases, manipulations solely in the cortex dramatically affect thalamocortical targeting. Indeed, ectopic expression of Fgf8 in either the subplate or cortical plate further revealed that thalamocortical axons (TCAs) are responsive to guidance cues present in both the subplate and cortical plate [26]. The interplay between subcortical and cortical mechanisms in determining eventual thalamocortical connectivity thus remains to be resolved.
To get a better understanding of the interactions between areal patterning and thalamic axon guidance, we have used the Sema6A gene trap mouse. As a consequence of Sema6A disruption, a large fraction of thalamic projections gets derailed within the ventral telencephalon [28]. As these mice survive to adulthood they provide a unique model with normal cortical patterning but altered thalamic input during embryonic life. Our study reveals a changing pattern of thalamocortical development in the Sema6A mutants, drawing attention to the spectacular capacity of the cortex for altering and organizing its initial sensory representation. In particular, our findings suggest that thalamic axons from the dorsal lateral geniculate nucleus (dLGN) can target their correct area even if they arrive there days later than normal, through alternate subcortical routes. They also indicate that dLGN axons can out-compete invading axons from inappropriate thalamic nuclei, establishing a surprisingly normal adult cortical representation.
Sema6A is broadly expressed in the thalamus at embryonic day (E)14.5, a time when thalamic neurons are extending axons towards their cortical targets [28], with expression highest in the dorsolateral aspect (n = 5; Figure 1A and 1B). Sema6A is also strongly expressed in the amygdala and the ventral telencephalon, and weakly expressed in the neocortex at this age, localized to the most superficial compartments (Figure 1A and 1B). At late embryonic stages, Sema6A is also expressed by deep cortical plate neurons, eventually layer 5 (unpublished data). Staining with the axonal marker placental alkaline phosphatase (PLAP), encoded on the gene trap cassette in this Sema6A allele, revealed thalamic axons extending from the thalamus through the internal capsule and towards the neocortex (Figure 1C and 1D).
A previous study using PLAP staining showed that many thalamic axons were misrouted in the Sema6A−/− brains at embryonic stages [28]. To further examine the guidance of TCAs in the absence of functional Sema6A protein, we performed carbocyanine dye tracing studies. Broad injections of DiI in the thalamus (including the dorsolateral aspect) at E15.5 revealed a prominent derailment of thalamic axons at the surface of the ventral telencephalon and amygdala in Sema6A−/− embryos (n = 4/4; Figure 2D and 2G–2I), compared to the normal route of navigation through the internal capsule towards the neocortex observed in wild-type embryos (n = 4/4; Figure 2A and 2C). The derailment of a large proportion of TCA fibers at the ventral telencephalon in Sema6A−/− brains at E16.5 was confirmed by neurofilament (NF) immunohistochemistry (n = 3; Figure 2E and 2F) and by PLAP staining (Figure S1A and S1C). Overlaid consecutive serial sections of Sema6A−/− brains at E17.5 stained for NF reveal more clearly the extent of the TCA derailment (n = 3; Figure 2H).
To identify the precise origin of the misrouted thalamic axons within the thalamus, we placed small DiI crystals at the site of the derailed fibers near the ventral surface of the telencephalon (Figure 3A) in wild-type (n = 4) and Sema6A−/− (n = 4) postnatal day (P)0 brains. Whereas no back-labeled cells were observed in any thalamic nuclei in wild-type brains (Figure 3B and 3C), several axon bundles were retrogradely traced to the dorsolateral aspect of the thalamus in Sema6A−/− brains (Figure 3D and 3E). Retrogradely labeled cells were specifically found in the presumptive dLGN (Figure 3G and 3H). Some labeled bundles were also observed ascending laterally towards the cortex (Figure 3E). Moreover, in Sema6A−/− brains, at more-caudal telencephalic levels, DiI-labeled axons were observed running through the intermediate zone of the primary visual cortical area (Figure 3F and 3I; and unpublished data), suggesting that some dLGN axons that follow this abnormal route might still reach the visual cortex. Indeed, whereas a DiA crystal placed in the internal capsule zone of wild-type brains at E17.5 back-labeled cells throughout the dorsal thalamus (n = 2/2; Figure S2A–S2C), a similarly placed DiA crystal in Sema6A−/− brains at the same age strikingly did not label cells in the dLGN (n = 2/2; Figure S2D–S2F).
To further study the trajectories of the dLGN projections, we placed small DiI crystals into the dLGN of wild-type and Sema6A−/− brains at P0. At this age, wild-type and heterozygous dLGN axons extended through the internal capsule and arrived to the visual cortex in a normal fasciculation pattern (n = 8/8; Figure 4A–4D). In Sema6A−/− brains, in contrast, dLGN axons projected ventrally to the most superficial regions of the ventral telencephalon. Despite this, a small number of dLGN axon terminals were observed in visual cortex (Figure 4G; n = 9/9). These results demonstrate that the vast majority of visual TCAs are severely affected by the lack of Sema6A function in vivo, but some still reach presumptive visual cortex by birth, possibly by alternate routes.
To determine whether the initial derailment of dLGN axons could affect the general topographical arrangement between neocortex and thalamus, we performed multiple cortical dye placements at E16.5 and P0. We placed DiI and DiA crystals into the putative visual and somatosensory cortices, respectively, in Sema6A+/− and Sema6A−/− mutant brains (Figure 5). In Sema6A+/− animals (n = 6/6 E16.5; n = 5/5 P0), a DiI crystal placement in the occipital neocortex labeled thalamic cell bodies and cortical axons in the dLGN (Figure 5A and 5B). Placements of DiA in the parietal neocortex labeled cells and axon terminals in a more medial thalamic domain, where the ventrobasal (VB) complex is located (Figure 5A and 5B). Interestingly, similar placements performed in Sema6A−/− embryos showed a lateral to medial shift in the thalamocortical and reciprocal corticothalamic connectivity (n = 8/8 E16.5; n = 5/5 P0; Figure 5C and 5D). Specifically, placements into the occipital cortex labeled large numbers of cells in the lateral part of VB in Sema6A−/− brains, whereas few, if any cells at E16.5 were back-labeled in the dLGN (Figure 5I–5L; 96.56 ± 4.6% of thalamic cells back-labeled from occipital cortex in E16.5 Sema6A−/− brains were found in the VB, compared to just 1.25 ± 2.25% of back-labeled thalamic cells in Sema6A+/− brains, p < 0.0001). At P0, an increase in the number of cells back-labeled in dLGN was observed in the mutants, though this was still far fewer than in wild-type animals (Figure 5K and 5L; 16.24 ± 6.9% of thalamic cells back-labeled from occipital cortex in P0 Sema6A−/− brains were found in the dLGN, compared to 91.61 ± 3.17% of back-labeled thalamic cells in Sema6A+/− brains, p < 0.0001). Placements of DiA in parietal cortex in Sema6A−/− brains back-labeled VB cells, though this labeling was more medial within this nucleus than in Sema6A+/− animals (Figure 5C and 5D). Moreover, the areas back-labeled within VB by placements into occipital and parietal cortex were contiguous but showed minimal overlap. We did not observe any double-labeled cells.
Taken together, these data show that in Sema6A−/− embryos and neonates, axons of the dLGN specifically are misrouted to the surface of the ventral telencephalon, and the topography of thalamocortical projections from VB is expanded caudally into the unoccupied occipital cortex.
Because Sema6A is also expressed in retinal ganglion cells, we wanted to test whether abnormal development of retinal projections in Sema6A−/− embryos might secondarily cause the defects observed in the projection of visual TCAs. We injected DiI and DiA crystals in the eye and V1, respectively, in wild-type and Sema6A mutant mice embryos at E16.5. In wild-type embryos (n = 3), retinogeniculate projections were observed in the ventral lateral geniculate nucleus (vLGN) and dLGN (Figure 5E and 5F). At the dLGN, this projection overlapped with back-labeled cells and cortical axons labeled from the visual cortical injections (Figure 5F). However, in Sema6A−/− embryos (n = 3/3), although the interconnectivity between visual cortex and thalamus was shifted medially, retinogeniculate projections still arrived to the vLGN and dLGN in a normal fashion (Figure 5G and 5H). Thus, the defects observed in the development of visual TCAs are not due to abnormal development of the projection from the retina or to abnormal differentiation of the LGN, which can still act as a specific target for retinal axons.
Similarly, to test whether changes in cortical patterning could account for the topographical defect observed in TCAs connectivity in Sema6A−/− mutants, we performed in situ hybridization of specific cortical area markers on both wild-type (n = 9) and Sema6A−/− (n = 8) brains at P0. In wild-type brains EphA7 is expressed both rostrally and caudally in the cortex, but is largely absent from parietal cortex, whereas EphrinA5 is expressed in a complementary pattern (Figure S3A and S3C). We did not find any changes in the expression of EphA7 and EphrinA5 in Sema6A−/− brains (Figure S3B and S3D). The expression of other cortical area markers (Cad6, Cad8, and RZRβ) was also unaltered in Sema6A−/− brains at P0 (unpublished data). At earlier (E14.5 and E16.5) and later (P7) developmental stages, there were also no differences observed in the expression of cortical area markers between wild-type and Sema6A−/− brains (unpublished data). These data indicate that intrinsic cortical patterning is unaltered in Sema6A mutants.
We investigated thalamocortical connectivity at P4 using multiple cortical dye placements. Small crystals of DiI and DiA were placed in the primary visual (V1) and primary somatosensory (S1) cortices, respectively, in wild-type and Sema6A−/− brains (Figure 6). In wild-type brains (n = 2/2), a DiI placement in V1 labeled thalamic cell bodies and cortical axons in the dLGN, whereas a DiA placement in S1 labeled cell bodies in the VB (Figure 6A and 6B). Interestingly, in Sema6A−/− brains (n = 2/2), a DiI placement in V1 mainly labeled thalamic cell bodies in the dLGN (Figures 6C, 6D, and 5J; 79.49 ± 25.3% of thalamic cells back-labeled from occipital cortex in P4 Sema6A−/− brains were found in the dLGN, compared to 99.49 ± 1.51% of back-labeled thalamic cells in wild-type brains, p = 0.0428), whereas far fewer cells were labeled in VB, in sharp contrast to results from similar tracing experiments at E16.5 and P0 (Figure 5K. In Sema6A−/− brains at P4, just 20.5 ± 25.3% of cells back-labeled from the occipital cortex were found in the VB, compared to 96.56 ± 4.68% and 83.65 ± 6.9% in Sema6A−/− brains at E16.5 and P0, respectively, p < 0.0001). A DiA placement in S1 labeled cell bodies in the VB as expected (Figure 6C and 6D). This apparent rapid recovery of the normal thalamocortical connectivity was observed despite a persistent misprojection of TCAs from the dLGN to the ventral telencephalon in Sema6A−/− brains at P4 (unpublished data). Although the topographical sorting of TCAs was apparently restored at this stage, fewer back-labeled cells were detected in the dLGN after visual cortical dye placements in Sema6A−/− compared to wild-type brains (Figure 6B and 6D).
To characterize the extent to which any shift in thalamocortical connectivity persists in the adult Sema6A−/− mouse, we performed in vivo stereotaxic tracing studies. We injected red and green cholera toxin (CT) dyes in V1 and S1, respectively, in wild-type (n = 6) and Sema6A−/− (n = 8) mice at P60 (Figure 6E, 6F, 6H, and 6I). At this adult stage, injections of red and green CT showed a normal topographical arrangement of the thalamocortical connectivity in Sema6A−/− brains (Figure 6I) when compared with wild-type brains (Figure 6F).
We also used two different colors of retrograde tracing microspheres, injected in V1 (red) and S1 (green) in wild-type or heterozygous (n = 5) and Sema6A−/− (n = 6) mice at P60. The back-labeled thalamic cells of each color were plotted in the corresponding thalamic nuclei, dLGN or VB, in wild-type (Figure 6G) and Sema6A−/− brains (Figure 6J). These data demonstrate that the early embryonic shift in TCAs connectivity, observed at E16.5 and P0, is partially recovered at P4 and totally compensated in the adult Sema6A−/− mouse.
To investigate whether the misrouted TCAs from the dLGN persist in the adult mouse, we performed PLAP staining studies to reveal the pathway of Sema6A-positive axons in both Sema6A+/− (n = 4) and Sema6A−/− (n = 4) brains. Sema6A is expressed in oligodendrocytes in adults, and this staining thus labels all myelinated fibers. Surprisingly, in Sema6A−/− brains at P60, a misrouted bundle of axons was still observed at the ventral-most region of the telencephalon (Figure 7F, 7L, and 7M) in a similar location to the misrouted TCAs shown at earlier developmental stages. Similar ectopic bundles were never observed in any heterozygous adult brains (Figure 7A).
Interestingly, we could follow some PLAP-labeled axons up to the level of the visual cortex in Sema6A−/− brains (Figure 7H–7M). Misrouted labeled axons appear to follow one of two alternate routes: (1) they either turn laterally through the amygdala and join the external capsule, or (2) they continue to project ventrally and extend along the superficial margin of the telencephalon (Figure 7L–7N). In some cases, abnormal bundles could be followed either through the marginal zone or the external capsule up to the cortex (Figure 7L). At more caudal regions, misrouted axons can be seen close to the pial surface of the cortex, including the visual cortex, of Sema6A−/− brains (Figure 7K).
To confirm the recovery pathways observed in PLAP-stained adult Sema6A−/− brains, a solution of DiI was injected into the primary visual cortex of wild-type or Sema6A−/− brains at P7. These injections back-labeled cell bodies in the dLGN specifically of both wild-type (n = 4; unpublished data) and Sema6A−/− brains (n = 3; Figure 8A). In wild-type animals, the axons of these back-labeled cells could be followed as they projected through the internal capsule to the cortex (unpublished data). In contrast, in Sema6A−/− brains, many labeled axons were observed projecting ventrally and rostrally in the ventral telencephalon, turning laterally to reach the external capsule at more rostral levels (Figure 8B–8D). Additionally, many axons were also seen to project in a more canonical fashion through the internal capsule, with some axons initially projecting ventrally before turning dorsally again to loop back and enter the internal capsule (Figure 8D7), suggesting that many initially misrouted dLGN axons may extend a branch at later stages and extend through the internal capsule to reach the cortex.
To investigate whether the massive misrouting of TCAs during early embryonic stages has an impact on the development of the dLGN nucleus and visual cortex, we performed a series of histochemical studies in adult Sema6A−/− mice. Using both Nissl and cytochrome oxidase staining, we observed a significant reduction in the volume of the dLGN in Sema6A−/− brains (n = 4) compared with wild-type littermates (n = 4) at P30 (Figure 9A–9E). These data suggest that dLGN neurons whose axons do not reach their target on time die during development. To further test this possibility, we performed an immunostaining against a cleaved caspase-3 to detect apoptotic cells in the thalamus of wild-type and Sema6A−/− brains at P4. At this stage, we observed an almost 3-fold increase in the number of apoptotic cells in the dLGN of Sema6A−/− brains (n = 5) compared to controls (n = 6; Figure 9F–9J), suggesting that many of the dLGN neurons, whose axons were misrouted, do indeed die during development.
We next investigated whether the early lack of dLGN afferents to the visual cortex together with the transient invasion of somatosensory input to this cortical area might affect the final relative representation of cortical areas in Sema6A−/− adult mice. We examined the cortical area occupied by S1 and V1 in tangential sections stained for serotonin (5HT) immunoreactivity in wild-type (n = 5) and Sema6A−/− (n = 5) brains at P7. Although we observed no changes between wild-type and Sema6A−/− brains in the relative position of these cortical areas (Figure 9K and 9L), we observed a significant reduction in the size of V1 in Sema6A−/− mouse brains (Figure 9L and 9M). Moreover, we observed a consistent change in the shape of the V1 cortical domain in Sema6A−/− mouse brains compared to wild-type littermates (Figures 9K and 9L). No changes were observed in the position and dimensions of the barrel field in Sema6A−/− mice. Together, these results strongly suggest that the reduction in the size of dLGN in Sema6A−/− mice leads to a reduction in the size of V1.
Our study of the Sema6A mutants revealed an initial subcortical pathfinding defect of thalamic axons specifically from the dLGN. This results in expansion of somatosensory thalamic axons into presumptive visual cortex during embryonic stages. Due to the viability of these mutants, we were able to assess the secondary consequences of early misrouting of the visual axons on postnatal cortical specification and adult thalamocortical topography. Remarkably, many dLGN axons are able to find their way to visual cortex during early postnatal stages, following alternate routes, and can establish almost normal patterns of thalamocortical connectivity in the adult. The general implications of these findings for principles of thalamic axon guidance and cortical arealization are discussed below.
The failure of dLGN axons to arrive to the occipital cortex in Sema6A−/− brains at embryonic stages results in the dramatic expansion of the domain of VB axons into this region. Importantly, we observe no changes in cortical gene expression patterns at early stages, indicating that the removal of Sema6A from the cortex does not affect global cortical patterning. This early caudal shift of thalamocortical targeting has also been observed in the other mutants with misprojected dLGN axons (i.e., Ebf1, Dlx1–2 double mutants [16]; reviewed in [2,18]). We observed back-labeling of VB from injections placed in occipital cortex, but did not observe a dramatic shift in connectivity from more rostral injection sites in parietal cortex (the normal position of S1), which still back-labeled VB (though more medially). This is more consistent with a caudal expansion of the innervation zone of VB axons in Sema6A mutants than with an overall shift of all thalamic connections.
A current model of thalamic axon pathfinding proposes an essential role for intermediate targets, in the ventral telencephalon (vTel), in guiding TCAs to specific cortical areas [2,17,18,20,29–31]. Mutations in a number of genes expressed predominantly in the vTel (Ebf1, Dlx1/2) affect thalamic projections subcortically in a manner that seems to be passively carried through in their projections to the cortex, at least at birth [16]. Analysis of the Sema6A mutants at postnatal stages reveals, however, that initially misrouted axons from the dLGN can eventually make appropriate connections to visual cortex. Remarkably, many of these projections seem to occur through alternate routes, either via the external capsule or a superficial route along the outside of the telencephalon. It is also possible that some projections are made via collaterals through the internal capsule that arise at later stages. These findings demonstrate that correct subcortical axonal sorting is not required for eventual projection to a specific cortical area and, further, that the normal temporal sequence of arrival of thalamic axons to the cortex is also not essential for correct targeting. In addition, they show that subcortical sorting is not sufficient to permanently determine connectivity as the initial shift in cortical targeting of VB axons that is apparent at embryonic stages can be corrected after birth. These conclusions are consistent with a growing body of research demonstrating the existence of cortical guidance cues for thalamic axon rearrangements [11,13,26,32] and suggest that the actions of these signals may be effective at a distance [22] to selectively attract misrouted dLGN axons to the appropriate cortical area. They also suggest that guidance cues within the neocortex exist not just in the subplate, but also across the developing cortical layers [26], allowing navigation even in the marginal zone, as demonstrated by the pathway follow by some of the misrouted LGN axons in the Sema6a−/− mutants.
The interpretation that subcortical sorting does not determine final cortical targeting would seem to be challenged by a number of other mutants that show early, global defects in subcortical thalamic projections, accompanied by later, highly specific defects in thalamocortical connectivity. For example, in double Ephrin-A5;EphA4 mutants [30] and in mutants in either CHL1 or Npn1 [20], rostral thalamic axons project more caudally than normally both subcortically and up to the cortex itself at embryonic stages. In both cases, a defect in thalamocortical connectivity is also apparent at postnatal stages, involving excess connectivity of one thalamic nucleus with a particular cortical area, although it is much more selective, and differs between Ephrin-A5;EphA4 and CHL1 or Npn1 mutants. In both cases, the early defect was interpreted as the cause of the later defect, but this has not been shown directly and the selective (and different) nature of the defects at later stages suggests that most of the early misrouting has in fact been corrected and that the postnatal connectivity defects are more likely to reflect later functions of these genes in the cortex itself.
Overall, these studies and our data are thus consistent with a model in which subcortical sorting of thalamic axons is coordinated with eventual cortical targeting, possibly using the same cues at both levels. However, subcortical targeting does not appear to be either strictly necessary or sufficient to determine final connectivity patterns as additional mechanisms exist to restore thalamocortical connectivity to a specific cortical area when alterations during embryonic development occur.
The recovery of the dLGN projection to visual cortex, in spite of previous occupation of this territory by VB axons suggests that dLGN axons have an advantage in the innervation of that particular cortical area. This must be in addition to selective axon guidance to this region as arrival of VB axons to this area is clearly not sufficient to enable them to make permanent connections, at least when faced with competition from later-arriving dLGN axons. A model to explain this would be that dLGN axons and presumptive visual cortex express some matching label(s) that confer this advantage. One candidate for such a cue is the neurotrophin NT-3, which is specifically required for dLGN axons to invade the cortical plate in V1 [33]. NT-3 has been shown to be most strongly expressed in presumptive visual cortex (V1) from around P0 [34], while its receptor TrkC, is selectively expressed by neurons in the dLGN. If such a matched cue is essential then VB axons that at early stages project into the subplate of occipital cortex may not be able to invade the cortical plate, allowing later-arriving dLGN axons to do so. Indeed, if the function of NT-3 in this context shares similarities with trophic signaling [35] then dLGN axons might actively secrete factors that promote withdrawal of VB axons. Axon–axon interactions mediated by surface receptors and cell adhesion molecules [36] might also actively mediate segregation of visual and somatosensory axons [37].
Activity-dependent mechanisms mediating the competitive advantage of dLGN axons for presumptive visual cortex must also be considered, especially as the process takes place during the first few postnatal days, by which time thalamic axons have normally entered into the cortex and formed fully functional synapses [32,38,39]. A number of studies have examined the potential role of electrical activity in areal targeting of thalamic axons. Intracranial infusion of the sodium channel blocker tetrodotoxin (TTX) caused dLGN axons to inappropriately innervate the subplate of cortical areas that they would normally bypass [40]. This could be taken as an instructive role for patterned activity in establishing areal connectivity but could alternatively be explained by an earlier effect of TTX on biochemical signaling pathways downstream of guidance receptors [41], or by feedback onto the expression levels of guidance molecules [42]. This interpretation is more consistent with the known specificity of thalamic axon targeting from the earliest stages [10,11,43] and the lack of effects in areal targeting observed in embryonic SNAP-25 mutants [39,44].
Finally, although our study demonstrates spectacular plasticity of thalamocortical connectivity during early postnatal life, there are some changes in the cortical architecture that persist into adulthood. The reduction in size and change in shape of V1 in Sema6A mutants, which are far more subtle than those observed in enucleation experiments [1,5,45,46], suggest that they may be an interesting model to study some less well-characterized processes, including the separation of the termination zones of primary thalamic axons into discrete areas, the innervation of intervening areas by axons from secondary nuclei, the formation of distinct borders and the hierarchical dependence of secondary and higher-order areas on correct specification of primary areas (reviewed in: [47,48]).
All animal procedures were performed to relevant national and international licensing agreements and in accordance with institutional guidelines. Sema6A mutants were identified in a gene trap screen, as described previously [49]. Insertion of the gene trap vector pGT1PFS into intron 17 results in a fusion of upstream exons of Sema6A with TM-β-galactosidase-neomycin phosphotransferase. This fusion protein is sequestered intracellularly [28]. PLAP is cotranscribed but translated independently from an internal ribosome entry site. No wild-type transcripts are produced from this allele [28].
Brains from E16.5 (n = 18), P0 (n = 45), P4 (n = 8), P7 (n = 6), and P30 (n = 26) were used in the study. To label thalamic and corticofugal fibers, single crystals of 1,1′-dioctadecyl-3,3,3′,3′-tetramethylindocarbocyanine perchlorate (DiI) and 4-(4-(dihexadecylamino) styryl)-N-methylpyridinium iodide (DiA) (Molecular Probes) were placed with a stainless steel electrode into the visual and somatosensory dorsal thalamic nuclei or the visual and somatosensory cerebral cortex of both hemispheres of each brain. After injections, brains were kept in 2% paraformaldehyde for between 3 wk and 2 mo at room temperature in the dark. Back-labeling of dLGN neurons and their axons in P7 animals was performed under hypothermia-induced anesthesia. A small incision was made in the scalp to reveal the skull, and a fine needle was used to pierce the skull above the primary visual cortex. A Hamilton syringe was used to inject 0.5 μl of a 10% solution of DiI in absolute ethanol, into the primary visual cortex. The scalp was bonded with tissue adhesive (Dermabond) and animals were allowed to survive for 24–48 h to allow for adequate retrograde labeling before being sacrificed. Dissected brains were postfixed for 24 h in 4% PFA at 4 °C. Brains were washed in PBS (0.1 M, pH 7.4), embedded into 4% agarose (Sigma), and cut at 100 μm with a Vibroslicer (Leica, VT1000S). Sections were counterstained with 2.5 μg/ml of bis-benzimide (Sigma) or with 0.5 μg/ml DAPI (4′-6-diamidino-2-phenylindole), mounted in PBS/glycerol or AquaPolymount (Polysciences) onto slides, and analyzed using an epifluorescence microscope (Leica, DMR, or Zeiss) and a laser scanning confocal microscope (Leica, DMRE).
Mice were perfused with 4% paraformaldehyde or a mixture of 1% paraformaldehyde/1.5% glutaraldehyde (for the cytochrome oxidase staining) in PBS. Brains were removed, postfixed in the same fixative overnight at 4 °C and embedded in 4% agarose. Serial 50- or 100-μm sections were cut on a vibratome (Leica; VT1000S) and processed for PLAP staining as previously described [28]. Alternatively, following perfusion, brains were postfixed in the same fixative for 3 h and cryoprotected with 30% sucrose in PBS. Serial 40-μm sections were cut on a freezing microtome and processed for Nissl staining (0.5% cresyl violet solution; Sema6A+/−: P0, n = 2; P30, n = 6, and Sema6A−/−: P0, n = 2; P30, n = 6). For cytochrome oxidase staining, cortical hemispheres were dissected from adult mice (Sema6A+/−, n = 9; Sema6A−/−, n = 9), postfixed between glass slides and cryoprotected before sectioning and processing. For immunohistochemistry, dissected brains were postfixed in 4% paraformaldehyde for 24 h, washed in PBS, embedded in 4% agarose, and sectioned (40–60 μm) on a vibratome. P 4 Sema6A+/+ (n = 7) and Sema6A−/− (n = 5) mouse brains were treated for immunofluorescence with rabbit antibody to cleaved caspase-3 (1:200; Cell Signaling Technologies). Similarly, immunofluorescence with mouse antibody to neurofilament (1:100; DHSB) was detected on sections from E16.5 Sema6A+/+ (n = 2) and Sema6A−/− (n = 2) mouse brains. Tangential sections were cut from flattened cortical hemispheres of P7 Sema6A+/+ (n = 5) and Sema6A knockout (KO; n = 10) mouse brains and incubated with antibody to serotonin (1:50,000; Immunostar), which was then detected with biotinylated secondary antibodies using the Elite ABC kit (Vector). Results were documented using a digital camera (Leica DC500; Canon Powershot S40) or an epifluorescence microscope (Zeiss) and digital camera (Olympus), and the images compiled with Adobe Photoshop 8.0 or Adobe Photoshop CS software.
Green and red fluorescent latex microspheres (Lumaflor) were used to labeled axonal projection from somatosensory and visual cortical areas, respectively, to the corresponding thalamic nuclei in Sema6A+/+ (n = 4) and Sema6A−/− (n = 4) mice. Animals were anesthetized with 2.7 mg/kg Hypnovel (Roche), Hypnorm (Janssen), and distilled H2O mixture (1:1:2 volume ratio), which was delivered intraperitoneally, and placed in a stereotaxis frame. After the skin was disinfected and incised, a microdrill was used to perform a craniotomy. Glass micropipettes (Clark Electromedical Instruments) and a binocular stereo-microscope (Zeiss) were used to inject a single injection of 0.3–1.0 μl of CT or microspheres into S1 or V1. Animals were allowed to survive for 24 to 48 h to permit adequate retrograde transport of the CT or microspheres to thalamic cell somata.
In situ hybridization was performed on 50-μm vibratome sections of E14.5 Sema6A (wild-type [WT]: n = 2, Sema6A+/− [HT]: n = 4 and KO: n = 4), P0 Sema6A (HT: n = 9, KO: n = 8), and P7 Sema6A (HT: n = 4, KO: n = 4) mouse brains, as previously described [50]. The following digoxigenin-labeled RNA probes were used: Sema6a (a gift from W. Snider); EphA7, EphrinA5, Cadherin6, and RZRβ (kindly provided by J. Rubenstein, with permission from the original researchers); Cadherin8 (241–1,481 of mouse Cad8; GenBank accession number X95600; obtained by reverse transcription [RT]-PCR).
The number of cells in the dLGN and the VB back-labeled from the occipital cortex were manually counted in consecutive 100-μm sections of E16.5 Sema6A+/− (n = 30 sections, 4 animals) and Sema6A−/− (n = 41 sections, 6 animals), P0 Sema6A+/− (n = 11 sections, 2 animals) and Sema6A−/− (n = 10 sections, 2 animals), and P4 wild-type (n = 9 sections, 2 animals) and Sema6A−/− (n = 12 sections, 2 animals) brains. The numbers of cells back-labeled to either the dLGN or VB for animals of a given age were compared using the Wilcoxon two-sample test and found to be significant at 99.9% confidence limits at E16.5 and P0. As the absolute number of cells back-labeled is dependent on the size of dye crystal used, we also analyzed the proportion of labeled cells in either the dLGN or VB of a given section. The proportional values were Arcsine transformed for statistical analysis by Wilcoxon two-sample tests. The area of the dLGN and vLGN thalamic nuclei was measured in 40-μm cytochrome oxidase serial sections from Sema6A+/− (n = 5) and Sema6A−/− (n = 4) brains using SigmaScan Pro software (SigmaScan). The volume of the dLGN and vLGN was calculated using the Cavalieri method. The relative area of V1 was measured in tangential sections of P7 of Sema6A+/− (n = 5) and Sema6A−/− (n = 10) brains, stained for serotonin immunohistochemistry, using Cell A software (Soft Image System).
|
10.1371/journal.pgen.1006096 | Functional Assessment of Genetic Variants with Outcomes Adapted to Clinical Decision-Making | Understanding the medical effect of an ever-growing number of human variants detected is a long term challenge in genetic counseling. Functional assays, based on in vitro or in vivo evaluations of the variant effects, provide essential information, but they require robust statistical validation, as well as adapted outputs, to be implemented in the clinical decision-making process. Here, we assessed 25 pathogenic and 15 neutral missense variants of the BRCA1 breast/ovarian cancer susceptibility gene in four BRCA1 functional assays. Next, we developed a novel approach that refines the variant ranking in these functional assays. Lastly, we developed a computational system that provides a probabilistic classification of variants, adapted to clinical interpretation. Using this system, the best functional assay exhibits a variant classification accuracy estimated at 93%. Additional theoretical simulations highlight the benefit of this ready-to-use system in the classification of variants after functional assessment, which should facilitate the consideration of functional evidences in the decision-making process after genetic testing. Finally, we demonstrate the versatility of the system with the classification of siRNAs tested for human cell growth inhibition in high throughput screening.
| Human genetics has entered a new age with the advent of next generation sequencing. However, this great advance also comes with new concerns. Currently, the extensive use of multi-gene panels, whole exome and whole genome sequencing, is generating an ever-growing number of new DNA sequence variations detected in the disease-predisposing genes of human patients. The pathogenic or neutral status of these variants needs to be known before planning any medical act or follow-up. We show here that the status of the variants identified in the BRCA1 breast/ovarian cancer susceptibility gene can be estimated thanks to experimental systems using yeast cells and a novel computational model. Importantly, this model provides a probabilistic classification of variants, opening the possibility to integrate results from functional assays into clinical decision-making. Moreover, our computational model is directly compatible with all kinds of experimental system without any requirement for skills in statistics thanks to ready-to-use online tools. We believe that this work is a step forward in the clinical interpretation of human genetic variants.
| Genetic tests, that aim to identify disease-associated germline variants in the genome of patients and relatives, have greatly expanded these last years, together with the number of predisposing genes scrutinized [1]. Genetic tests are proposed by genetic counselors to identify the carriers of genetic variants and to define appropriate clinical follow-ups and treatments for these carriers. The detection of a variant can have severe psychological and physical consequences for the tested patients, depending on whether the variant is known to be pathogenic (associated with disease development), neutral (not related to disease development) or of unknown significance (VUS). Thus, clinical decision-making after genetic testing requires the establishment of reliable variant classifications. The best support is to use methods that attribute a probability of pathogenicity for each variant identified. Because genetic/epidemiological methods, such as co-segregation, case-control, co-occurrence and familial data analyses, provide such probabilities [2], they remain the gold standard in clinical decision-making after genetic testing (see an example in S1 Table). However, genetic/epidemiological methods are time consuming, as they require a substantial amount of observations. Moreover, they are unsuitable for a large number of variants identified, for instance when the number of known carriers is rare. As genetic tests are evolving towards the use of multi-gene panels, whole exome and whole genome sequencing [3], the number of VUS detected is inevitably increasing [1], which stresses the need to improve variant classification [3].
Functional assays have been designed to circumvent the limitations of genetic/epidemiological methods. The generic "functional assay" term refers to in vitro and in vivo systems, able to classify VUS by assessing their influence on protein function or conformation [4]. Functional assays have been widely developed for genes involved in cancers [5] and BRCA1 has become the leading gene analyzed, with 23 different assays proposed, to date [6]. However, despite the genuine interest for strategies that alleviate the limitations of genetic/epidemiological methods, the main challenge of functional assessment remains in its inclusion into clinical decision-making. Indeed, most of the functional assays lack statistical validation [4]. Moreover, analyses are usually based on visually defined cut-offs [6]. Finally, except in rare cases [7,8], the resulting variant classifications lack the probability of pathogenicity provided by genetic/epidemiological methods.
Here, we used experimental as well as computational approaches to overcome these limitations. We evaluated the clinical utility of four different BRCA1 functional assays, designed in yeast cells, by assessing 40 BRCA1 missense mutations, previously classified by genetic/epidemiological methods. To interpret these results, we developed a novel approach, referred to as "Mann-Whitney-Wilcoxon (MWW) method", that defines a non-arbitrary best cut-off value between the neutral and pathogenic variants and that refines variant ranking in data from functional assays. We also developed a computational system that transforms the dual classification between "pathogenic" or "neutral", provided by the non-arbitrary best cut-off, to a probabilistic classification adapted to clinical decision-making. This system of classification, referred to as "probability system", uses the fluctuation of the best cut-off to derive probabilities of pathogenicity for each assessed variant. We show the benefit of our computational model, coupling the MWW method and the probability system, using the experimental data from the four BRCA1 functional assays and using theoretical simulations. We also illustrate that our model is adapted to experimental systems far beyond the genetic variant assessment, with the probabilistic classification of small interfering RNAs (siRNAs) tested for human cell growth inhibition in high throughput screening.
The colony size assay is a BRCA1 functional assay, that has been designed in the yeast model organism, which allows rapid, large-scale and cost-effective variant assessment [9], but has never been subjected to clinical validation yet. In this functional assay, expression of the full length wild type (WT) BRCA1 protein in yeast, induces a growth defect [9–11]. Indeed, after 63 hours of growth on an agarose plate, a single yeast cell gives rise to a colony varying between 5,000 and 21,000 cells (Fig 1A, BRCA1), while colonies reach several millions of cells without protein expression (Fig 1A, Vector control). To ascertain the utility of this assay in clinical medicine, we selected 40 BRCA1 missense mutations, according to their neutral or pathogenic classification by genetic/epidemiological methods (S1 Fig and S2 Table). We confirm that pathogenic missense mutations restore the proliferation rate of yeast cells [10,11]. Indeed, pathogenic mutations have a global tendency to give rise to the biggest colonies, while colony sizes arising from neutral mutations remain close to those of the WT BRCA1 reference (Fig 1A). However, the Colony Size assay does not fully discriminate between pathogenic and neutral mutants. Indeed, variant medians appeared to continuously decrease from M1689R (highest median) to V1804D (lowest median), without clear gap between the pathogenic and neutral regions. Moreover, the neutral M1652T mutation is clearly within the pathogenic sector and the pathogenic R1699W mutation slightly overlaps the neutral region. In such situations, it is critical to have a sound evaluation of the sensitivity, specificity and accuracy of the assay (see the definitions in the S1 Text), which depends on a non-arbitrary and optimal cut-off setting.
The standard method is based on the Youden's index, a classical approach to compute the sensitivity and specificity in a dataset. Using this, the cut-off of 17,910 cells per colony gives the best combined sensitivity and specificity, with 96% (24/25) and 93% (14/15) respectively (Table 1 and S2A Fig). In total, 95% (38/40) of the mutations are correctly classified. The M1652T neutral mutation is misclassified as pathogenic and the pathogenic R1699W mutation is misclassified as neutral (S3 Table). From now on, we refer to "experimental best cut-off", "experimental sensitivity", "experimental specificity" and "experimental accuracy" as the best cut-off, sensitivity, specificity and accuracy obtained from the experimental data.
The disadvantage of the standard method is that mutations are characterized by a single value, here by the median of colony sizes, which can lead to paradoxes in the mutant classification. For instance, the neutral I1858L mutation displays a median of cells per colony higher than the median of the neutral T1720A mutation. Thus, in the mutant ranking, I1858L is closer to the pathogenic group of mutations than T1720A (arrows in Fig 1A). However, T1720A has three values out of nine over the experimental best cut-off, which are thus in the pathogenic area, while I1858L has none (S3A Fig). Therefore, in terms of dispersion range, T1720A could be considered as "more pathogenic" than I1858L. To overcome such paradoxes in variant classification, we developed a nonparametric approach to define the best non-arbitrary cut-off value, that takes into account more information from distributions than the median value alone. This method is based on the MWW test [12–14]. Since the p value of this test provides a quantification of the overlap between two distributions (S4 Fig), we compared each mutant distribution to the WT BRCA1 distribution. The p values obtained defined relative positions of the mutant distributions using the WT BRCA1 distribution as a reference position (Fig 1B and S4 Table). Contrary to the standard method described above, the cut-off used to compute the sensitivity and specificity parameters is a p value. Any mutant with a p value below the p value cut-off, indicates a mutant classified as pathogenic. In contrast, a mutant distribution with a p value over the p value cut-off is considered as neutral. Strikingly, the MWW method solves the paradoxes observed with the standard method, since T1720A is closer to the pathogenic group of mutations than I1858L (arrows in Fig 1B). Moreover, the experimental sensitivity and specificity remains unchanged (Table 1 and S2E Fig). This confirms that the M1652T and R1699W mutations cannot be correctly classified by the Colony Size assay, even when using more information from the experimental data than the variant medians alone. However, it also emphasizes that the variant classification, provided by the MWW method, does not diminish the high sensitivity and specificity of the assay. From this, we conclude that the MWW method is a reliable alternative to the standard method to define a non-arbitrary cut-off in data from functional assessments.
Recently, two-component models have been proposed for the probabilistic classification of variants based on functional assessment. These parametric models require the normal distribution of the neutral and pathogenic values [7,8]. However, as shown in S5A Fig, the Colony Size assay is poorly compatible with such models, due to the bimodal distribution of the pathogenic values. Therefore, we designed an alternative nonparametric and more versatile system of classification. This system is based on the fact that the best cut-off is a random variable that fluctuates, depending on the experimental values. We asked what the variant classification would be, using the Colony Size assay, taking this fluctuation into account. For this, we performed sampling with replacement (bootstrap) of the colony size values, by randomly choosing 9 values among the 9 from each mutant, and 36 values among the 36 from the BRCA1 reference control. Next, using this new set of sampled data, we applied the standard or MWW method to obtain the best cut-off. We repeated this procedure a large number of times, which allowed us to define a best cut-off distribution for the standard and MWW methods (S5 Table). We also used a third method, referred to as "standard with reference method". It is similar to the standard method, except that the best cut-off distribution obtained includes the fluctuation of the WT BRCA1 reference, as explained in the S1 Text. Notably, the standard with reference method allows an additional comparison with the MWW method, which also includes the fluctuation of the WT BRCA1 reference. Finally, we designed the probability system of classification. This system allows to assign a probability of pathogenicity to each assessed variant, using the best cut-off fluctuation (Fig 2A and S6 Fig). The rationale is that the farther a variant is from the core of the best cut-off fluctuation, the more robust is its classification as either pathogenic or neutral. A probability close to 1 indicates that the variant can be classified as pathogenic, with a low risk of misclassification as neutral due to the fluctuation of the best cut-off. A probability close to 0 indicates that the variant can be classified as neutral, with a low risk of misclassification as pathogenic due to the fluctuation of the best cut-off. Finally, a probability of 0.5 designates no preferential classification as either neutral or pathogenic (variant completely unknown). With such probabilities, the five-class nomenclature proposed by Plon et al [26]. (S1 Table) can be directly applied to functional assays. Probabilities obtained for the Colony Size assay are shown in Fig 2B. Strikingly, a level of uncertainty was generated, notably with variants classified as "uncertain" (class 3). This highlights the critical influence of the best cut-off fluctuation in variant classification. In addition, the MWW method exhibits the best accuracy, with 37/40 mutations correctly classified versus 36/40 for the standard and standard with reference methods. When including the number of misclassified mutations, the MWW method shows a balance of 35 mutations, ex-aequo with the two other methods (S6 Table). Altogether, these results confirm the possibility to use the MWW method in variant classification. In addition, the probability system seems to be an effective and simple way to obtain a probabilistic classification of variants in functional assessment.
We validated three other functional assays, by assessing the same 40 mutations used in the Colony Size assay. The Liquid Medium assay monitors the growth defect of yeast cells expressing BRCA1 (S7 and S8 Figs), as in the Colony Size assay, but in liquid instead of solid medium [11]. The Spot Formation assay is derived from the observation that the BRCA1-mCherry fusion protein accumulates in a single aggregate in the nucleus of yeast cells. This aggregate is referred to as "spot" due to its visual signature using fluorescent microscopy. We previously showed that pathogenic missense mutations decrease the proportion of cells showing one spot [11]. Here, we confirmed this effect (S9 and S10 Figs). The last assay tested was the Yeast Localization assay. Whereas cytoplasmic spots are rare in yeast cells expressing the WT BRCA1 protein, this event has a tendency to increase in the presence of pathogenic mutations [11]. Here, we confirmed this effect (S11 and S12 Figs). However, albeit promising, none of these three assays provided a better discrimination than the Colony Size assay, to distinguish between pathogenic and neutral variants. This was notably shown by the experimental sensitivity and specificity computed (Table 1 and S3 Table).
We took advantage of the experimental differences among the four assays (recapitulated in S7 Table) to detect potential flaws in the MWW method. In contrast, the MWW method constantly overcomes the incoherent ranking generated by the standard method (see examples in S7, S9 and S11 Figs). This is achieved without reducing the experimental accuracy compared to the standard method (Table 1), except for a minor decrease in the Liquid Medium assay (88% versus 90%). Also, no flaws were detected in the probability system, which would result in an unexpected high level of misclassifications (Fig 2B). Interestingly, accuracy of the MWW method is globally better than in the standard or standard with reference method, with the best accuracy of 93% in the Colony Size assay, 83% in the Spot formation assay, and with the best ex-aequo accuracy of 73% in the Yeast Localization assay (Fig 2B). Variant misclassification was slightly higher in the MWW method, compared to the two other methods, with one more misclassification in the Colony Size and in the Spot Formation assays, one less in the Liquid Medium assay, and ex-aequo in the Yeast Localization assay ("Total number of variants misclassified" column in S6 Table), even if the balance between accuracy and misclassification maintains the MWW method as the best one, ex-aequo with the standard method ("Balance" column in S6 Table). Finally, contrary to the MWW method, the standard method suffers from a lack of sensitivity in the Yeast Localization assay, since none of the pathogenic mutations are classified as class 5 (Fig 2B). Overall, the analysis of four functional assays did not reveal any major flaw in the probability system of classification. In addition, the results obtained with the MWW method confirm the possibility to classify variants using more information from the variant distribution than the median value alone.
To complete the detection of potential flaws in our classification model, we analyzed theoretical situations. A reference situation was designed, similar to that in the Colony Size assay (S8 Table). Next, different parameters were scrutinized: the position of the pathogenic mutations (S13 Fig), neutral mutations (S14 Fig), or WT BRCA1 reference (S15 Fig), the initial sensitivity and specificity of the assay before using the probability system (S16 Fig), the number of neutral and pathogenic variants used (S17 Fig), the number of values in the variants and in the WT reference distributions (S18 Fig), and the range of the variant and WT reference distributions (S19 Fig). Results are recapitulated in S9 Table and summarized in Table 2. The standard with reference method shows strong usage limitations, notably when the WT reference exhibits a negative median or a median close to zero (Table 2 and S15E Fig, middle panel). Interestingly, the MWW method is not affected by such situations. The main limitation detected is an extreme situation in which the WT reference distribution falls outside of the range of the neutral and pathogenic distributions (e.g., S15A Fig, left panel), which impairs the sensitivity of the probability system of classification (Table 2 and S15E Fig, right panel). Except for this extreme situation, we confirm the efficient behavior of our classification model, coupling the MWW method and the probability system: (1) when the pathogenic and neutral distributions are strictly identical, all the mutations are classified as class 3 (Table 2 and S13D Fig, right panel), (2) the sensitivity and specificity of the probability system of classification increase when pathogenic mutations move away from the WT BRCA1 reference distribution (S13D Fig, right panel), and (3) when pathogenic mutations are contaminated by neutral mutations (experimental specificity reduced), the sensitivity of the probability system of classification is decreased (Table 2 and S16E–S16G Fig, right panel), and vice versa. This last result is an important criterion for classification, since unknown mutants that would be located in a pathogenic region containing neutral mutations, could not be formally classified as pathogenic. Therefore, it is noteworthy that the experimental sensitivity and specificity values are taken into account by our classification model. Interestingly, the model is poorly sensitive to the number of neutral or pathogenic mutations used to validate a given assay (S17E–S17G Fig, right panel), as long as the number of values in the dataset is high enough (S18E–S18G Fig, right panel). Supplemental information is provided in the S1 Text. This notably includes an extensive analysis of the best cut-off fluctuation, which explains the lack of sensitivity of the standard method, mentioned above in the Yeast localization assay (Fig 2B) and also shown in theoretical situations (see the legend of S13 Fig). It also contains specific procedures for variant classification (e.g., Bayesian inference, combination of functional results, assessment of VUS), as well as procedures to fit the proposed model to other situations. It finally includes the ProClass toolbox that generates the probabilistic classification of variants, adapted to most kind of functional assays.
We wondered if the classification model developed for genetic variants could be easily extended to other decision-making situations. The analysis of theoretical situations showed that variant classification remains accurate when only one neutral and one pathogenic variant are available (S17E–S17G Fig). This indicates that the fluctuation of the best cut-off supports decision-making in situations represented by a limited number of positive and negative controls. To confirm this, we analyzed data from 406 genes targeted by small interfering RNAs (siRNAs), screened for their capability to inhibit the proliferation of a human prostate tumoral cell line (Fig 3). The "No siRNA" control, the siKIF11 positive control and the siGOLGA2 and siGL2 negative controls were treated as WT reference, pathogenic and neutral variants, respectively. Structure of the data is reported in S7 Table. As in the BRCA1 functional assays, the experimental accuracy was not impaired using the MWW method, compared to the standard method (Table 1). In addition, no flaws were detected in the probability system, since the accuracy remained at 1, whatever the standard, standard with reference or MWW method used (Fig 3C). Finally, the advantage of the MWW method is again highlighted in the final classification of the screened siRNAs. Indeed, in the siRNA ranking, based on the median values, siGTSE1 is closer to the negative controls than siITGA2 (Fig 3A). By taking the distribution of these two siRNAs into account, the MWW method switches their ranking position (Fig 3B), so that siGTSE1 is finally classified as "unclear effect on cell growth inhibition" (Fig 3C, MWW method), instead of "no cell growth inhibition" (Fig 3C, standard and standard with reference methods). Thus, this demonstrates that our probabilistic model is also adapted to the classification of experimental data far beyond the functional assessment of genetic variants.
We provide the statistical validation of four BRCA1 functional assays, as well as a classification model that facilitates the incorporation of functional assay results into clinical decision-making. The probabilistic model is based on the fluctuation of the best cut-off, which is driven by the fluctuation of the experimental data. Thus, the variant classification provided reflects the robustness of a cut-off-based decision-making towards data fluctuation. The model has the advantage to be nonparametric, easy to handle and easy to adapt to most kind of functional assays. Moreover, among the variants incorporated in functional assays, the model only depends on those previously classified by genetic/epidemiological methods as pathogenic or neutral. It is not influenced by unknown variants, meaning that the subsequent incorporation of unknown variants in a functional assay does not require a new analysis of the best cut-off fluctuation. These features of our model contrast with parametric models, proposed for variant classification [7,8].
We achieved a widespread analysis of the best cut-off fluctuation dedicated to decision-making (completed in the S1 Text). This analysis is focused on the classification of genetic variants, but it is also valid for other decision-making situations compatible with our classification model, such as high throughput siRNA screenings. Using many different kinds of data structures (four BRCA1 functional assays, one siRNA screen and 93 theoretical situations), three different methods of best cut-off fluctuation were scrutinized: the standard, the standard with reference and the MWW methods. From this study, we conclude that the standard with reference method is poorly compatible with a versatile classification model, due to important lacks of accuracy when the WT reference exhibits a negative median or a median close to zero (S15E Fig, middle panel). The standard method has the advantage to support decision-making in experimental situations devoid of a WT reference. The MWW method has the advantage to use more information from the distribution of the classified elements than the median value alone. This refines the ranking and the final probabilistic classification. Contrary to the standard method, the MWW method is adapted to experimental situations in which the neutral and pathogenic variants (or the negatives and positives controls) are represented by a single value if the WT reference encompasses a significant number of different values (S18E Fig, compare the left and right panels for the mutant with one value). However, the MWW method is poorly adapted to experimental situations where the WT reference distribution is more or less outside of the range of the neutral and pathogenic distributions (e.g., S15A Fig, left panel). Thus, we propose to prioritize the MWW method if the data are compatible with this method, notably if the WT reference is well embedded in the neutral/negative distributions, or if the WT reference is between the neutral/negative and the pathogenic/positive distributions, and to use the standard method otherwise. The different methods are proposed in the ProClass toolbox available online (see the S1 Text).
Interestingly, none of the four yeast assays is able to correctly classify the R1699W pathogenic and the K45Q neutral variants. Pathogenicity of R1699W has been long-established in independent studies, using different genetic/epidemiological methods [15–17], confirming that yeast cells are unable to detect the deleterious impact of R1699W [18]. This emphasizes that the mechanism of R1699W, leading to tumor development, is different from the other pathogenic missense variants of BRCA1. It is probably related to a protein-binding defect without major BRCA1 structural destabilization [19]. The classification of K45Q has been established by a single epidemiological study [20], with little evidences of neutrality (e.g., probability of being pathogenic of 11% by family history prediction). However, the absence of any functional impact has been confirmed in three different functional assays using mammalian cells [21,22], which stresses a specific effect of K45Q in yeast cells, that remains to be explained.
Finally, this work showed that the yeast organism can be used to classify variants positioned in both Nter and Cter parts of BRCA1. Among the four assays analyzed, the Colony Size assay is the most accurate (93%) and the most robust to data fluctuation (one class 3 variant). The Liquid Medium and Yeast Localization assays may also be attractive for diagnosis due to the absence of false negative results detected, notably when using the MWW method. Interestingly, the Yeast Localization assay allows the identification of pathogenic variants that delocalize the BRCA1 protein into the cytoplasm. If confirmed in human cells, this assay could define subcategories in the pathogenic variants of BRCA1, based on different cellular mechanisms leading to tumor development.
All plasmids are derived from pJL48 [11], a modified version of pESC-URA (Agilent Technologies), in which the MYC epitope has been removed by SalI-XhoI digestion and vector ligation. In this plasmid expression of the cDNA is controlled by the GAL1 promoter, inducible by galactose and repressed by glucose. The backbone of the human BRCA1 (MIM# 113705) cDNA used, corresponds to the AY888184.1 GenBank sequence with a TGA stop codon instead of TAG. To facilitate the cloning of BRCA1 missense mutations, silent mutations were inserted in the cDNA to generate 4 new restriction sites: SalI (c.1020A>C + c.1023T>C), AvrII (c.4662A>T), FseI (c.4839T>G + c.4842A>G + c.4845T>C) and XhoI (c.5502C>T). Of note, the WT BRCA1 and BRCA1-mCherry plasmids used in this study (pPT60 and pPT63 respectively, see S10 Table) are different from the pJL45 and pGM40 plasmids, used in our previous publication [11], by the addition of the 4 restriction sites. The 40 missense mutations were generated by targeted mutagenesis (Genscript Company, Piscataway, NJ, USA) on intermediate plasmids. Next, we inserted the mutated cDNA fragment into the pPT60 and pPT63 plasmids by a single digestion—ligation step. All resulting plasmid constructs were verified by sequencing the promoter, the full cDNA and the terminator.
Transformation of the Saccharomyces cerevisiae haploid BY4741 or YKR082W-GFP strains were performed as previously described [11]. The strains generated are referenced in S11 Table. To facilitate the description, we referred to the cells transformed with pESC-URA as the name of the cDNA inserted into the plasmid. Thus, "BRCA1" refers to yeast cells transformed with the plasmid containing the WT BRCA1 cDNA; "M18T" refers to yeast cells transformed with the plasmid containing the M18T mutated version of the BRCA1 cDNA; and "vector" refers to yeast cells transformed with the same plasmid without inserted cDNA. Three independent transformants per strain, also referred to as "clones", were selected after each transformation. We observed that lithium acetate transformation can result in diploidisation of haploid cells. To control this, the ploidy of each clone was verified by FACS analysis, using the yeast strain BY4741 (haploid) and BY4743 (diploid) as a control. Next, in the different assays, cells were grown in glycerol-lactate medium (GL-URA) as previously described [11]. Addition of galactose in the medium (GAL) induced the expression of BRCA1, while addition of glucose (GLU) strongly repressed the expression of BRCA1.
Cells were grown in log phase in YPD medium (1% yeast extract, 2% peptone, 2% dextrose, 60 μM Adenine, 8 μM NaOH). 107 cells were collected and put at 4°C to block the cell cycle. Cells were centrifuged at 4°C and resuspended in 70% ethanol. After 1 hour incubation at room temperature (RT), cells were centrifuged and resuspended in freshly made sodium citrate [50 mM] pH7. Sonication was performed to dissociate cell aggregates (Vibracell and probe CV33 (Bioblock Scientific, Illkirch, France), pulse 30%, time 15 seconds). Cells were centrifuged and resuspended in sodium citrate [50 mM] pH7 + RNAse A [0.25 mg/ml]. After 1 hour incubation at 50°C, cells were centrifuged, resuspended in sodium citrate [50 mM] pH7 + Propidium Iodide [16 μg/ml] and analyzed using an Accuri (BD Bioscience, San Jose, CA, USA).
This assay was previously named "small colony phenotype" (SCP) assay. The method already published [11] was slightly improved as follows: (1) GL-URA+galactose and GL-URA+glucose plates were incubated 63 hours and 50 hours respectively (instead of 52 hours), and (2) the biggest colony of each plate, representing the size of at least five other colonies on the plate, was chosen for cell counting. This prevents the choice of rare but extremely big colonies (outliers). For the simultaneous assessment of 10 variants by a single technician, the time required between the delivery of the intermediate plasmids (see above) and the final results is 20 days.
This method already published [11] was slightly improved as follows: during glucose induction, cells were diluted at 0.5×106 cells/ml (instead of 106 cells/ml) for the 15 hour culture time at 30°C. Galactose induction conditions remained as before. For the simultaneous assessment of 10 variants by a single technician, the time required between the delivery of the intermediate plasmids (see above) and the final results is 20 days.
This method already described [11] was slightly improved as follows. Briefly, Nup133-GFP cells, expressing the WT or mutated BRCA1 protein fused to mCherry, were induced for 4 hours with galactose before analysis using live fluorescent microscopy. The previously named "yeast localization phenotype" (YLP) assay [11] was subdivided into two assays in this study. The Spot Formation assay monitors the proportion of cells showing a single aggregate of WT or mutated BRCA1, visible in fluorescent microscopy, without considering the intracellular localization. This aggregate is also referred to as "spot". Cells with several aggregates were not considered in this assay. The Yeast Localization assay monitors the proportion of spot volumes localized in the cytoplasm of yeast cells. Picture acquisitions were previously described [11]. For each clone, at least three fields, containing at least 100 cells, were acquired. For the Spot Formation assay, the number of cells showing one spot was manually counted. Next, the proportion of cells containing one spot was computed by dividing the number of cells showing one spot to the total number of cells (one value per clone). For the Yeast Localization assay, images of the three fields were deconvoluted [23] and the volume Volij of each spot i, in the field j, was measured using the 3D Object Counter plugin [24] of ImageJ. Next, each spot was manually categorized as "inside" or "outside" the nucleus. Finally, the proportion of volume outside the nucleus was computed using the formula (ΣiΣjVolij/outside) / (ΣiΣjVolij/outside + ΣiΣjVolij/inside), which led to one value per each clone assessed. This proportion quantifies the cytoplasmic localization of the mCherry protein fused to BRCA1. For the simultaneous assessment of 10 variants by a single technician, using the Spot Formation and Yeast Localization assays, the time required between the delivery of the intermediate plasmids (see above) and the final results is 21 days.
IGR-CaP1 epithelial cells, derived from a human prostate primary tumor [25], were plated in 384-well plates at 750 cells/well, were allowed to adhere overnight and then were transfected with a single siRNA from a siRNA library targeting 406 different genes. siKIF11, siGL2 and siGOLGA2 were used as controls. After 72 hours, cells were fixed and nuclei were stained with DAPI. Images were acquired with an INCell 2000 automated wide-field system (GE Healthcare, Little Chalfont, UK) and cell counts were quantified in each well with the INCell Analyzer workstation software (GE Healthcare). The pictures analyzed represent 20% of the well surface, which corresponds to an average of 150 cells initially plated for this surface.
Statistical and computational methods, as well as R source codes, are provided in the S1 Text.
|
10.1371/journal.pcbi.1002229 | Conformation Regulation of the X Chromosome Inactivation Center: A Model | X-Chromosome Inactivation (XCI) is the process whereby one, randomly chosen X becomes transcriptionally silenced in female cells. XCI is governed by the Xic, a locus on the X encompassing an array of genes which interact with each other and with key molecular factors. The mechanism, though, establishing the fate of the X's, and the corresponding alternative modifications of the Xic architecture, is still mysterious. In this study, by use of computer simulations, we explore the scenario where chromatin conformations emerge from its interaction with diffusing molecular factors. Our aim is to understand the physical mechanisms whereby stable, non-random conformations are established on the Xic's, how complex architectural changes are reliably regulated, and how they lead to opposite structures on the two alleles. In particular, comparison against current experimental data indicates that a few key cis-regulatory regions orchestrate the organization of the Xic, and that two major molecular regulators are involved.
| In mammal female cells X-Chromosome Inactivation (XCI) is the vital process whereby one X, randomly chosen, is silenced to compensate dosage of X products with respect to males. XCI is governed by a region on the X, the X Inactivation Centre (Xic), which undergoes a sequence of conformational modifications during the process. The two Xic are exposed, though, to the same environment, and it is obscure how they attain different architectures. By use of computer simulations of a molecular model, here we individuate general physical mechanisms whereby random Brownian molecules can assemble chromatin stable architectures, reliably regulate conformational changes, and establish opposite transformations on identical alleles. In the case-study of the murine Xic, our analysis highlights the existence of a few key regulatory regions and molecular factors. It also predicts, e.g., the effects of genetic modifications in the locus, which are compared with current deletion/insertion experiments. The physical mechanisms we describe are rooted in thermodynamics and could be relevant well beyond XCI.
| X-Chromosome Inactivation (XCI) is the vital process occurring in female mammalian cells whereby one randomly selected X is transcriptionally silenced to balance dosage with respect to males [1]–[4]. XCI is regulated by a region on the X chromosome, the X inactivation center (), which encompasses a key group of neighboring non-coding genes (see Fig. 1.A) including, e.g., , , and [1]–[4]. The fate of the X is determined by its gene which is strongly upregulated on the future inactive X and repressed on the other X. In turn, is negatively regulated by , and positively regulated by , , and other factors [5]–[7].
Before random XCI starts, a complex epigenetic program, coupling transcription and chromatin remodelling [8], [9] to pluripotency factors [10]–[12], produces a state where the has the same spatial conformation on the two X chromosomes [13] and both alleles are just weakly active. Upon XCI, an unknown symmetry breaking mechanism determines the opposite behaviour of the two , and induces alternative modifications of the three-dimensional conformation of their [13], [14]. Finally, on the designated inactive X further chromatin reorganizations occur as a heterochromatic compartment forms into which genes are recruited to be silenced [3], [15]. Several molecular factors are known to be involved in the process [3], [4], including noncoding transcripts, chromatin modifiers and organizers, such as CTCF (a Zn finger having arrays of binding sites on the Xic), Dnmt3a, Oct4 and other pluripotency factors [9]–[12], [16], [17]. Different models have been proposed to describe random XCI [18]–[22], but still none to elucidate its associated chromatin changes, whose nature remains mysterious.
To understand the principles of chromatin organization, within the murine case study, here we explore the scenario where chromatin conformations emerge from its interaction with diffusing molecular factors. We discuss general physical mechanisms whereby random Brownian molecules can: i) succeed in establishing stable, non random conformations on the chromosomes; ii) reliably regulate specific conformational changes; and iii) produce opposite transformations on identical alleles exposed to the same environment (“symmetry breaking”). We investigate by computer simulations a schematic model consisting of two identical polymers which interact with a concentration of diffusing molecules (see Fig. 1.B). In the light of current 3C data [13], the model poses that along each polymer three types of regions exist type-, and ) and predicts the existence of two types of regulatory molecules (type-A and B).
We show that the system thermodynamic stable states fall in distinct classes corresponding to different conformations. The polymers spontaneously select one of them according to molecule concentration/binding energy. Conformational changes are driven by thermodynamic phase transitions which act switch-like, regulated by given concentration/binding energy thresholds. The two polymers are exposed to the same environment, yet they can undergo alternative architectural modifications: we show that a symmetry breaking mechanisms is activated if the homotypic interaction between regulatory molecules rises above a threshold.
Comparison to experimental observations [1]–[5], [13], [21] suggests that the regions envisaged by the model can be approximately mapped along the sequence as illustrated in Fig. 1.B, while type-A and B complexes could be related to an activating and a blocking regulator of .
We represent the relevant region of each X chromosome (see scheme in Fig. 1.B) by a standard model of polymer physics, a self-avoiding bead chain [23]. In the light of current 3C data [13], we pose that along each polymer there are, for simplicity, two type- regions which have an array of binding sites for type-A Brownian molecular factors. Each polymer has also two type- regions with binding sites for a different kind of molecular factors (type-B). Finally, the polymers have a type- region whose binding sites can be bound by either type-A or B molecules. Thus, type-A molecules (resp. type-B) can bridge a type- (resp. type-) and a type- site. For simplicity, with no loss of generality, we consider the case where the two types of molecules have the same concentration, , and the same affinity, , for all binding regions. Similarly, we assume that type- and type- regions have the same number of binding sites, , than type-. The value of is fixed to have a total binding site number of the order of known binding molecules. As CTCF is a general chromatin organizer which has been associated to XCI and its binding sites have been well characterized [17], we use it as an example (and set ). For simplicity, is here also the length of the intervening inert sequences between them. Type-A (resp. type-B) molecules can bind, with multiple valency, each other with affinity (resp. ); we set and, considering the number of binding domains of CTCF, the valency to four.
We investigate by Monte Carlo (MC) simulations the conformations of the system as they spontaneously emerge when the three control parameters, , are varied. For computational purposes, the system lives in a cubic lattice with a lattice spacing , whose value corresponds to the typical size of a DNA binding site, and can be roughly estimated to be . The volume concentration of molecules in our model, , can be related to molar concentrations : , being the Avogadro number (details in Text S1). Thus, for instance, a typical nuclear protein concentration of would correspond to . Below we consider concentrations in the range and binding energies in the weak biochemical scale (a few units in ). Finally, conversion of MC time unit to real time is obtained by imposing that the diffusion constant of our polymers is of the order of measured chromatin diffusion constants (see Text S1 for details).
We first show that diffusing molecules can produce a looped conformation on each polymer where type- and type- stably interact with type- region. The process is based on a thermodynamic mechanism (a phase transition, in the thermodynamic limit) which acts switch-like when concentration/affinity of binding molecules rise above a threshold [24].
Before describing our MC results in details, we illustrate the underlying mechanisms. A single, say, type A molecule forms a bridge between type- and type- regions via the stochastic double encounter of the molecule with its binding sites. This is, though, an unlikely event, especially if molecule concentration, (or , see below), is small. And the half-life of such a bridge is short when weak biochemical interactions are considered. Thus, on average the regions float away from each other (see pictorial representation in the bottom panel of Fig. 2, “Open State”). At higher (or ), however, many a molecule can bind type-/ regions and stabilize the conformation via a positive feedback mechanism as their bridges reinforce each other and facilitate the formation of additional bridges. The concentration where such a positive feedback mechanism starts winning marks the threshold above which stable contacts are established (pictorial representation in the bottom panel of Fig. 2, “Stable Interaction”).
This pictorial scenario summarizes our MC results. For sake of simplicity, we consider first the case where molecule mutual interaction is turned off, , and set as initial configuration of the polymers a randomly open conformation. We measure the interaction order parameter, , where (resp. ) is the probability to have, on a polymer, a contact of a type- (resp. type-) with type- region. If neither type- nor type- regions are in contact with , the order parameter is zero, ; if only one pair is stably interacting then ; finally, if both type- and type- loops are established. Fig. 2 top panel shows the MC time evolution of for two values of : if is small, remains indefinitely close to zero, , as no stable contact is statistically possible; instead, if is high enough, grows to a value close to one, , showing that both the type- and loops are formed.
In the space of the control parameters, , a sharp line separates the two regimes, as shown in Fig. 2 bottom panel: when or are small, contacts cannot be stable and ; conversely, above the transition line the two loops conformation is reliably established on each polymer, and . Such a line marks the boundary between two thermodynamic phases [25]: it corresponds to the point where the entropy loss due to loop formation is compensated by the energy gain obtained from the establishment of the corresponding bridges.
The discovery of such a switch-like behaviour can also explain how loop formation can be sharply and reliably regulated in the cell by increasing the concentration of specific molecular mediators or the affinity to their DNA target sites, e.g., by chromatin or molecule modifications.
The position of the transition line is also dependent on the number of available binding sites, , since, schematically, the overall binding energy scale is . Thus, non-linear threshold effects in genetic deletion/insertions of the locus exist.
From Monte Carlo results we can predict concentration (or energy) thresholds in real nuclei. For instance, in vitro measures of CTCF DNA binding energies give , a typical value for TFs [26], [27]: an extrapolation from Fig. 2 then predicts a threshold , corresponding to a typical nuclear protein molar concentration (see Text S1).
Finally, the mechanism leading to stable loop formation has to be fast enough to serve functional purposes. In our model we find that stable interactions are established on scales of the order of minutes (see Fig. 2 top panel and Text S1), a range consistent with biological expectations.
The mechanism to induce conformational changes illustrated above acts “symmetrically” on the two polymers. Now we show that molecule homotypic interaction, , can break the polymer symmetry via a different thermodynamic mechanism. More precisely, if (and , see below) is above a critical threshold, a single major aggregate of type A molecules and a single one of type B are formed because of homotypic binding cooperativity: in facts, the energy gain in forming a single cluster of A/B molecules (which maximizes the number of possible chemical bonds) compensates, if is large enough, the corresponding entropy reduction. The single, say, type A aggregate will then randomly bind just one polymer, leaving the other one “naked” (pictorial representation in the bottom panel of Fig. 3, “Symmetry Breaking”).
Type-A and B aggregates bind opposite polymers because A and B molecules compete for binding sites in the type- region. Hence, if a fluctuation increases the presence of, say, A molecules on one polymer, cooperativity tends to favor their assembling at that site and B molecules are expelled; in turn, the depletion of A around the other polymer favors the assembling of B molecules on it. On the polymer where the A cluster binds the type- region, the B-related loci can no longer be stably linked, and their loop opens; the opposite situation happens on the other polymer.
The above scenario results from our MC simulations. We measured the symmetry breaking order parameter, , where is the average local concentration of molecules around the type- region of polymer . The parameter is close to zero if an equal amount of A molecules is present around the two polymers, whereas it approaches one if the symmetry is spontaneously broken ( and behave analogously). Fig. 3 top panel shows the time evolution of from an initial configuration corresponding to the symmetric state (schematic picture in the bottom panel of Fig. 3, “Stable Interaction”) where each polymer has two stable loops as seen before: if is small, remains close to zero at all times and the system remains in a symmetric state; conversely, if is high enough, approaches one because A molecules reside mostly around just one, randomly chosen polymer and the symmetry is broken (schematic picture in the bottom panel of Fig. 3, “Symmetry Breaking”). The phase diagram of Fig. 3 bottom panel shows that the symmetry breaking mechanism is switch-like too: in the space, as soon as a narrow transition line is crossed the system switches from a symmetrical polymer state to a broken polymer symmetry state. More details are in the Text S1.
For sake of simplicity, we considered the case where the concentration/DNA affinity of molecules A and B are the same. However, such an assumption does not affect our general results. The only condition for the Symmetry Breaking and Configurational Switch mechanisms to be triggered is that concentration/interaction energy of both types of molecules rise above the appropriate threshold.
As far as XCI is concerned, the predicted single B molecule aggregate is interpreted as an repressing factor (a Blocking Factor, BF) and designates the future active X. The A aggregate marks the X where transcription is enhanced and is interpreted as an activating factor (AF). Importantly, the thresholds predicted by our theory for the symmetry breaking mechanisms also fall in the correct biochemical range (see above and Fig. 3 bottom panel).
The time scale required to break the symmetry in a real nucleus can depend on a number of details. Our MC provides, thus, only a very rough order of magnitude estimate. As shown in Fig. 3 top panel, such a time scale is predicted to be around 10 hours, a value of the order of the time required for XCI initiation.
In males other processes could intervene, yet it is easy to see how the same two factors mechanism can work, i.e., why the only X is usually bound by the B aggregate (and not by A) to repress . In fact, the affinities of A and B molecules for the type- region are expected, in general, to be different: . Hence, if is larger than , it is thermodynamically convenient that B molecules bind the X, a difference of a few units in being sufficient to skew of orders of magnitudes the binding probability of A and B.
Finally, variants of the model can be considered to account for further biological details. For instance, additional molecular factors, or the effects on polymer colocalization can be discussed (see Text S1), but no relevant changes to the present scenario are found.
Our schematic model (Fig. 1.B) predicts that two kinds of molecular regulators, type-A and B molecules, interact with a set of specific regions along the polymers. Current 3C data [13] suggest that our type- and type- regions map respectively in the area 5′ and 3′ to , while the type- region is in between.
We showed that in our model only three classes of stable conformational states exist (see Fig. 4 A,B,C). The system spontaneously falls in one of them, according to molecule concentration and homotypic interaction, and . State changes are regulated by a “conformation” and by a “symmetry breaking” switch, related to two distinct thermodynamic phase transitions [25]. The switches are controlled by changing and above/below specific threshold values. Their on/off nature can explain how a sharp regulation of nuclear architecture and stochastic choice of fate can be reliably obtained by simple strategies, such as protein upregulation or chromatin modification. Importantly, the model predicts energy/concentration thresholds which are in the expected biological range (weak biochemical energies, fractions of concentrations).
We now discuss how the present scenario can recapitulate in a unified framework important experimental results on XCI.
Before XCI, the conformation is found to be identical on the two X's [13]: and genes are looped onto a “buffer” region; similarly, , and the “buffer” form a second hub with . Upon XCI, on the future active X, the --buffer hub opens while remains in contact with . On the other X, instead, the - interactions is lost whereas and remain in contact.
Our model rationalizes how those elements are sharply regulated to recognize each other and to form stable interactions based on weak biochemical bonds. It can also explain how the same physical elements later at XCI spontaneously break the X symmetry. The molecular aggregate bound, in our model, to the type- regions (which should encompass the - area) is interpreted as a factor related to silencing (i.e., to its Blocking Factor, BF [1], [2], [4]) and designates the future active X; the different aggregate bound to type- regions, encompassing the area of the other X would be linked to an activating factor (AF) [5], [6], [22]. The link between architectural changes and choice of fate emerges here naturally.
During XCI establishment, the inactive X undergoes further architectural reorganization [3], [14], [15]. The mechanistic details of those conformational changes are still not understood, but they could involve mechanisms as those illustrated here.
Other interesting models have been proposed for “counting&choice” at XCI, but still none had focused on the spatial organization, including our original Symmetry Breaking theory [19]. In the approach of ref. [21], each X chromosome is assumed to have an independent probability to initiate inactivation. Two competing factors exist: an X-linked XCI-activator and an XCI-inhibitor produced by autosomes. In a male XY cell the XCI-activator concentration is too low to initiate the inactivation of the only X; in female XX cells the initial XCI-activator concentration is, instead, above the threshold needed to start XCI. As soon as one X is inactivated, the XCI-activator concentration falls down to the levels found in males, and thus the other X remains active. A different model [22] poses that two types of sites are present on the X: “XCI-init” which is responsible for the initiation of inactivation of the X bearing it, and “XCI-repres” sites which inhibit the action of “XCI-init”. Each active X produces molecules, say molecules, which bind to some autosomal sites. If these sites are saturated, the autosomes produce a set of molecules , which, with a “Symmetry Breaking” mechanism [19], self-assemble into a single molecular factor and inhibit the activity of “XCI-repres” sites on one of the two X, determining its inactivation. As the availability of the signal is reduced, it is no longer sufficient to saturate the autosomal receptors, and the remaining X remains active.
The mechanisms for conformational changes we discussed here are rooted in thermodynamics and are, thus, very robust to difference in molecular details. They could apply then to all the mentioned models for “counting&choice”. An interesting question concerns the applicability of those models to mammals other than mice. Important differences have emerged, for instance, between human and mice XCI [28], [29]. As stated above, the mechanisms we discussed for architecture in mice stem cells are very robust, yet data on other organisms are still too scarce to decide whether such mechanisms might apply elsewhere.
The phenotype of key deletions along the (see reviews in [1], [2], [4], [19], [22] and ref.s therein) can be explained by our model. The deletion [30] removes encompassing and part of /. In heterozygous females the deleted X is always inactivated. In males it leads to the inactivation of the only X; the shorter the deletion considered within the (see , , , [31]), the smaller the fraction of ectopic X inactivations in a population.
Those deletions, in our model, map into sites where the “blocking factor” (BF) binds (and blocks inactivation of that X): removes a large portion of binding sites, thus the deleted X has a strongly reduced affinity for the BF (w.r.t. the wild type X) which does not bind there; the shorter the deletion, the weaker the effect. So, in heterozygously deleted females a skewed random XCI occurs, whereas in males the only X can be inactivated. These deletions can also impact the formation of the BF itself because the involved regions possibly encode some of its components.
Heterozygous [32] and [33] deletions in females also result in the inactivation of the deleted X. Their homozygous counterpart produces, though, “aberrant counting/chaotic choice”, i.e., presence of two active or inactive X's in a fraction of the cell population [18]. While that cannot be easily rationalized by other models (see, e.g., [21]), in our framework it is originated simply because the BF can fail to bind at all [34].
is deletion including , and , which in heterozygous causes a skewed XCI, as only the Wild Type X gets inactivated [21]. In the frame of our model could have a double effect: on the one hand, it hinders the binding of the AF and BF to the deleted X, by removing a number of their binding sites; on the other it affects especially the BF, since it removes the genes which are presumably linked to some of the BF components. Thus, the overall effect will be that while the deleted X remains active (as it lacks ), the BF is depleted and the AF wins the competition for binding the Wild Type chromosome, which is then inactivated.
Transgenic insertions are also interesting [35]. One of the predictions of our model is the highly non-linear effect of deletion/insertion, due to the “switch-like” nature of the underlying thermodynamic mechanism. The insertion experiments of ref. [35] support this view: long transgenes can cause inactivation on male ES cells only when they are present in multiple copies, while single insertions do not have appreciable effects. The outcome of other deletions/insertions, such as [36], - [37], [5], [6], etc., are similarly explained (see Text S1). XCI in diploid cells with more than two X and in polyploid cells [21] can be understood as well in our scenario (see Text S1), but additional biological hypotheses are required, since key pieces of information are still missing.
In summary, we illustrated physical switch-like mechanisms establishing conformational changes and symmetry breaking in a polymer model. For clarity, we included just the required minimal ingredients, but our model can accommodate more realistic molecular details. It can be mapped into the region of X chromosomes to explain their complex self-organization and other important aspects of random XCI, such as the deep connection between architectural changes and choice of fate, reconciling within a single framework a variety of experimental evidences. The on-off character of the underlying mechanisms can also explain how sharp and reliable regulation of XCI can be attained by simple strategies, such as gene upregulation or chromatin modification.
It supports a picture where random XCI could be governed by a few core molecular elements and basic physical processes. Two main groups of molecular factors are envisaged to control the process and to produce an activating and a blocking factor for . The specific polymer regions in our model emerge as key cis-regulators which orchestrates functional contacts along the . Experiments targeted at that area could test their role. The model also predicts threshold effects of, e.g., genetic deletions of the regulatory regions.
The precise nature of factors and sequences involved at XCI could differ from the minimal one considered here, yet the thermodynamic mechanisms we discussed are robust and independent of the specific molecular details. Similar mechanisms could be, thus, relevant to XCI and, more generally, to other nuclear processes requiring, for example, chromatin spatial reorganizations [38]–[40] or alternative choices [41].
|
10.1371/journal.pntd.0005779 | Analytical sensitivity and specificity of a loop-mediated isothermal amplification (LAMP) kit prototype for detection of Trypanosoma cruzi DNA in human blood samples | This study aimed to assess analytical parameters of a prototype LAMP kit that was designed for detection of Trypanosoma cruzi DNA in human blood. The prototype is based on the amplification of the highly repetitive satellite sequence of T.cruzi in microtubes containing dried reagents on the inside of the caps. The reaction is carried out at 65°C during 40 minutes. Calcein allows direct detection of amplified products with the naked eye. Inclusivity and selectivity were tested in purified DNA from Trypanosoma cruzi stocks belonging to the six discrete typing units (DTUs), in DNA from other protozoan parasites and in human DNA. Analytical sensitivity was estimated in serial dilutions of DNA samples from Sylvio X10 (Tc I) and CL Brener (Tc VI) stocks, as well as from EDTA-treated or heparinized blood samples spiked with known amounts of cultured epimastigotes (CL Brener). LAMP sensitivity was compared after DNA extraction using commercial fiberglass columns or after “Boil & Spin” rapid preparation. Moreover, the same DNA and EDTA-blood spiked samples were subjected to standardized qPCR based on the satellite DNA sequence for comparative purposes. A panel of peripheral blood specimens belonging to Chagas disease patients, including acute, congenital, chronic and reactivated cases (N = 23), as well as seronegative controls (N = 10) were evaluated by LAMP in comparison to qPCR. LAMP was able to amplify DNAs from T. cruzi stocks representative of the six DTUs, whereas it did not amplify DNAs from Leishmania sp, T. brucei sp, T. rangeli KPN+ and KPN-, P. falciparum and non-infected human DNA. Analytical sensitivity was 1x10-2 fg/μL of both CL Brener and Sylvio X10 DNAs, whereas qPCR detected up to 1x 10−1 fg/μL of CL Brener DNA and 1 fg/μl of Sylvio X10 DNA. LAMP detected 1x10-2 parasite equivalents/mL in spiked EDTA blood and 1x10-1 par.eq/mL in spiked heparinized blood using fiberglass columns for DNA extraction, whereas qPCR detected 1x10-2 par.eq./mL in EDTA blood. Boil & Spin extraction allowed detection of 1x10-2 par.eq /mL in spiked EDTA blood and 1 par.eq/ml in heparinized blood. LAMP was able to detect T.cruzi infection in peripheral blood samples collected from well-characterised seropositive patients, including acute, congenital, chronic and reactivated Chagas disease. To our knowledge, this is the first report of a prototype LAMP kit with appropriate analytical sensitivity for diagnosis of Chagas disease patients, and potentially useful for monitoring treatment response.
| Trypanosoma cruzi, a parasite transmitted to humans from hematophagous insects, causes Chagas Disease, a Neglected Tropical Disease with public health impact, affecting 7 million people in Latin America. Although mainly related to low income populations inhabiting rural environments, migrations have conveyed Chagas Disease to urban areas of endemic and non-endemic countries. It often presents non-specific symptoms, and direct, low cost microscopy-based diagnosis only detects acute infections, missing a high proportion of cases. Serology is the “gold standard” diagnostic technique for chronic stages and needs the concordance of at least two different assays to confirm infection. In this context, we aimed to evaluate the analytical sensitivity and specificity of a prototype kit based on a novel and rapid molecular biology reaction, named Loop mediated isothermal amplification (LAMP), using standardized Real Time PCR as a comparator. To our knowledge, this is the first LAMP prototype kit with an analytical performance appropriate for human diagnosis of Chagas disease and potentially useful for monitoring treatment response. Its simple handling using basic laboratory devices will enable point-of-care diagnosis and screening for congenital infection at birth as well as early detection of acute infections due to oral contamination.
| Chagas disease, caused by the parasite Trypanosoma cruzi, remains a major concern in 21 endemic countries of Latin America, where infection is acquired mainly through the triatomine insect vector. Due to migration movements, it has spread over other continents, with 6 to 7 million people estimated to be infected. T. cruzi infection can also be transmitted by blood transfusion, the trans-placental route causing Congenital Chagas disease, oral contamination, organ transplantation and laboratory accident [1].
Two disease stages can be distinguished and the strategies for diagnosis are stage-dependent. Firstly, a short acute stage occurs with patent parasitemia that can be detected using conventional parasitological techniques, such as parasite microscopic observation in blood smears or microhematocrite, xenodiagnosis and hemoculture. However, these methods usually lack sensitivity and are operator dependent, and the last two mentioned techniques are cumbersome and their results can be acquired only several weeks after sample collection [2]. In a majority of acute cases, symptoms are not evident and thus the infection mostly goes undiagnosed; it enters in an indeterminate chronic period that may span life-time in around 70% of cases. In the remaining 30%, chronic stage leads to cardiomyopathy and/or digestive megasyndromes, causing death if untreated. As in the chronic phase, parasitemia is intermittent and low, diagnosis is largely made by serological tests. Due to the antigenic variability of the parasite, WHO’s guidelines recommend to perform at least two serological assays based on distinct antigen sets, which must agree for a conclusive diagnosis [1].
The different transmission modes, the disease phases and the high genetic variability of the parasite increase the difficulties of making diagnostic kits with most appropriate markers for the diverse Chagas disease epidemiological settings. Nucleic acid amplification strategies have opened new options to detect T. cruzi infection and evaluate anti-parasitic chemotherapy.
Diagnostic assays for Chagas disease need improvement. The development of diagnostic test for the following areas have been identified as a priority: acute phase, including oral and congenital transmission and monitoring of anti-parasitic treatment response [2, 3,]. The use of Loop-mediated isothermal amplification (LAMP) of DNA has been proposed as an outstanding approach to bridge some of these gaps [3]. LAMP is a platform developed by Eiken Chemical Company of Japan (http://www.eiken.co.jp/en). This technology detects known genes from different pathogens [4] [5] [6]. It has the following characteristics: (1) only one enzyme is used and the amplification reaction proceeds under isothermal conditions [7] [8]); (2) extremely high specificity because of the use of four primers recognizing six distinct regions on the target; (3) high amplification efficiency, with DNA being amplified 109−1010 times within 15–60 minutes of incubation; and (4) it produces tremendous amounts of amplification product, making simple visual detection possible [8] [9]. Due to the above mentioned characteristics, LAMP can be performed in basic laboratories without the need for specialized infrastructure and it is appropriate for field applications and point-of-care diagnosis. Several of these kits are quite mature, such as the Tuberculosis and Malaria assays which are already commercialized Loopamp Assays. LAMP Assays—HUMAN Diagnostics Worldwide. Available: https://www.human.de/products/molecular-dx/isothermal-amplification/lamp-assays/)
A LAMP method for detection of T. cruzi DNA was previously designed based on the 18S ribosomal RNA (rRNA) gene and evaluated in DNA samples extracted from internal organs of triatomine vectors [10]. However, its low analytical sensitivity (100 fg per reaction tube) has not allowed its application to diagnosis of Chagas disease in humans.
LAMP tests based on amplification of multicopy repetitive sequences, such as the mobile element RIME of T. brucei [11] and the satellite DNA sequence of T. vivax [12] have been used to improve sensitivity. Accordingly, a novel LAMP assay based on the highly repetitive satellite DNA sequence of T.cruzi was developed to create a prototype kit for detection of this parasite in human blood. The present study aimed to assess the analytical performance of this kit prototype on DNA extracted from EDTA-treated as well as from heparinized human blood and explore its performance to detect T.cruzi DNA in blood samples from Chagas disease patients.
Informed written consent was obtained from all healthy donors and Chagas disease patients included in the study before blood collection, after permission of the IRB of the participating institutions, in agreement with argentine legislation in force (Blood Donation Law N° 22990, Res. N°1409/15)
Peripheral blood samples from a total of 23 well-characterised Chagas disease patients and 10 seronegative controls were tested by LAMP and qPCR. Four clinical groups with T.cruzi infection were evaluated, namely Group CI: Samples from five newborns/neonates born to Chagas disease women; Group AI-TXRI: samples from three transplanted seronegative receptors of organs from a seropositive donor that acquired acute T.cruzi infection; Group CCD: samples from ten chronic Chagas disease patients; Group RCD: samples from five chronic Chagas disease patients undergoing clinical reactivation due to immunosuppression after organ transplantation. All cases and controls were admitted at Health Centers of Argentina.
T. cruzi stocks belonging to the six discrete typing units (DTUs) [13] were used in the present study. Sylvio X10 (Tc I) and CL-Brener (Tc VI) are available at INGEBI since year 2009 and their identity is periodically assessed by multiplex Real Time PCR [14]. JG1 (Tc II), M6241 (Tc III) and M4167 (Tc IV) were kindly provided by Dr Constança Britto and Otacilio Moreira (Instituto Oswaldo Cruz, Rio de Janeiro, Brasil) and PAH179 (Tc V) by Dr Patricio Diosque (Instituto de Parasitología Experimental—IPE, University of Salta, Salta, Argentina). Strains were cultured using LIT medium and DNAs were purified using phenol-chloroform extraction followed by ethanol precipitation.
DNAs from Leishmania sp. and T.rangeli were kindly provided by Dr. Concepción Puerta from Pontificia Universidad javeriana—PUJ University, Bogotá, Colombia and DNA from T. brucei sp. and P. falciparum by Dr Yasuyoshi Mori from Eiken Chemical Company, Japan.
All DNAs were measured with microvolume UV-Vis spectrophotometer (Nanodrop 1000, Thermo Scientific, with nd_1000-v.3.8.1 software).
DNA was extracted from EDTA blood and heparinized blood samples, using two different preparation procedures:
a) Blood based-DNA for LAMP was obtained from 200 μL of EDTA-blood using the High Pure PCR Template Preparation Kit with packed fiberglass filter tubes, cell lysis buffer and proteinase K (RAS extraction kit, Roche Applied Sciences, Mannheim, Germany) with one additional step prior to the addition of 100 μL of elution buffer, which consists of spinning the columns after the second wash to eliminate any traces of isopropanol and b) a Boil & Spin rapid procedure modified according to the type of blood sample: b1) Boil & Spin of EDTA-blood samples: 200 μL of blood were mixed with 200 μL of 0.5% SDS in double distilled water in a 1.5 mL microtube with screwcap and o-ring by vortex (10 seconds), then heated in a thermoblock at 95°C for 5 min. The tube content was spun down for 5 minutes at the maximum speed (13,300 rpm) on a bench top centrifuge. The supernatant was pipetted into a new labeled 1.5 mL flat cap microtube, ready to use immediately or stored at −20°C for up to 48 h, prior to use in the LAMP reaction, with only one freeze-thaw cycle before testing. b2) Boil & Spin of heparinized blood samples: 300 μL of blood and 15 μL of 10% SDS in double distilled water were thoroughly mixed in a 1.5 mL microtube with screwcap and o-ring by vortex (30 seconds). 100 μL of the solution were withdrawn, transferred to a new tube containing 400 μL of sterile water, then heated in a thermoblock at 90°C for 10 minutes. The tube content was spun down for 3 minutes at maximum speed (13,300 rpm). The supernatant was pipetted into a new labeled 1.5 mL flat cap microtube and the previous step was repeated once. The supernatant was used immediately for the LAMP reaction.
Blood based-DNA for qPCR was obtained from 200 μL of EDTA-blood using the the RAS extraction kit as detailed for LAMP assays.
This product is based on the nucleic acid amplification method, LAMP, developed by Eiken Chemical Co., Ltd. Japan, using as molecular target the repetitive satellite DNA sequence of T. cruzi. The primer sequences were designed after alignment analysis of selected database sequences to have a relatively well-conserved sequence recognized in all six described DTUs. The nucleotide sequence of the primers used in the study are not provided, since the assay is being developed into a commercial product. T. cruzi Loopamp kits can be obtained from Eiken (http://www.eiken.co.jp/en/inquiry.html.).
The reaction tube contains strand displacement Bst (Bacillus stearothermophilus) DNA polymerase, deoxynucleotide triphosphates (dATP, dCTP, dGTP and dTTP), calcein and T. cruzi-specific primers.
These reagents are in dried form on the inside of the cap of the reaction tube and are stable for one year at 30°C. A negative control (NC, distilled water) is provided in the kit. The positive control was 1fg/μL of DNA from CL-Brener stock (Tc VI).
The LAMP reaction was performed as follows: 30 μL of sample DNA extract was dispensed into each LAMP tube, and the cap was closed. Each tube was flicked down to collect the solution at the bottom and placed upside down during two minutes to reconstitute the dried reagent. It was inverted five times to mix the content followed by a spin down.
Incubation of the reaction was carried out at 65°C for 40 minutes for isothermal amplification, followed by a step at 80°C for five minutes for enzyme inactivation using different devices: 1) Rotor Gene 6000 thermocycler (Corbett Life Science, Cambridgeshire, UK); Perkin-Elmer 9600 (Thermofisher, USA) and Genie III Fluorimeter Instrument (Optigene, Horsham, West Sussex, UK).The results of amplification were visualized using different strategies: i) Real-time fluorescence data was obtained on the Rotor Gene 6000 FAM channel (excitation at 470 nm and detection at 510 nm, auto-optimisated gain level -2, 33 after 40 cycles of 60 seconds each at 65°C to achieve an isothermal reaction followed by a hold on 80°C during 5 minutes) ii) Genie III Instrument (excitation at 470 nm, detection at 510 nm, low gain level 4, specific for calcein; the run profile settings for amplification were 65°C, 40 min. To end reaction, the anneal settings starts at 98°C with a ramp rate of 0,06°C/second to reach 80°C during 5 minutes in order to inactivate the enzyme iii) visualization of fluorescence by the naked eye iv) in some cases, 1.2% agarose gel electrophoresis with TAE (Tris-acetate with EDTA), ethidium bromide and GelPilot loading dye 5x, final dilution 1x, run at 80 V (Qiagen, Hilden, Germany) was carried out to corroborate the typical ladder profile of multiple bands of LAMP products [15].
The LAMP method was compared to a standardized qPCR assay that consisted of a duplex qPCR with TaqMan probes targeted to T. cruzi satellite DNA and Internal Amplification control (IAC), as described by Duffy and coauthors [16] on the basis of the T.cruzi primer and probe sequences published by Piron and coauthors [17], and validated following international guidelines [16,18]. The qPCR reactions were carried out in duplicates with 5 μL of eluted DNA in a final volume of 20 μL containing 2X FastStart Universal Probe Master Mix (Roche Diagnostics GmbHCorp., Mannheim, Germany), 10 μM primers cruzi 1 and cruzi 2 and probe cruzi 3 [19]; 5 μM primer IAC FW, primer IAC Rv and probe IAC-Tq [14]. Cycling conditions were a first step of 10 minutes at 95°C followed by 40 cycles at 95°C for 15 seconds and 58°C for 1 minute. The amplifications were carried out in an ABI7500 (Applied Biosystems, Foster City, CA, USA) or a Rotor-Gene 6000 (Corbett Life Science, Cambridgeshire, United Kingdom) thermocycler equipments. All reactions included a strong positive control (SPC): 10 fg/μL of CL Brener DNA and a weak positive control (WPC): 1 fg/μL of CL-Brener DNA.
Analytical sensitivity was estimated on ten-fold serial dilutions done in triplicate from three independent DNA samples of Sylvio X10 (Tc 1) and CL Brener (Tc VI) stocks (Fig 3).
The prototype LAMP kit detected T.cruzi DNA at concentrations ≥ 1 x 10−2 fg/μL (0.3 fg per test) in both CL Brener and Sylvio X10 stocks. DNA samples spanning 1 fg/μL to 1 x 10−3 fg/μL were also assayed by qPCR for comparative purposes (S2 and S3 Figs). The qPCR detected up to 1 x 10−1 fg/μL of CL Brener DNA (S2 Fig) and 1 fg/μL of Sylvio X10 DNA, whereas 1 x 10-1fg/μL of the latter was detectable in only one aliquot (S3 Fig, aliquot A1).
The analytical sensitivity of LAMP was tested in EDTA and heparinized-blood samples spiked with known parasite loads (par.eq./mL).
EDTA blood samples spiked with CL Brener DNA showed positive LAMP results when using the RAS extraction kit (6 x 10−4 par.eq. per test; Fig 4A) and the Boil & Spin preparation method (3 x 10−4 par.eq. per test; Fig 4B).
The qPCR was carried out in spiked EDTA blood samples following the standardized procedure, which gave a sensitivity of 1 x 10−2 par.eq./mL (S1 Table)
A panel of peripheral blood specimens belonging to Chagas disease patients (N = 23), as well as seronegative controls (N = 10) were evaluated by LAMP in comparison to qPCR. LAMP allowed detection of T.cruzi DNA in all congenital Chagas disease (CI), seronegative receptors of organs from seropositive donors with acute infection (AI-TXRID) and reactivated Chagas disease (RCD) patients, in agreement with qPCR positivity (Table 2). However, in a panel of samples from ten chronic Chagas disease patients (CCD), LAMP detected T.cruzi DNA in four, whereas qPCR was positive in three, suggesting higher sensitivity of LAMP with respect to qPCR. LAMP gave negative results in samples from ten seronegative patients in agreement with qPCR results (Table 2, NI).
Examples of the tested cases are illustrated in Fig 6 (LAMP results) and S5 Fig (qPCR results).
Clinical molecular diagnosis of Chagas disease is important for: (i) early diagnosis of congenital transmission in newborns when presence of maternal anti-T. cruzi antibodies may deliver false positive results [20], (ii) early detection of infection in transplant receptors of organs from a Chagas positive donor [21], (iii) monitoring of parasite reactivation in chronically infected patients immune-suppressed due to organ transplantation [22] or AIDS [23], and (iv) the evaluation of new treatments in clinical trials, because detection of serological negative conversion in seropositive patients showing a favorable treatment outcome is impractical from a study time perspective [24] [25].
This work aimed to assess the performance of a LAMP kit prototype targeted to satellite T. cruzi DNA, a highly repetitive and conserved sequence in all characterized T.cruzi strains.
The prototype kit has the advantage of using reagents in dried form on the inside of the cap of the reaction tubes. Moreover, the use of calcein allows direct visualization of amplification by the naked eye. Calcein is included in the reaction tubes in a quenched state, bound to manganese ions. Once the LAMP reaction starts, pyrophosphate ions that are generated bind to the manganese ions, releasing calcein and generating fluorescent light. Furthermore, the presence of magnesium ions in the buffer system enhance calcein fluorescence [9].
Analytical sensitivity and specificity of the kit prototype was carried out using purified DNA from different parasite stocks representative of the six T. cruzi DTUs, as well as from human blood samples anti-coagulated with EDTA or with heparin and spiked with known quantities of T. cruzi cells. Furthermore, the prototype was evaluated in a blind panel of peripheral blood samples from well characterized Chagas disease patients at different stages and with different clinical manifestations, as well as seronegative donors as non-infected controls (Table 2 and S5 Fig).
The LAMP kit was able to amplify DNA from all tested T.cruzi stocks by the naked eye and by Real Time detection of fluorescence in a Rotor Gene Corbett Thermocycler (Table 1).
Analytical sensitivity was calculated for Tc I and Tc VI stocks, among other parasite strains, and was higher than the analytical sensitivity observed with the same stocks using standardized qPCR based on the same target [16]. The kit does not intend to discriminate between DTUs. Variations in Cts among T.cruzi stocks can be explained due to the heterogeneity in copy numbers of satellite repeats [19], however sensitivity was high for all tested strains. Indeed, satellite DNA has been successful as a molecular target for LAMP assays to detect infections by T. vivax [12] and to implement a direct dried blood based diagnostic tool for Human African trypanosomiasis [11] [26].
LAMP was specific for T. cruzi DNA since negative findings were obtained even with high concentrations (up to 1 pg/μL) of DNAs from T. rangeli (KPN+ and KPN- stocks), P. falciparum, several Leishmania species, and T. brucei sub-species.
The use of EDTA as anticoagulant proved to be suitable in our study, although it has not been recommended for LAMP, due to the fact that EDTA competes for manganese ions with the pyrophosphate ions generated once the reaction starts [27]. Nevertheless, LAMP analysis carried out using DNA obtained from EDTA-blood samples has also been reported [26]. Since an easy and quick detection is desirable for application of a LAMP prototype in point-of-care diagnosis, EDTA-blood samples are appropriate since they are routinely withdrawn in most health care centers; i.e. for hemogram reports. In fact, we detected up to 1 x 10−2 par. eq./mL in extracts from spiked EDTA blood samples using fiberglass commercial columns as well as after Boil & Spin, in accordance to the sensitivity detected by qPCR.
In contrast, when LAMP experiments were performed in heparinized spiked blood using RAS columns, sensitivity was one order below. Moreover, LAMP results obtained using 30 μL of Boil & Spin DNA preparations from heparinized blood could only be detected after agarose gel electrophoresis, whereas observation of fluorescence with the naked eye was not possible: reddish appearance of the contents of tubes hampered visualization (Fig 5B). Direct visualization was possible only after diluting the DNA extracts at least 1:100 times. However, the analytical sensitivity in those diluted samples was poor in comparison to the detection limit achieved using the other methods; it was only 1 par.eq./mL (S4 Fig). Consequently, the Boil & Spin procedure used in our study needs to be revised for improvement. It performed well in other Loopamp kits, such as those developed for Pan/ P. falciparum (27, 28) and Trypanosoma brucei detection (Standard operating procedures for the Loopamp Trypanosoma Available: https://www.finddx.org/wp-content/uploads/2016/06/HAT-LAMP-SOP_13JUN16.pdf).
The analytical sensitivity of this LAMP assay was superior to that previously obtained using a LAMP procedure based on the 18S rDNA gene [10]. The mentioned study was done using DNA from Tulahuen strain (Tc VI), and analytical sensitivity was 100 fg per test, using a Real-time turbidimeter, detection under UV light and direct visualization by the naked eye after 60 minutes of incubation.
To our knowledge, this is the first report of a LAMP kit with similar analytical sensitivity than Real Time PCR. It was validated using the same T.cruzi strains [16, 28] and using different methods for visualization of amplification. Our data provide evidence of the usefulness of this LAMP kit for molecular diagnosis of Chagas disease. Prospective analysis of clinical specimens will allow establish its performance in different epidemiological and clinical scenarios, such as early diagnosis of congenital infection, POC detection of acute infections due to oral transmission or in seronegative recipients of organs from seropositive donors, early detection of reactivation due to immunosuppression due to organ transplantation or AIDS [3,20,21,22,23,29]. Furthermore, the persistence of positive LAMP results in patients under etiological treatment could be useful to assess treatment failure [3, 24, 25, 30, 31,32].
|
10.1371/journal.pntd.0007673 | Transcriptional blood signatures for active and amphotericin B treated visceral leishmaniasis in India | Amphotericin B provides improved therapy for visceral leishmaniasis (VL) caused by Leishmania donovani, with single dose liposomal-encapsulated Ambisome providing the best cure rates. The VL elimination program aims to reduce the incidence rate in the Indian subcontinent to <1/10,000 population/year. Ability to predict which asymptomatic individuals (e.g. anti-leishmanial IgG and/or Leishmania-specific modified Quantiferon positive) will progress to clinical VL would help in monitoring disease outbreaks. Here we examined whole blood transcriptional profiles associated with asymptomatic infection, active disease, and in treated cases. Two independent microarray experiments were performed, with analysis focussed primarily on differentially expressed genes (DEGs) concordant across both experiments. No DEGs were identified for IgG or Quantiferon positive asymptomatic groups compared to negative healthy endemic controls. We therefore concentrated on comparing concordant DEGs from active cases with all healthy controls, and in examining differences in the transcriptome following different regimens of drug treatment. In these comparisons 6 major themes emerged: (i) expression of genes and enrichment of gene sets associated with erythrocyte function in active cases; (ii) strong evidence for enrichment of gene sets involved in cell cycle in comparing active cases with healthy controls; (iii) identification of IFNG encoding interferon-γ as the major hub gene in concordant gene expression patterns across experiments comparing active cases with healthy controls or with treated cases; (iv) enrichment for interleukin signalling (IL-1/3/4/6/7/8) and a prominent role for CXCL10/9/11 and chemokine signalling pathways in comparing active cases with treated cases; (v) the novel identification of Aryl Hydrocarbon Receptor signalling as a significant canonical pathway when comparing active cases with healthy controls or with treated cases; and (vi) global expression profiling support for more effective cure at day 30 post-treatment with a single dose of liposomal encapsulated amphotericin B compared to multi-dose non-liposomal amphotericin B treatment over 30 days. (296 words; 300 words allowed).
| Visceral leishmaniasis (VL), also known as kala-azar, is a potentially fatal disease caused by intracellular parasites of the Leishmania donovani complex. VL is a serious public health problem in rural India, causing high morbidity and mortality, as well as major costs to local and national health budgets. Amphotericin B provides improved therapy for VL with single dose liposomal-encapsulated Ambisome, now affordable through WHO-negotiated price reductions, providing the best cure rates. The VL elimination program aims to reduce the incidence rate in the Indian subcontinent to <1/10,000 population/year. By assessing immune responses to parasites in people infected with L. donovani, but with different clinical status, we can determine the requirements for immune cell development and predict which asymptomatic individuals, for example healthy individuals with high anti-leishmanial antibody levels, will progress to clinical VL. This will help in monitoring disease outbreaks. In this study we looked at global gene expression patterns in whole blood to try to understand more about asymptomatic infection, active VL, and the progress to cure in cases treated with single or multi-dose amphotericin B. The signatures of gene expression identified aid in our understanding of disease pathogenesis and provide novel targets for therapeutic intervention in the future.
| Visceral leishmaniasis (VL), also known as kala-azar, is a potentially fatal disease caused by obligate intracellular parasites of the Leishmania donovani complex. VL is a serious public health problem in indigenous and rural populations in India, causing high morbidity and mortality, as well as major costs to both local and national health budgets. The estimated annual global incidence of VL is 200,000 to 400,000, with up to 50,000 deaths annually occurring principally in India, Bangladesh, Sudan, South Sudan, Ethiopia and Brazil [1]. In India, improvements in drug therapy have been afforded through the introduction of amphotericin B treatment, with single dose liposomal encapsulated Ambisome providing the best cure rates and now being used as the preferred treatment regime in the VL elimination program [2]. However, with the potential development of drug resistance to each new therapeutic approach [3], there remains a continuing need for improved and more accurate methods of early diagnosis, as well as ability to monitor responses to treatment and to predict disease outcome. These objectives are also important in relation to the World Health Organization-supported VL elimination initiative in the Indian subcontinent, which aims at reducing the incidence rate of VL in the region to below 1 per 10,000 population per year by 2020 [4]. Monitoring disease outbreaks in the context of the elimination program will be an important goal, including the ability to determine which individuals displaying asymptomatic disease, as monitored by anti-leishmanial IgG [5, 6] and/or Leishmania-specific modified Quantiferon responses [7], will progress to clinical VL disease [6].
In recent years, the use of whole blood transcriptional profiling in humans has provided a better understanding of the host response to infectious disease, leading to the identification of blood signatures and potential biomarkers for use in diagnosis, prognosis and treatment monitoring (reviewed [8]). Pioneering studies using this approach were successful in identifying a neutrophil-driven interferon (IFN)-inducible blood transcriptional signature for active tuberculosis that involved both IFN-γ and type I IFN-α/β signalling [9] and was subsequently confirmed in multiple countries world-wide (reviewed [8]). This neutrophil-driven interferon signature was present in active disease but absent in both latent infection and in healthy controls [9]. While an IFN-inducible signature was also identified in patients with the autoimmune disease systemic lupus erythematosus, there were differences in the signatures that also distinguished the two profiles from each other [9]. Viral infections [10] and bacterial infections like melioidosis [11] are also broadly characterised by IFN-inducible gene expression, but whole blood signatures have been identified that are able to discriminate between bacterial and viral infections [10, 12], as well as between different viral infections [10]. In HIV, blood transcriptional signatures have been identified that distinguish between rapid compared to slow progression to disease [13]. Blood signatures have also been identified which distinguish between children who acquire dengue virus fever compared to those who develop dengue haemorrhagic fever [14, 15]. There are also signatures that distinguish between pulmonary and extra-pulmonary tuberculosis [16], as well as between pulmonary tuberculosis, pulmonary sarcoidosis, pneumonias and lung cancers [17]. A transcriptional signature that can be used to monitor treatment response is also a valuable goal in infectious disease. Again, studies from two cohorts followed longitudinally in South Africa show that the transcriptional signature of active tuberculosis disease rapidly diminishes with successful treatment [18, 19].
More recently, whole blood transcriptional profiling has been used to study human host responses to protozoan pathogens such as malaria [20, 21] and Chagas disease caused by Trypanosoma cruzi [22, 23]. Expression profiling has also been applied in the context of animal models [24–26] and in human studies [15, 27–29] of the leishmaniases. In particular, whole blood transcriptomics was used to compare expression profiles in patients with active VL caused by L. infantum with asymptomatic infected individuals, patients under remission from VL, and controls [27]. While VL patients exhibited profiles reflecting activation of T cells via MHC Class I signalling and type I interferon, patients in remission showed heterogeneous profiles associated with T cell activation, type I interferon signalling, cell cycle, activation of Notch signalling, and an increased proportion of B cells. Asymptomatics (as determined by a positive delayed type hypersensitivity response to leishmanial antigen) and uninfected individuals exhibited similar gene expression profiles. Here we also set out to determine whole blood transcriptional profiles that might distinguish unifected individuals from asymptomatic infection or active disease caused by L. donovani in India, as well as to monitor the changes in transcriptional profiles that accompanied drug treatment. Whilst we were unable to detect a signature that distinguished asymptomatic (IgG antibody positive [5, 6], or modified Quantiferon positive [7]) individuals from healthy endemic controls who were negative by these two assays, we were able to determine the transcriptional profile of active VL cases, and to demonstrate interesting differences in return to baseline between patients treated with non-liposomal compared to liposomal-encapsulated (Ambisome) amphotericin B.
The enrolment of human subjects complies with the principles laid down in the Helsinki declaration. Institutional ethical approval (reference numbers: Dean/2012-2013/89) was obtained from the ethical review board of Banaras Hindu University (BHU), Varanasi, India. Informed written consent was obtained from each participant at the time of enrolment, or from their legal guardian if they were under 18 years old. Only patients who had not previously received treatment and who agreed to participate in the study were enrolled. All clinical treatment and follow-up records were maintained using standardised case report forms on an electronic server. All patient data were analysed anonymously.
In this study two independent microarray experiments were performed. For experiment 1, samples were collected between February and April 2011. For experiment 2, samples were collected between April and July 2012. Samples were collected at the Kala-azar Medical Research Center (KAMRC), Muzaffarpur, Bihar, India, or in nearby field sites for some asymptomatic individuals and endemic controls. Active VL cases were diagnosed by experienced clinicians based on clinical signs, including fever (>2 weeks), splenomegaly, positive serology for recombinant antigen (r)-K39 and/or by microscopic demonstration of Leishmania amastigotes in splenic aspirate smears. VL patients were treated according to routine clinical care with either (a) experiment 1: 0.75 mg/kg non-liposomal amphotericin B daily for 15 days (N = 3), or on alternate days over 30 days (N = 7), by infusion (i.e. 15 doses in all; total dose 11.25 mg/kg over 30 days); or (b) experiment 2: 10 mg/kg of Ambisome (liposome-encapsulated amphotericin B) as a single dose by infusion. Blood samples were collected pre- (N = 10 experiment 1; N = 11 experiment 2) and post- (day 30; N = 10 experiment 1; N = 11 experiment 2) treatment. There were 9 paired pre-/post-treatment samples for experiment 1; 10 for experiment 2. Healthy control subjects included (i) asymptomatic individuals (N = 2 experiment 1; N = 6 experiment 2) who had sustained high anti-leishmanial antibody levels by direct agglutination test (DAT titer ≥1:25,600) over two annual surveys prior to blood collection for profiling [6]; (ii) asymptomatic individuals (N = 8 experiment 1; N = 9 experiment 2) who were positive by Leishmania-specific modified quantiferon assays [7] over two annual surveys prior to blood collection for profiling; and (iii) Serology (DAT titer ≤1:1600) and quantiferon negative healthy endemic controls (N = 6 experiment 1; N = 10 experiment 2) who were negative by both assays over two annual surveys prior to blood collection for profiling. Sample sizes are for post-QC samples used in expression profiling studies (see below). Further clinical and demographic details on participants are provided in S1 Table. The work flow for data analysis is provided in S1 Fig.
Whole blood (5 mL) collected by venepuncture was immediately placed into Paxgene tubes (QIAGEN GmbH, Germany) and stored at -80°C for later processing for RNA. RNA was extracted using PAXgene Blood RNA kits (QIAGEN GmbH, Germany) according to manufacturer’s instructions. RNA integrity and purity were checked using Tape Station 4200 (Agilent Technologies, USA). Samples used for beadchip analysis had RNA integrity (RIN) mean±SD values 6.75±0.67 (range 5.5–7.7). Globin mRNA was depleted using GLOBINclear-Human kits (ThermoFischer Scientific, USA). RNA was reverse transcribed and biotin-labelled using the Illumina TotalPrep RNA Amplification kit (ThermoFischer Scientific, USA). The resulting biotinylated cRNA was hybridised to Illumina HT12v4 Expression BeadChips, specifically HumanHT-12_V4_0_R2_15002873_B, containing 47,323 genome wide gene probes, and 887 control probes. Samples from different control or clinical groups were distributed evenly across 3 (experiment 1) or 4 (experiment 2) beadchips. All RNA preparation and processing of samples over beadchips was carried out at Sandor Lifesciences Pvt. Ltd. (Hyderabad, India).
All data analysis was carried out in R Version 3.4.3 (Smooth Sidewalk - https://www.r-project.org/) and RStudio (version 1.1.383). The Bioconductor package Lumi [30] was used to read in raw expression values and perform quality control. Background correction and quantile normalisation of the data was carried out using the Bioconductor package Limma [31]. Pre-processing of the microarray data and removal of non-expressed (detection P-value > 0.05 in all arrays) and poor quality probes previously shown to have unreliable annotation [32] provided 21,959 and 23,466 probe sets which passed QC in experiments 1 and 2, respectively. Principal components analysis (PCA) and unsupervised cluster analysis (Pearson’s correlation coefficient; hclust = complete) of normalised data was performed in R. Data was visualised using the R packages ggplot2 (3.1.2) [33] and pheatmap (1.0.12) [34]. Differential expression analysis using linear modelling and empirical Bayes methods was carried out in the Bioconductor package Limma [31] for comparisons between control and clinical groups, as indicated. The threshold for differential expression was a log2-fold-change ≥1 (i.e. ≥2-fold) and/or Benjamini-Hochberg [35] adjusted p-value (Padj) ≤0.05, as indicated. Genes achieving these thresholds were taken forward in analyses using the gene set enrichment tool Enrichr [36], and using Ingenuity Pathway Analysis (IPA) (Ingenuity Systems, www.ingenuity.com) to identify canonical pathways, upstream regulators, and gene networks. Enrichr [36, 37] accesses a wide range of open access databases to identify terms (pathways/processes/disease states) for which the gene set is enriched. Input to Enrichr comprised lists of DEGs for specific between-group comparisons, as indicated, and did not include expression level data for individual genes. Enrichr uses four scores to report enrichment: a p-value (reported here as Pnominal) calculated using Fisher’s exact test; a q-value (reported here as Padj) which is the Benjamin-Hochberg adjusted p-value; a rank or z-score of the deviation from the expected rank by the Fisher’s exact test; and a combined score which is a combination of the p-value and z-score calculated by multiplying the two scores using the formula c = ln(p)*z. This z-score and the combined score correct for biases in ranking of term lists based solely on Fisher’s exact test [37] and outperform other enrichment methods in benchmarking studies [36]. The Enrichr z-score is not an activation score. IPA uses the Ingenuity Knowledge Base, an extensive database comprising biological pathways and functional annotations derived from the interactions between genes, proteins, complexes, drugs, tissues and disease, to carry out all its analyses. Benjamini-Hochberg correction was applied where applicable and Padj ≤ 0.05 was used to filter all results. Canonical pathway predicts known biological pathways that are changing based on the pattern of gene expression. The p-value uses Fisher’s Exact Test and does not consider the directional effect of one molecule on another, or the direction of change of molecules in the dataset. The significance level is the most important metric. The Z-score in IPA canonical pathway analysis is an activation z-score that takes account of known directional effects of one molecule on another or on a process, and the direction of change of molecules in the dataset. However, just because a pathway does not have a good z-score does not make it uninteresting. Upstream Regulator Analysis within IPA was employed to predict if there were any endogenous genes/cytokines/transcription factors which may be responsible for the observed gene expression patterns. If an upstream regulator is identified, an activation Z-score is calculated based on the fold change values of its target genes within the dataset. A Z-score ≥2 suggests that an upstream regulator is activated whereas a Z-score ≤-2 suggests it is inhibited, with active VL cases being the experimental group of baseline comparator. IPA also generates a “Top Tox List” pathway which provides an indication of toxic or pathogenic pathways that could be amenable to therapeutic intervention. Networks were constructed in IPA using the “Connect” option under the “Build” functionality. Genes with no previously documented interactions were removed from the diagram and the functions of each network were inferred from the remaining connected genes in each time-point. Nonparametric Gene Set Enrichment Analysis (GSEA [38]) using expression values to rank genes by their differential expression between two phenotypes was also used as an additional tool to ensure that important differences were not missed due to stringency of parametric methods, especially in the comparing the uninfected healthy endemic control groups with healthy antibody positive or healthy quantiferon positive groups. For this analysis we compared our data to the Blood Transcription Module gene list for antibody responses to vaccines ([39] in addition to the canonical pathway (CP) collection of the GSEA-MSigDB C2 curated gene sets (C2) [38]. GSEA was run for 1,000 permutations using weighted enrichment statistic and signal-to-noise ranking metric.
Two independent microarray experiments were carried out to compare transcriptional profiles across clinical groups that included active VL cases pre-treatment (N = 10 experiment 1; N = 11 experiment 2), drug treated VL cases (N = 10 experiment 1; N = 11 experiment 2), modified quantiferon [7] positive asymptomatic individuals (N = 8 experiment 1; N = 9 experiment 2), high Leishmania-specific antibody positive (by DAT) asymptomatic individuals (N = 2 experiment 1; N = 6 experiment 2), and endemic healthy controls (N = 6 experiment 1; N = 10 experiment 2) who were both modified quantiferon negative and antibody negative by DAT. PCA of the top 500 most variable probes (Fig 1) across all pairwise comparisons of samples showed that principal component 1 (PC1) accounted for 45% (experiment 1; Fig 1A and 1B) and 31% (experiment 2; Fig 1D and 1E) of the variation and resolved active cases compared to endemic healthy control and asymptomatic groups. The latter were not well resolved from each other in either experiment. Treated patients sat intermediate between, and overlapping with, both active cases and control/asymptomatic groups in experiment 1 but showed greater overlap with control/asymptomatic groups in experiment 2 (see below). This is particularly apparent when comparing plots of PC1 by PC3 (Fig 1B and 1E). Unsupervised hierarchical cluster analysis also (Fig 1C and 1F) provided discrete clusters of active cases compared to control and asymptomatic individuals, with treated cases interspersed with both active cases and control groups and not falling into a single discrete cluster in either experiment.
Consistent with the PCA plots (Fig 1), there were no differentially expressed probes representing genes (i.e. Benjamini-Hochberg [35] Padj ≤0.05) when comparing either modified quantiferon positive asymptomatic individuals with endemic healthy controls, or when comparing high antibody titer individuals with endemic healthy controls, in either experiment 1 or experiment 2. The additional non-parametric analysis carried out using GSEA also failed to identify gene sets enriched in controls compared to high antibody titre or modified quantiferon positive asymptomatic individuals that could be replicated across the two experiments (S2 and S3 Tables). For the analyses presented below, these groups were therefore analysed as one group referred to as “controls” or “healthy controls” in all further differential expression analyses.
Differential expression analysis focused on the comparison of (i) active VL cases versus controls, (ii) treated VL cases versus controls, and (iii) active versus treated VL cases. Log2-fold-change in experiment 1 was highly correlated with log2-fold-change in experiment 2 across all probes. Table 1 shows the number of differentially expressed probes representing genes in experiment 1 and experiment 2 for each comparison as well as the number of differentially expressed probes that replicated and were concordant for direction of effect between the two cohorts. At Padj ≤0.05, there are 2,584 concordant differentially expressed probes in common when comparing active cases with controls, 37 concordant probes when comparing treated cases with controls, and 221 concordant probes when comparing active and treated cases. At the more stringent threshold of ≥2-fold change there were 439, 8, and 42 concordant probes for these comparisons, respectively.
Of note, we found a greater number of transcriptional differences between treated cases and controls in experiment 1 compared to experiment 2 (Table 1; differentially expressed probes are 1132 and 126, respectively, at Padj ≤0.05). One explanation for this could be the different treatment regimen employed in the two cohorts. VL patients of the first experiment were treated with 15 doses of a non-liposomal form of amphotericin B over 30 days. In experiment 2 patients received a single dose of liposomal amphotericin B, which has shown better efficacy for the treatment of VL [40, 41]. The effect of treatment regimen on whole blood transcriptional profiles is further indicated by the comparison of active and treated cases. In this case, fewer differences in transcriptional regulation are observed between active and treated cases in experiment 1 as opposed to experiment 2 (Table 1; differentially expressed probes are 654 and 1317, respectively, at Padj ≤0.05), in which patients have received a more efficacious therapy. These findings agree with the PCA results (Fig 1), in which treated cases of the experiment 1 cohort form a more discrete group between active cases and controls (Fig 1A and 1B) whereas treated cases of the experiment 2 cohort are grouped more closely to controls (Fig 1D and 1E).
Due to the small number of concordant differentially expressed genes identified for the treated cases versus controls (Table 1; 37 at Padj ≤0.05, 8 at Padj ≤0.05 and ≥2-fold change), only the concordant differentially expressed gene sets for active cases versus controls and active cases versus treated cases were used in subsequent pathway and gene set enrichment analyses.
S1 and S2 Data provide spreadsheets of the data from experiments 1 and 2 respectively for all concordant DEGs that were significant at Padj<0.05).
Heatmaps were generated for individual expression levels for probes representing the top 10 concordant genes expressed at a higher (“induced”) level (Fig 2A), and the top 10 concordant genes expressed at a lower (“repressed”) level (Fig 2B), in active cases compared to controls in experiment 1. Heatmaps for the same “induced” and “repressed” probes/genes in experiment 2 are presented in Fig 2C and 2D. Of note 8/10 “repressed” genes were also in the top 10 most highly differentially expressed “repressed” genes in experiment 2; all 10 genes achieved ≥2-fold change in both experiments. Amongst these 10 most “repressed” genes were: peptidase inhibitor 3 (PI3), a known antimicrobial peptide for bacteria and fungi that is upregulated by lipopolysaccharide and cytokines; the C-C chemokine ligand 23 (CCL23; represented by 2 probes) which acts as a chemoattractant for resting (but not active) T cells, monocytes, and to a lesser extent neutrophils; G-protein-coupled C-C motif chemokine receptor 3 (CCR3) which binds CCL10 (eotaxin), CCL26 (eotaxin-3), CCL7 (MCP3), CCL13 (MCP4) and CCL5 (RANTES) that likewise act as chemoattractants for eosinophils, monocytes and neutrophils; ALOX15 which is a lipoxygenase known to regulate inflammation and immunity; and the G-protein-coupled prostaglandin D2 receptor 2 (PTGDR2 alias GPR44) that is preferentially expressed in CD4 effector T helper 2 (Th2) cells and mediates pro-inflammatory chemotaxis of eosinophils, basophils and Th2 cells. For the “induced” genes (Fig 2A and 2C), only 2/10 (the top 2 in both experiments) were also in the top 10 “induced” genes in experiment 2, but all achieved ≥2-fold change in both experiments. In addition to type I interferon inducible 27 (IFI27) and complement C1q B chain (C1QB) genes, there was a bias amongst the most strongly “induced” genes towards genes involved in erythrocyte function, including: glycophorin B (GYPB), a major sialoglycoprotein of the human erythrocyte membrane; Rh D blood group antigens (RHD); hemoglobin subunit delta (HBD); 5'-aminolevulinate synthase 2 (ALAS2) an erythroid-specific enzyme located in the mitochondrion and involved in heme biosynthesis; carbonic anhydrase 1 (CA1) which is found at its highest level in erythrocytes; atypical chemokine receptor 1 (Duffy blood group) (ACHR1 alias DARC) known for its role as the erythrocyte receptor for Plasmodium vivax and P. knowlesi; and 2,3-diphosphoglycerate (2,3-DPG) (BPGM) found at high concentrations in red blood cells where it binds to and decreases the oxygen affinity of haemoglobin.
To gain a more global picture of the impact of differential gene expression, the 391 genes represented by 439 probes that were concordant for differential gene expression (Padj ≤0.05; ≥2-fold change) between active cases and controls in experiments 1 and 2 were taken forward in Ingenuity Pathway (IPA) and gene-set enrichment (Enrichr) analyses. IPA network analysis indicated that 254 of these genes are joined in a single network (Fig 3), with IFNG as the major hub gene (i.e. with most connections to other genes in the network), and other major hub genes including CCNA2, CXCL10, SPI1, SNCA, CHEK1, MCM2, AURKB, RARA, CDK1, CDC20, and FOXM1. The top Ingenuity Canonical Pathways for the 391 genes that achieved Padj <0.05 and ≥2-fold change (Table 2) were Estrogen-mediated S-phase Entry (P = 6.46x10-5; Padj = 0.019; z-score 2), Mitotic Roles of Polo-Like Kinases (P = 1.20x10-4; Padj = 0.019; z-score 1.89), Aryl Hydrocarbon Receptor (AHR) Signalling (P = 1.51x10-4; Padj = 0.019; z-score 1.89), and Heme Biosynthesis II (P = 3.72x10-4; Padj = 0.035). Although not achieving Padj≤0.05, identification of the Th2 pathway (P = 1.22x10-3; Padj = 0.074; z-score -0.82) and Activation of Th1 and Th2 Pathway (P = 1.32x10-3; Padj = 0.074) as nominally significant canonical pathways is consistent with prior knowledge of immune responses to leishmaniasis.
Activation of AHR signalling (z-score 1.89) as a top Ingenuity Canonical Pathway is reflective of increasing recognition of the role of AHR signalling in immunity, including the ability of AHR ligands to significantly induce cell secretion of IL-10 and inhibit IL-1β and IL-6 production in dendritic cells, and to promote IL-10 production and suppress IL-17 expression in CD4(+) T cells [42–44]. It is also reflected in the identification of RARA, CCNA2 and CHEK1 genes from the AHR pathway (Table 2) as major hub genes (Fig 3). AHR signalling was also identified as top in the Ingenuity “Top Tox List” pathway (P = 4.16x10-4) indicative of its role as a toxic pathology endpoint that could be amenable to therapeutic intervention. Schematic representation of the core AHR canonical pathway overlaid with concordant gene expression data (Padj<0.05) for experiment 1 (Fig 4) for active cases relative to healthy controls shows differential gene expression that includes core players AHR and the AHR nuclear translocator (ARNT) in the AHR pathway, as well as for key phase I metabolising enzymes (CYPB1, ALDH5A1, ALD3B1 and ALD3A2). The full pathway, including cross-talk between AHR and other signalling pathways that lead to noncanonical mechanisms of action of AHR and its ligands, overlaid with expression data from experiments 1 (S2 Fig) and 2 (S3 Fig), highlight a total of 28 concordant genes that all achieve differential gene expression at Padj<0.05. These demonstrate the interplay between the top IPA-identified canonical pathways, with AHR function influencing cell proliferation and estrogen receptor signalling pathways, while heme derivatives biliverdin and bilirubin are known to act as endogenous ligands for AHR [45, 46]. Identification of Mitotic Roles of Polo-Like Kinases as a top canonical pathway is indicative of cell proliferative activity that is consistent with CDC20 and CDK1 (Table 2) as major hub genes in the network (Fig 3), and with identification of the cyclin-dependent kinase inhibitor CDKN1A as the top inhibited upstream regulator (Activation z-score = -2.764; P = 5.4x10-26) in IPA.
Using Enrichr (S4 Table), signalling pathways involved in cell cycle predominated amongst the top pathways using the Reactome 2016 (“Cell cycle_Homo sapiens”, “Cell Cycle, Mitotic_Homo sapiens”, and multiple other pathways involved in cell cycle), WikiPathways 2016 (“Cell Cycle Homo sapiens”), KEGG 2016 (“Cell cycle_Homo sapiens”), and NCI-Nature 2016 (“Aurora B signalling”, “Aurora A signalling” and the “FOXM1 transcription factor network for Homo sapiens”, all of which play key roles in cell cycle progression) databases. CDK1 was also identified as the top PPI Hub Protein using Enrichr (S4 Table). Consistent with our top 10 “induced” gene list, other database comparisons using Enrichr (S4 Table) identified gene sets associated with erythrocyte function including “erythroid cell” (Jensen Tissues Table), “abnormal erythrocyte morphology” and multiple other erythrocyte-related phenotypes (MGI Mammalian Phenotype 2017), “CD71+Early Erythroid” (Human Gene Atlas), “congenital haemolytic anaemia” (Jensen Diseases), and “Haemoglobin’s Chaperone pathway” (BIOCARTA_2016).
Heatmaps were generated for individual expression levels for the top 10 concordant genes expressed at a higher level (Fig 5A), and the top 10 concordant genes expressed at a lower level (Fig 5B), in active cases compared to treated cases in experiment 1. Heatmaps were also generated for the same “induced” and “repressed” probes/genes in experiment 2 (Fig 5C and 5D). In this case, 6/10 and 7/10 top genes from experiment 1 were also in the top 10 most highly differentially expressed genes for “induced” and “repressed” gene sets in experiment 2, respectively, and all achieved fold-change >2 in both experiments. Amongst the 10 most “repressed” genes in experiments 1 and 2 were 3 genes also observed in the comparison of active cases with healthy controls: peptidase inhibitor 3 (PI3), as noted above known as an antimicrobial peptide for bacteria and fungi; ALPL which encodes an alkaline phosphatase that plays a role in bone mineralization; and CACNA2D3 which encodes the alpha2delta3 subunit of the voltage-dependent calcium channel complex. Of additional interest in this comparison were “repressed” genes: CHI3L1 which encodes a chitinase-like protein that lacks chitinase activity but is secreted by activated macrophages and neutrophils; EMR3 (ADGRE3) encoding an adhesion G protein-coupled receptor expressed predominantly in cells of the immune system and playing a role in myeloid-myeloid interactions during inflammation; and MMP25 that encodes matrix metallopeptidase 25 which inactivates alpha-1 proteinase inhibitor produced by activated neutrophils during inflammation thereby facilitating transendothelial migration of neutrophils to inflammatory sites. Of interest amongst the top 10 “induced” genes in both experiments were: CXCL10 encoding a chemokine of the CXC subfamily that is a ligand for CXCR3, binding to which results in stimulation of monocyte, natural killer and T-cell migration; IFNG encoding interferon-γ, well known for its role in macrophage activation for anti-leishmanial activity; and GBP1 that encodes a guanylate binding protein induced by interferon.
As there were only 42 concordant genes that achieved ≥2-fold change in expression, a more global picture of the impact of differential gene expression was obtained by performing IPA and Enrichr analyses using the full set of 210 genes represented by 221 probes that were concordant for differential gene expression at Padj ≤0.05. IPA network analysis indicated that 85 of these genes are joined in a single network (Fig 6), with IFNG as the major hub gene (i.e. with most connections to other genes in the network), and other major hub genes including STAT1, SPI1, RARA, NOTCH1 and MAPK3. The top canonical pathways included pathogenesis of multiple sclerosis (Table 2), consistent with interconnections between CXCL10/CXCL9/CXCL11 and major hub genes IFNG and STAT1 (Fig 6), and the Notch signalling pathway. Aryl hydrocarbon receptor signalling was also identified as a canonical pathway in this analysis at a nominal P = 0.003 (Table 2). Enrichment for chemokine signalling and Notch signalling pathways were also supported by analyses undertaken using Enrichr (Reactome 2016; WikiPathways 2016, and KEGG 2016 pathways; S5 Table). Consistent with this were top LINCS_L1000_Ligand_Perturbations_Up (S5 Table) for which perturbations of TNFA, IFNG, IL1, IFNA, and HGF were all significant at Padj<0.01. These ligand perturbations were all associated with differential expression at CXCL10, and commonly also at CXCL11, CXCL9, and STAT1. The major cell types associated with the treatment response were CD14+ monocytes and CD33+ myeloid cell populations (Human Gene Atlas; S5 Table).
As noted above, we found more differentially expressed probes between treated cases and controls, along with fewer differentially expressed probes between active and treated cases, in experiment 1 compared to experiment 2 (Table 1). We hypothesize that this is due to more effective treatment using liposome encapsulated amphotericin B in experiment 2 compared to the non-liposomal form of the drug employed during experiment 1. We therefore examined the genes that were discordant between active cases and treated cases across the two experiments to understand differences in the cure response. In support of the more efficient cure rate in experiment 2, 7/10 of the top “repressed” genes (namely: OLIG1, OLIG2, PTGDR2 alias GPR44, CCR3, CCL23, ALOX15, SLC29A1) identified as differentially expressed between active cases and treated cases in experiment 2 but not experiment 1 were the same genes that were most repressed in the concordant genes comparing active cases with healthy controls. In comparison, 0/10 of the top “repressed” genes identified as differentially expressed between active cases and treated cases in experiment 1 (but not experiment 2) matched the comparison of concordant genes for active cases and healthy controls. That is, treated cases in experiment 2 were behaving more like healthy controls than were treated cases in experiment 1.
To gain a more global picture of differential gene expression that might inform mechanistic differences in cure rates between the two therapeutic regimes, the 417 genes (from 443 probes) that were differentially expressed between active cases and treated cases in experiment 1 but not experiment 2, and the 988 genes (from 1096 probes) that were differentially expressed between active cases and treated cases in experiment 2 but not experiment 1, were analysed in Enrichr for gene-set enrichment. S6 and S7 Tables present details of the pathways and gene sets that contrast molecular events that characterise the two different treatment groups. These are summarised in Table 3.
For the 988 genes that were differentially expressed between active cases and treated cases in experiment 2 but not in experiment 1 (S6 Table) signalling pathways involved in cell cycle predominated amongst the top pathways using the Reactome 2016 (“Cell cycle_Homo sapiens”), Wiki Pathways 2016 (“Cell Cycle Homo sapiens”), KEGG 2016 (“Cell cycle_Homo sapiens”), and NCI-Nature 2016 (“Aurora B signalling”) databases. In every case there were multiple other pathways involved in cell cycle that achieved rank z-scores <-1 and Padj ≤0.01. This pattern recapitulates the results obtained in the earlier comparison of concordant genes for active cases and healthy controls (S4 Table), with CDK1 again identified as the top PPI Hub Protein for this gene set (S7 Table). Consistent with an enhanced rate of cure, multiple immune response signalling pathways (Table 3 and S6 Table) were also identified in this gene set, including IL-1, IL-3, IL-4, IL-6, IL-7 and IL-8 signalling pathways (Reactome 2016, Wiki Pathways 2016, and NCI Nature 2016 databases), Delta-Notch signalling (Wiki Pathways database), chemokine signalling (Wiki Pathways 2016 and KEGG 2016 databases) including specifically IL-8/CXCR2-mediated and CXCR4-mediated signalling (NCI Nature 2016 database), and Fc gamma R-mediated phagocytosis Homo sapiens (KEGG 2016 database). Of note, IL-4 was identified as the most significantly down-regulated perturbed ligand pathway (LINCS_L1000_Perturbed_Down; rank z-score -1.8, Padj 6.36x10-11) in this set of genes differentially expressed in active versus treated cases in experiment 2 but not experiment 1. None of these databases showed significant gene set enrichment when interrogated with the 417 genes identified as differentially expressed between active cases and treated controls in experiment 1 but not in experiment 2, i.e. they are not present in Table 3 or S7 Table which compare other enriched gene sets showing differences of interest between experiments 1 and 2. For example, all PPI Hub Proteins identified as significant for the 988 genes that were differentially expressed between active cases and treated cases in experiment 2 but not experiment 1 were related to cell cycle (S7 Table). In contrast, the 5 significant matches to gene sets for PPI Hub Proteins for the 417 genes differentially expressed between active cases and treated cases in experiment 1 but not experiment 2 included the inhibitor of NFκB NFKBIA and the SMAD-signalling pathway gene SMAD9 which transduces signals from members of the TGFβ family. Mutations in NFKBIA are associated with T-cell immunodeficiency [47]. SMAD9 (aliases SMAD8, SMAD8A, SMAD8B, SMAD8/9) transduces signals following ligation of TGFβ family members known as bone morphogenesis proteins (BMPs) to specific BMP (TGFβ family) receptors. Enrichr identified enrichment for a gene set matching genes differentially expressed in BMP4-treated cells (SILAC-Phosphoproteomic Database; P = 5.4x10-5, Padj = 0.003, rank z-score = -1.74) from the 417 but not the 988 genes (S7 Table). Another difference was enrichment of the “CD71+Early Erythroid” (Human Gene Atlas) gene set in the 417 genes, while the 988 genes were enriched for gene sets (S7 Table Human Gene Atlas database) associated with B lymphoblasts, CD105+ endothelial, CD33+ myeloid, and CD14+ monocytes but not erythroid cells. Overall these analyses of discordant gene sets between experiments 1 and 2 are consistent with our hypothesis that patients treated with a single dose of liposomal amphotericin B (experiment 2) were at a more advanced stage of cure at day 30 post treatment than patients treated with multi-dose non-liposomal amphotericin B (experiment 1).
In this study we have analysed whole blood transcriptomic data to further understand the pathogenesis of VL. One original goal of the study was to identify transcriptomic signatures that might differentiate asymptomatic infections from uninfected controls. In our attempt to achieve this we compared both modified quantiferon positive asymptomatic individuals and high antibody positive asymptomatic individuals with healthy endemic controls who were negative for these assays. In the event, we did not find signatures that would be diagnostic for either of these asymptomatic groups compared to negative controls. This was despite longitudinal epidemiological evidence from our study area showing that high antibody individuals are the group at most risk of progressing to clinical VL [6]. However, in that study we observed that high antibody individuals progressed to clinical VL within one year. In our study we selected individuals who had sustained high DAT titres for more than two annual surveys. Hence, we were effectively selecting for a subset of asymptomatic individuals who were resistant to progression to clinical disease. Similarly, there was no significant difference in the odds of progression to clinical disease in individuals who were positive by the modified quantiferon assay [6], and we found no evidence for a whole blood transcriptional signature to distinguish these individuals from uninfected endemic healthy controls. Our results therefore mirror those of Gardinassi and coworkers [27] who likewise found no significant differences in whole blood transcriptional signatures between asymptomatic individuals infected with L. infantum in Brazil, as determined by positive delayed type hypersensitivity to leishmanial skin-test antigen, and uninfected endemic controls. A more detailed longitudinal study will be required to detect transcriptional signatures early after exposure to L. donovani or L. infantum to identify signatures that may be predictive of progression to disease in asymptomatic individuals positive for antibody or cellular immunity to leishmanial antigens. In India it may be particularly interesting to identify signatures for those high titre DAT antibody individuals who progress to disease within 9 months from those who do not.
Our failure to identify signatures to detect asymptomatic infection meant that our attention focussed on understanding disease pathogenesis by comparing whole blood transcriptomes from active cases with all healthy controls, and in examining differences in the transcriptome following different regimens of drug treatment. In these comparisons 6 major themes emerged: (i) expression of genes and enrichment of gene sets associated with erythrocyte function in active cases; (ii) strong evidence for enrichment of gene sets involved in cell cycle in comparing active cases with healthy controls (or with more effective cure in experiment 2); (iii) identification of IFNG encoding interferon-γ as the major hub gene in concordant gene expression patterns across experiments comparing active cases with healthy controls or with treated cases; (iv) enrichment for interleukin signalling (IL-1/3/4/6/7/8) and a prominent role for CXCL10/9/11 and chemokine signalling pathways in the comparison of active cases with treated cases; (v) the novel identification of AHR signalling as a significant IPA canonical pathway identified from concordant gene expression patterns across experiments comparing active cases with healthy controls or with treated cases; and (vi) global expression profiling support for more effective cure at day 30 post-treatment with a single dose of liposomal encapsulated amphotericin B compared to multi-dose treatment over 30 days.
Interesting in our analysis of top differentially expressed genes and enriched gene sets/pathways between active cases and healthy controls was the predominance of gene sets associated with erythroid cells and function. A recent systematic review [48] found that anaemia has an overall prevalence higher than 90% in VL. Pathogenesis of anaemia based on clinical observations included the presence of anti-erythrocyte antibodies, dysfunction in erythropoiesis, and hemophagocytosis in spleen or bone marrow. Of these, the authors of this review conclude that hemophagocytosis is the most likely cause [48]. The results of our study indicate differential regulation of gene sets associated with abnormal erythrocyte morphology, erythropoiesis, erythrocyte physiology, erythrocyte osmotic lysis, along with decreased haematocrit, spherocytosis and reticulocytosis. The gene sets defining these erythrocyte phenotypes therefore suggest mechanisms other than just hemophagocytosis and could provide important signatures to monitor clinical cure. This is especially relevant given our observation that erythroid related genes were present amongst the discordant genes that were differentially expressed between active cases and cases treated with multi-dose amphotericin B (experiment 1) in which the degree of clinical cure was not as progressed for the same period of treatment with a single dose of liposomal amphotericin B (experiment 2).
Many of the individual cell-cycle and immune-related (e.g. Notch signalling, interleukin and chemokine signalling) signalling pathways that were perturbed in active cases relative to cured cases or healthy controls were also observed in the similar study of whole blood expression profiling carried out by Gardinassi and coworkers [27] in relation to VL caused by L. infantum in Brazil. However, a common feature of both the comparison of active cases with healthy controls, and of active cases with treated cases, in our study was the identification of IFNG encoding interferon-γ as the major hub gene. This was not itself surprising since interferon-γ plays a key role in activating macrophages to kill L. donovani parasites [49]. Studies across the leishmaniases have generally supported the notion that type 1 immune responses and the production of interferon-γ are vital for macrophage activation and parasite elimination [50–52]. It was interesting in our study that transcript levels for IFNG were higher in active cases than treated cases, where enhanced interferon-γ responses might have been expected to accompany drug cure. Nonetheless, it concurs with our observations that CD4+ T cells in whole blood from active VL patients and treated patients secrete high levels of interferon-γ following stimulation with crude Leishmania antigen [53, 54], the difference being that only active VL cases secreted IL-10 concurrently with interferon-γ [54]. The higher transcript abundance for IFNG in active compared to treated cases in our study suggests return to baseline with treatment in the latter. Gardinassi and coworkers similarly found higher transcript levels for IFNG in active compared to treated cases [27].
Accompanying the central role of IFNG as a hub gene when comparing active cases with treated cases was evidence for perturbation of multiple cytokines, including IFNG, IFNA, IL-1, IL-6, and TNF, all of which were supported by differentially expressed gene signatures that generally included CXCL10/11/9 and STAT1. This CXCL10/11/9 chemokine gene expression signature also accounted for the identification of “pathogenesis of multiple sclerosis” [55] as the top disease-related canonical pathway identified using IPA, consistent with a proinflammatory response contributing to disease pathology in active VL. “Pathogenesis of multiple sclerosis” was also identified as a top canonical pathway in spleen tissue and splenic macrophages from L. donovani infected hamsters [26], a study in which the authors also noted high interferon-γ expression that was ineffective in directing macrophage activation and parasite killing. STAT1 is a transcription factor activated by ligation of interferon-γ receptors. CXCL10/11/9 are all induced by interferon-γ, all bind to CXCR3, and between them have multiple roles as chemoattractants for monocytes and macrophages, T cells, NK cells, and dendritic cells, and in promoting T cell adhesion. CXCL10 and CXCL9 were also identified as the most highly “induced” genes in comparing lesion transcript profiles with normal skin of patients with American cutaneous leishmaniasis, consistent with their roles in inflammatory cell recruitment [28]. Cxcl9, Gbp1 (encoding the interferon-γ-induced guanylate binding protein GBP1 identified here as one of the top 10 induced genes when comparing active versus treated cases), and Ifng were also identified as part of a common signature of 26 genes upregulated in blood, spleen and liver throughout the course of experimental infection with L. donovani in susceptible BALB/c mice, with Cxcl9 and Gbp1 reported as hub genes from a STRING analysis [24].
Given the many studies that have identified the importance of regulatory IL-10 in VL pathogenesis [54, 56–59], it was of some interest in our study that IL10 was not identified as a top differentially expressed gene or as a significantly enriched signalling pathway in either comparison of active cases with healthy controls, or of active cases with treated cases. Nor did we observed perturbation of IL10R as has been reported in experimental transcriptional profiling studies of VL [24]. Indeed, downregulated expression of the type 2 cytokine gene IL4 was the strongest response associated with effective cure in liposome-encapsulated amphotericin B treated cases, in line with previous studies showing that IL-4 levels were two-fold higher in VL patients who had failed treatment compared to previously untreated patients, whereas IL-10 levels were comparable in both [58].
One novel observation of our study was identification of AHR signalling as the top canonical pathway when comparing transcriptomes between active cases and healthy controls or treated cases. Through crosstalk between signalling pathways, AHR ligands have been shown to significantly induce IL-10 secretion and inhibit IL-1β and IL-6 production in dendritic cells, and to promote IL-10 production and suppress IL-17 expression in CD4+ T cells [42–44]. IL-17 is a potent activator of neutrophils, both through lineage expansion and through their recruitment by regulating chemokine expression. While IL-17 perturbation was not identified in our whole blood transcriptional profiles associated with human VL, evidence from murine models [60] demonstrate a strong role for IL-17 and neutrophils in parasite clearance from liver and spleen. Duthie and coworkers [59] have shown that both IL-10 and IL-17 cytokines are elevated in the serum of active VL patients, reverting to baseline levels with standard antimonial treatments. AHR activation has also been shown to inhibit inflammation through upregulation of IL-22 [61], another cytokine that has been shown to be significantly higher in Leishmania antigen stimulated peripheral blood mononuclear cells from active VL cases compared to treated cases [62]. AHR activation during VL may underpin the complex regulation of pro- and anti-inflammatory responses during disease pathogenesis and during response to therapy.
Of potential translational importance in our study was the additional identification of AHR signalling pathway at the top of the Ingenuity “Top Tox List” indicative of its role as a toxic pathology endpoint that could be amenable to therapeutic intervention. AHR locates to the cytoplasm in a stable complex that includes HSP90 observed as a differentially regulated gene in our comparison of active cases with healthy controls. Ligand binding occurs in the cytoplasm and triggers AHR translocation to the nucleus where it binds with ARNT to act as a transcription factor. Both AHR and ARNT were differentially expressed between active VL cases and controls in our study. The AHR response was first associated with xenobiotic induction of metabolizing enzymes, such as the induction of cytochrome P450, family 1, subfamily A, polypeptide 1 (Cyp1a1) following exposure to the polychlorinated dibenzo-p-dioxin 2,3,7,8-Tetrachlorodibenzo-p-dioxin [63]. Multiple AHR ligands are known to induce a “gene battery” of metabolizing enzymes involved in oxidative stress response, cell cycle and apoptosis [64], amongst which are CYP1B1, ALD3B1, ALD3B and ALDH5A1 that were differentially expressed between active VL cases and healthy controls. Transcriptomic profiling of M. tuberculosis infected macrophages uncovered evidence for the generation of endogenous AHR ligands through induction of enzymes controlling tryptophan catabolism [65]. The generation of endogenous AHR ligands may likewise explain the role of AHR signalling in VL. For example, heme derivatives biliverdin and bilirubin have both been shown to act as endogenous ligands for AHR, as have arachidonic acid metabolites such as prostaglandins and leukotrienes [45, 46]. The former would be consistent with the strong perturbation of erythrocyte function between active VL cases and controls observed in our study. Importantly, addition of exogenous AHR ligands enhanced M. tuberculosis infection associated AHR transactivation to stimulate expression of AHR target genes, including IL-1β and IL-23 which stimulate T cell subsets to produce IL-22. This suggests that administration of exogenous ligands could be used as a therapeutic intervention, especially in the knowledge that different exogenous AHR ligands can modulate either regulatory T cell or inflammatory T helper 17 cell differentiation in a ligand-specific fashion to suppress or exacerbate autoimmune disease [66].
One of the potential benefits of gene expression profiling is the identification of gene signatures that could be used in the diagnosis of disease and in the monitoring of treatment efficacy. In this respect it is remarkable that 9 of the top 10 DEGs found to be more highly expressed in active cases compared to healthy controls in our study (i.e. all except DARC) were also found to be more highly expressed in active cases compared to healthy controls in the Brazilian whole blood expression profiling study of L. infantum [27]. Similarly, 5 (PI3, CCR3, OLIG1, CACNA2D3, ALPL) of the top 10 DEGs found to be reduced in expression in active cases relative to healthy controls were also found to be reduced in expression in active L. infantum cases. More extensive cross-matching of the gene lists from the two studies identifies larger sets of concordant genes to be used as signatures for VL disease that cross the divides of geography and species and could be tested in other regions endemic for VL disease. In relation to treatment monitoring, 6 (CXCL10, ANKRD22, MT1G, IFNG, GBP1, SEPT4) of the top 10 genes retaining higher expression in cases relative to treated cases were also concordant across the two studies. While there was no concordance for the top 10 genes expressed at reduced levels in active cases compared to treated cases, this could reflect the effect of different treatment protocols (pentavalent antimony in Brazil versus two forms of Amphotericin B in India). As we observed in our comparison of different Amphotericin B treatment strategies, the rate of return to control levels of gene expression differs across treatments. Nevertheless, in our study we observed some concordance (PI3, CACNA2D3, ALPL) between the genes that were more highly expressed in all treated cases and healthy controls relative to active cases. A signature that combines gene markers of active disease with genes that represent return to healthy baseline would be valuable in the monitoring of treatment efficacy.
Overall, our study has made some novel observations in relation to gene signatures that accompany both active VL disease and clinical cure in treated cases that could provide translatable targets for the development of novel or drug repurposed therapeutic interventions. Furthermore, by studying in more detail the discordant gene patterns that accompanied treatment with single dose liposome encapsulated amphotericin B versus multi-dose non-liposomal amphotericin B we were able to define gene signatures that could be used to monitor progress towards clinical cure.
|
10.1371/journal.pntd.0004014 | Development of a Fluorescence-based Trypanosoma cruzi CYP51 Inhibition Assay for Effective Compound Triaging in Drug Discovery Programmes for Chagas Disease | Chagas disease, caused by the protozoan parasite Trypanosoma cruzi (T. cruzi), is a life threatening global health problem with only two drugs available for treatment (benznidazole and nifurtimox), both having variable efficacy in the chronic stage of the disease and high rates of adverse drug reactions. Inhibitors of sterol 14α-demethylase (CYP51) have proven effective against T. cruzi in vitro and in vivo in animal models of Chagas disease. Consequently two azole inhibitors of CYP51 (posaconazole and ravuconazole) have recently entered clinical development by the Drugs for Neglected Diseases initiative. Further new drug treatments for this disease are however still urgently required, particularly having a different mode of action to CYP51 in order to balance the overall risk in the drug discovery portfolio. This need has now been further strengthened by the very recent reports of treatment failure in the clinic for both posaconazole and ravuconazole. To this end and to prevent enrichment of drug candidates against a single target, there is a clear need for a robust high throughput assay for CYP51 inhibition in order to evaluate compounds active against T. cruzi arising from phenotypic screens. A high throughput fluorescence based functional assay using recombinantly expressed T. cruzi CYP51 (Tulahuen strain) is presented here that meets this requirement. This assay has proved valuable in prioritising medicinal chemistry resource on only those T. cruzi active series arising from a phenotypic screening campaign where it is clear that the predominant mode of action is likely not via inhibition of CYP51.
| Chagas disease, caused by the parasite Trypanosoma cruzi (T. cruzi), is endemic in Latin America and emerging in North America and Europe through human migration. It is a severe global health problem with 8–10 million people infected and an estimated 12,000 deaths annually. Current treatment options are poorly efficacious and have severe side effects. New drugs are therefore urgently required. Two of these potential new drugs, posaconazole and ravuconazole, both targeting an enzyme in T. cruzi called CYP51, have recently failed in clinical development. Therefore, in light of these recent clinical failures and in order to better balance the overall risk in the drug discovery portfolio for Chagas disease, it has become prudent to assess whether new chemical start points for drug discovery programmes have a mode of action predominantly driven by T. cruzi CYP51 inhibition. In this paper we report a fluorescence based assay to determine whether compounds inhibit T. cruzi CYP51. This provides a high throughput screen to help prioritise medicinal chemistry resource on those T. cruzi active new chemical series that do not have a mode of action predominantly driven by CYP51 inhibition.
| Chagas disease is a tropical parasitic disease caused by the flagellate eukaryotic (protozoan) parasite Trypanosoma cruzi (T. cruzi), endemic in Latin America and now emerging in North America and Europe through human migration. It is becoming a severe global health problem with approximately 8–10 million people infected, an estimated 12,000 deaths per year, and placing 100 million people at risk. Transmission to humans and other mammals is predominantly by an insect vector, the blood-sucking "kissing bugs" of the subfamily Triatominae (family Reduviidae) [1]. Transmission has also been reported to occur through contaminated food, blood transfusions and from mother to child.
Clinical Chagas disease can be classified into two distinct phases, acute and chronic. In the acute phase, lasting a few weeks, parasites begin to multiply in the organs and tissues. Symptoms are usually mild and non-specific with patients rarely being diagnosed. However, life-threatening myocarditis or meningoencephalitis can occur during the acute phase with a death rate for people in this phase of about ten percent. Ten to fifty percent of infected survivors develop chronic Chagas disease. People in the chronic phase can be asymptomatic for many years, with parasites generally undetectable in the blood. However, the disease causes organ and tissue damage, particularly potentially lethal cardiopathy and megacolon or megaoesophagus, caused by the sequential induction of inflammatory response to the parasite. Nitroheterocyclic compounds, benznidazole and nifurtimox, developed in the 1960’s [2], are currently the only two drugs used for the treatment of Chagas disease. Both have low efficacy in the chronic stage and, with prolonged dosing regimens, both drugs have significant side effects including skin irritation, neurotoxicity, and digestive system disorders [3]. Newer, safer and more efficacious treatments are therefore in desperate need.
Inhibition of sterol 14α-demethylase (CYP51) has been considered a viable target against T. cruzi for over 30 years [2,4,5,6,7,8]. Found in a broad variety of organisms including animals, plants, fungi and protozoa, this enzyme plays an essential role in the sterol biosynthetic pathway, catalysing the oxidative removal of the 14α-methyl group from sterol precursors such as lanosterol or eburicol [9]. The products of the pathway, cholesterol in humans or ergosterol in fungi, are required for the integrity of the eukaryotic cell membrane. These sterols are required for membrane function in T. cruzi. Inhibition of CYP51 activity is lethal as the T. cruzi parasites are unable to scavenge and utilise host cholesterol [10]. The CYP51 gene is known to be expressed in all stages of the parasite life cycle and indeed it has also been shown to be up-regulated in multiplying forms [9]. As with other members of the Cytochrome family, CYP51 is a haem containing protein located on the membrane of the endoplasmic reticulum that relies upon electron transfer by NADPH reductase for activation [11].
Azole inhibitors, which interfere with sterol biosynthesis, essential in eukaryotic cells, have already been used with success in humans in the treatment of fungal infections. Several of these drugs have been considered as possible treatments for Chagas disease [12]. Ketoconazole, fluconazole, itraconazole, ravuconazole and posaconazole are known to inhibit CYP51 in vitro, competitively binding to the haem within CYP51 and occupying the active site preventing any substrate from binding. Although ketoconazole and itraconazole have not demonstrated significant curative activity in humans with chronic Chagas disease [6], other azoles, with greater potency and improved pharmacokinetic properties, which have been shown to have potent activity against T. cruzi, including posaconazole [13,14] and ravuconazole (Fig 1), are in clinical development with the Drugs for Neglected Diseases initiative (DNDi).
To prevent enrichment of candidates against a single target, and thus reduce risk in the overall drug discovery portfolio for Chagas disease, it has therefore become necessary to evaluate and prioritise medicinal chemistry resource on new chemical series active against T. cruzi but with such activity not likely driven via T. cruzi CYP51 inhibition. Recent findings from clinical trials with posaconazole [15] and ravuconazole [16] has indicated re-emergence of parasitaemia in two thirds of patients once dosing has been completed, thus reinforcing the need to strengthen the overall drug discovery portfolio for Chagas disease with new chemical lead series not working via this mechanism of action.
Evaluating compounds as potential inhibitors of T. cruzi CYP51 has previously been demonstrated measuring the apparent dissociation constants (Kd) by spectral titration [4,17,18] utilising the shift of the haem iron soret band in response to binding [18]. One of the drawbacks to this methodology is that micromolar protein concentrations are required for screening causing potential interference with the optical properties and/or solubility of test compounds [4]. There are many potential reasons why affinity estimates measured by binding may not correlate with functional inhibition. These include allosteric sites, non or uncompetitive modes of inhibition or slow kinetics [19].
Inhibition of endogenous substrate lanosterol, eburicol and obtusifoliol has also been used as an in vitro tool using recombinant expressed human CYP51 enzyme [20]. In particular, measuring effect on CYP51 driven metabolism of lanosterol to follicular fluid meiosis activating sterol (FF-MAS) in the presence of test substances [21] is well established (Fig 2). However, FF-MAS detection requires mass spectrometry limiting the number of compounds that can be tested and consequently limiting the value of such an assay for triaging large numbers of phenotypic screening T. cruzi hits toward identifying modes of action away from CYP51.
Metabolism of fluorogenic probe substrate to a product, detectable by fluorescence is well established with recombinantly expressed cytochrome P450 enzymes (CYP’s) for the purpose of assessing possible drug-drug interactions [22]. Measuring CYP inhibition by this method provides a high throughput screening approach, avoiding time consuming analysis by mass spectrometry and minimising use of expensive substrates. The O-dealkylation of Vivid substrate benzyloxymethylocyanocoumarin (BOMCC) to fluorescent product cyanohydroxycoumarin (CHC) is commonly used to evaluate CYP3A4 activity in recombinantly expressed membrane preparations (Fig 2). Valuably, O-dealkylation activity in the presence of recombinantly expressed T. cruzi CYP51 was observed. This has enabled the creation of a fast, high-throughput, 96 and 384 well microtitre method to assess the inhibitory potential of compounds against T. cruzi CYP51, which is described in this paper.
Benzyloxymethyloxycyanocoumarin (BOMCC) was obtained from ThermoFisher Scientific. Fluconazole, ketoconazole, itraconazole, NADP, NADPH, glucose-6-phosphate, glucose-6-phosphate dehydrogenase, Cytochrome C from horse heart and sodium bicarbonate were obtained from Sigma Aldrich. 50 mM potassium phosphate (pH 7.4) buffer was prepared from dibasic and monobasic forms of potassium phosphate obtained from Sigma Aldrich.
E. coli membrane fractions containing the T. cruzi CYP51 (Tulahuen strain) (Bactosomes) were provided by Cypex Ltd.
The cDNAs coding for T. cruzi CYP51 and NADPH P450 reductase were synthesized and supplied in pUC vectors by Genescript. The CYP51 cDNA was cloned into the expression vector pCW with an ompA N terminal leader and the reductase was cloned into a pACYC184 derived expression vector with a pelB N terminal leader. E. coli JM109 was used for the co-expression of the recombinant proteins.
For each concentration of test compound the rate of fluorescence units per minute was calculated as a percentage of the average rate of the solvent only control wells.
The percentage of solvent control values was then plotted against the concentration range. Using the following 4 parameter fit equation, an IC50 value can be determined.
Accession number for T. cruzi cDNA is AY283022; T. cruzi reductase is DQ857724 and mosquito reductase is AY183375.
Despite the presence of the N terminal pelB leader sequence, the recombinant T. cruzi NADPH P450 reductase proved to be a cytosolic protein (as might be expected from previous data [10]) and was not present in the E.coli membrane fraction with the T. cruzi CYP51. Reconstitution of the cytosolic fraction containing the T. cruzi reductase with the membrane fraction containing the T. cruzi CYP51 did not result in active CYP51 so, in order to generate an active T. cruzi CYP51 system, the T. cruzi CYP51 was co-expressed with human (expression construct supplied by Cypex Ltd) or mosquito (expression construct supplied by the Liverpool School of Tropical Medicine) NADPH P450 reductase. Screening of CYP51 activity showed that the T. cruzi CYP51 co-expressed with the mosquito reductase gave the highest rate of BOMCC turnover. The mosquito reductase expression cassette was cloned from the pACYC vector into the pCW based CYP51 expression vector to allow both proteins to be expressed from the same plasmid resulting in a relatively higher level of mosquito reductase in the membrane fraction with a concomitant increase in BOMCC turnover. Supplementation of the T. cruzi CYP51 / mosquito NADPH P450 reductase bactosomes with partially purified mosquito cytochrome b5 at 10 fold excess over the CYP51 resulted in a further increase in BOMCC turnover. This bactosome preparation was then scaled up for all further work.
To determine kinetic parameters, 100 μM BOMCC O-dealkylase activity was assessed in incubations containing 25, 50, 75, 100, 150, 200, 250, 300, 400, 500 and 600 pmoles/mL T. cruzi CYP51 enzyme. Protein concentrations were normalised using control protein containing no CYP activity. Incubations were pre-warmed at 37°C before addition of NADPH regenerating system. Production of CHC was measured at 1 minute intervals at 37°C (Exc 410 nm, Em 460 nm) for 10 minutes (Fig 3a).
Km and Vmax parameters for BOMCC were determined by incubating 37 pmoles/mL T. cruzi CYP51 enzyme (0.24 mg/mL bactosomes) with 25, 50, 75, 100, 150, 200 and 250 μM BOMCC. Incubations were pre-warmed at 37°C before addition of NADPH regenerating system. Production of CHC was measured at 1 minute intervals at 37°C (Exc 410 nm, Em 460 nm) for 10 minutes (Fig 3b). Rates of BOMCC metabolism appeared linear up to 100 pmoles/mL with a Km value determined as 191± 55 μM. Assay conditions of 37 pmoles/mL of T. cruzi CYP51 in bactosomes and 100 μM BOMCC provided a good dynamic range of metabolism.
Metabolic activation of 100 μM BOMCC by T. cruzi CYP51 (37 pmoles/mL) was evaluated in the presence of 2% v/v DMSO, methanol or acetonitrile. Incubations were pre-warmed at 37°C before addition of NADPH regenerating system. Production of CHC was measured at 1 minute intervals at 37°C (Exc 410 nm, Em 460 nm) for 10 minutes. The average rate of fluorescence units (AFU) per minute was compared to a solvent only control (Fig 4a). There did not appear to be a significant decrease in metabolic activity in the presence of 2% v/v DMSO or 2% v/v methanol.
With kinetic parameters established, a microtitre-based assay to measure inhibition of recombinantly expressed T. cruzi CYP51 by test compounds was then implemented. Using a selection of ‘azole’ inhibitors, including those previously shown to competitively bind to CYP51, and known xenobiotic CYP450 inhibitors (Table 1), the assay was validated to ensure robustness. The incubation times and protein concentrations employed were within the linear range for each assay. Incubations containing 37 pmoles/mL of T. cruzi CYP51 in bactosomes, 100 μM BOMCC, 10, 3.3, 1.0, 0.33, 0.1, 0.033 or 0.01 μM test compound (2% v/v solvent) in 50 mM potassium phosphate buffer (pH 7.4), were pre-incubated at 37°C for 5 minutes. Upon addition of NADPH regenerating system (7.8 mg [28 μM] glucose-6-phosphate, 1.7 mg [2.2 μM] NADP, 6 units/mL glucose-6-phosphate dehydrogenase in 2% w/v sodium bicarbonate buffer) production of fluorescent metabolite CHC was measured (Exc 410 nm, Em 460 nm) at 1 minute intervals over a 10 minute period. Rates of metabolite production per minute were compared to uninhibited controls and plotted against test compound concentration to obtain an IC50. The assay was then miniaturized to 20 μL and moved from kinetic to single endpoint stopped readout (60 minute reaction before quenching with posaconazole to generate a long stable signal) to increase throughput.
Ketoconazole, itraconazole, posaconazole and miconazole all showed potent inhibition of T. cruzi CYP51 activation of substrate BOMCC with IC50 values of 0.014, 0.029, 0.048 and 0.057 μM, respectively. Indeed, it is likely that these compounds are even more potent than they appear to be as the T. cruzi CYP51 enzyme concentration used in the assay defines a minimum IC50 of approximately 0.02 μM. Fluconazole also inhibited activity (IC50 0.88 μM) although this was at least 20 fold less inhibitory than observed with other azole type compounds. Methimazole did not appear to inhibit activity (IC50 > 8 μM), perhaps as a result of an inability to adequately fill the active site and prevent substrate binding.
Compounds which did not appear to inhibit T. cruzi CYP51 in this assay included known CYP450 inhibitors ticlopidine (CYP2C19), sulphaphenazole (CYP2C9), sulphamethoxazole (CYP2C8/9), quinidine (CYP2D6) and furafylline (CYP1A2). This was expected as, unlike other drug metabolising CYP450s, CYP51 has a very narrow substrate specificity, being limited to endogenous sterols including eburicol and lanosterol [23].
As previously discussed, addition of an equivalent reductase enzyme (mosquito, A. gambiae) was required to deliver metabolic activity of T. cruzi CYP51. Due to the artificial nature of this pairing it was therefore necessary to confirm that any potent inhibition of CHC production was not the result of indirect inhibition of the electron transfer by NADPH Reductase. Thus, cytochrome c reductase activity of the bactosome preparation was monitored in the presence of all test inhibitor compounds individually. Incubations containing 0.82 pmoles/mL of T. cruzi CYP51 in bactosomes, 50 μM cytochrome c, 10, 3.3, 1.0, 0.33, 0.1, 0.033 or 0.01 μM test compound (2% solvent) in 50mM potassium phosphate buffer (pH 7.4), were pre-incubated at 37°C for 5 minutes. Absorbance at 550 nm was then measured for 3 minutes to ascertain a background level. After addition of NADPH (final concentration 80 μg/mL), absorbance was further measured over 5 minutes. The rate of reduction of cytochrome c at each test compound concentration was compared to an uninhibited control (Fig 4b). The assay was then miniaturized to 50 μL and moved from kinetic to single endpoint readout (60 minutes reaction) to increase throughput. No decrease in activity of cytochrome c reductase was observed for any of the test inhibitors. Furthermore, subsequent evaluation of a much larger set of T. cruzi CYP51 inhibitors has since been carried out and none have yet decreased cytochrome c reductase activity confirming that inhibition of BOMCC O-dealkylation observed is by direct inhibition of T. cruzi CYP51 and not by its effect on the cytochrome c reductase nor the regeneration system. This strongly suggests that the artificial nature of the pairing of T. cruzi CYP51 with A. gambiae NADPH reductase in the bactosomes used for this assay will not deliver false positives.
Identifying the fluorogenic probe BOMCC has enabled development of a fluorescence based high throughput functional assay to determine T. cruzi CYP51 inhibition, suitable for compound triaging of T. cruzi phenotypic screening hits. This assay is now embedded in our Chagas disease screening cascade to prioritise compound series for progression into hit to lead and lead optimisation.
Following a phenotypic screen of approximately 200,000 compounds against the T. cruzi parasite, inhibition of T. cruzi CYP51 was measured for compounds from 22 trypanocidal series and 16 singleton compounds. Compounds were derived from a variety of sources including diversity and target-centric libraries from the Dundee Drug Discovery Unit and GSK corporate collections. At least three compounds from each chemical series displaying as wide a range of T. cruzi activity as possible were tested (a total of 129 compounds; Fig 5a).
With the assumption that those series which show a good correlation between T. cruzi activity and T. cruzi CYP51 enzyme inhibition have a T. cruzi CYP51 mediated mode of action, a remarkable enrichment for the T. cruzi CYP51 mode of action was observed. Correlation between T. cruzi activity and T. cruzi CYP51 inhibition was evident for 11 out of the 22 hit series, with an additional two series showing sporadic association with T. cruzi CYP51 inhibition. Similarly high rates were observed with the singletons, with 11 out of 16 showing antiparasitic activity in line with T. cruzi CYP51 inhibition. 13/22 series and 11/16 singletons were therefore de-prioritised from further medicinal chemistry resourcing and focus given to those hits demonstrating at least one log unit divergence between T. cruzi activity and T. cruzi CYP51 activity, providing confidence that the predominant mode of action was not T. cruzi CYP51 driven. It is clearly important to look at the overall correlation between inhibition of T. cruzi CYP51 and antiparasitic activity for a given series to ensure that they are not tracking.
As it transpires, the majority of the chemical series removed from progression bear aza-heterocyclic groups which are already well known to exert inhibition of cytochromes through co-ordination to the haem iron (Fig 6) [9] and could have perhaps been eliminated “by- eye”. However, the great benefit of this assay to our drug discovery effort is that it has allowed identification of T. cruzi active series which do contain potential cytochrome binding motifs, but which do not inhibit T. cruzi CYP51. For instance, two particularly promising series (Fig 5b) having potent T. cruzi activity were found not to have an obvious correlation with T. cruzi CYP51 inhibition in spite of containing heterocycles with known potential for haem-binding in cytochromes. This was in stark contrast to a third series (Fig 5c) where T. cruzi activity tracked with T. cruzi CYP51 inhibition, akin to the known human CYP51 training set. The work described here has therefore enabled us to advance the first two series into hit-to-lead studies with a greater degree of confidence and oral activity with compounds from both series has subsequently been demonstrated in mouse models of Chagas disease.
The 96 well plate based assay used in this analysis has now been further miniaturized and adapted to automated liquid dispensers and microplate readers for high-throughput screening in 384 well endpoint format. The T. cruzi CYP51 FLINT assay (Fig 7a) can be run in 20 μL and the cytochrome c reductase absorbance assay (Fig 7b) in 50 μL final volumes. This allows a throughput increase to up to more than 10,000 wells in a single experiment with Z’ > 0.7 and with a significant reduction in screening cost and time. Sensitivity to described inhibitors was maintained during assay miniaturization.
|
10.1371/journal.pntd.0001369 | Larval Development of Aedes aegypti and Aedes albopictus in Peri-Urban Brackish Water and Its Implications for Transmission of Arboviral Diseases | Aedes aegypti (Linnaeus) and Aedes albopictus Skuse mosquitoes transmit serious human arboviral diseases including yellow fever, dengue and chikungunya in many tropical and sub-tropical countries. Females of the two species have adapted to undergo preimaginal development in natural or artificial collections of freshwater near human habitations and feed on human blood. While there is an effective vaccine against yellow fever, the control of dengue and chikungunya is mainly dependent on reducing freshwater preimaginal development habitats of the two vectors. We show here that Ae. aegypti and Ae. albopictus lay eggs and their larvae survive to emerge as adults in brackish water (water with <0.5 ppt or parts per thousand, 0.5–30 ppt and >30 ppt salt are termed fresh, brackish and saline respectively). Brackish water with salinity of 2 to 15 ppt in discarded plastic and glass containers, abandoned fishing boats and unused wells in coastal peri-urban environment were found to contain Ae. aegypti and Ae. albopictus larvae. Relatively high incidence of dengue in Jaffna city, Sri Lanka was observed in the vicinity of brackish water habitats containing Ae. aegypti larvae. These observations raise the possibility that brackish water-adapted Ae. aegypti and Ae. albopictus may play a hitherto unrecognized role in transmitting dengue, chikungunya and yellow fever in coastal urban areas. National and international health authorities therefore need to take the findings into consideration and extend their vector control efforts, which are presently focused on urban freshwater habitats, to include brackish water larval development habitats.
| Aedes aegypti and Aedes albopictus mosquitoes transmit arboviral disease like dengue and chikungunya that are of international concern. Control of dengue and chikungunya presently focuses on eliminating freshwater larval development habitats of the two mosquitoes in urban surroundings. We investigated the ability of the two mosquito species to lay eggs and undergo development into larvae, pupae and adults in brackish water, and examined brackish water collections in the peri-urban environment for the presence of larvae. The results confirmed their ability to lay eggs and for the eggs to develop into adults in brackish water. Their larvae were found in brackish water in discarded food/beverage containers and abandoned boats as well as disused wells. Such brackish water collections with larvae in Jaffna city, Sri Lanka were found near areas of high dengue incidence. This hitherto unappreciated potential contribution to arboviral disease transmission in urban areas is of global significance. National and international health authorities need to take these new findings into consideration in developing appropriate strategies for controlling diseases transmitted by the two mosquito species.
| Aedes aegypti (Linnaeus) (Diptera: Culicidae) is the principal tropical mosquito vector of arboviruses causing yellow fever, dengue and chikungunya [1]–[3]. The related Aedes albopictus Skuse is a secondary vector of dengue and chikungunya [1]–[4]. Unlike Ae. aegypti, Ae. albopictus possesses a diapausing egg stage to survive winters which has enabled it to spread to temperate regions and cause a chikungunya epidemic in northern Italy in 2007 [4].
Dengue is the most common arboviral disease of humans, with 50 million annual cases in more than 100 countries, an increasing incidence and spread worldwide, and 2.5 billion people at risk [5], [6]. About 500,000 persons require hospitalization every year for dengue hemorrhagic fever (DHF) and 2.5% of DHF cases are fatal [6]. Dengue is endemic in Sri Lanka with a high incidence in the northern and eastern districts of Jaffna and Batticaloa respectively [7]. Sri Lanka moreover experienced an epidemic of chikungunya in 2006–2007 [8]. Dengue is also endemic in Brunei and many other Southeast Asian countries [5], [9], [10]. There is presently no licensed vaccine or specific anti-viral drug for dengue [6], [9].Yellow fever, another flaviviral disease, is endemic in Africa and South America, has a zoonotic reservoir and is responsible for 200,000 cases and 30,000 deaths worldwide [11]. An effective vaccine is available against yellow fever, but it can potentially spread to Asia through increased global transport. Chikungunya, caused by an alphavirus, is endemic in Southeast Asia and has produced recent epidemics in Africa and South Asia [1], [2], [12]. Additional arboviral diseases with animal reservoirs are emerging as serious threats to human health [1], [13].
Ae. aegypti and Ae. albopictus have adapted to feed on humans and undergo larval and pupal development in natural ( e.g. rock pools, tree holes, leaf axils) and artificial (e.g. water tanks, blocked drains, decorative pots and discarded tyres and food/beverage containers) freshwater collections in the urban and peri-urban environment [1], [3], [5]. Ae. aegypti and Ae. albopictus are common anthropophagic mosquitoes in urban areas of Sri Lanka [14], [15]. Control of dengue and chikungunya in tropical countries is mainly achieved through surveillance for Ae. aegypti and Ae. albopictus larvae and eliminating larval development habitats, with the attendant use of insecticides and public education [5], [9], [10]. Larval control efforts invariably focus on freshwater development habitats of the two mosquito species near human habitations [5], [9], [10]. We investigated a hypothesis that Ae. aegypti and Ae. albopictus also undergo preimaginal development in brackish water in the peri-urban environment (water with <0.5 ppt or parts per thousand, 0.5–30 ppt and >30 ppt salt are termed fresh, brackish and saline respectively), with the potential thereby to make a hitherto unrecognized contribution to the transmission of dengue, chikungunya and other arboviral diseases.
Head of each household was explained the purpose and the objectives of the study and his/her informed consent was obtained orally to inspect wells for the presence of Aedes larvae and to measure the salinity of well water.
Aedes eggs and larvae for experimental studies were collected in Thirunelvely (inland, 9°41′ 03.49″N: 80° 01′ 14.49″E) and Batticaloa town (coastal, 7°43′ 35.81″N: 81° 42′ 4.04″E) in the Sri Lankan districts of Jaffna and Batticaloa respectively. Coastal areas of Jaffna city in the Jaffna district and Thannamunai (7°46′ 06.15″N: 81° 36′ 33.11″E) in the Batticaloa district and specific inland locations in the Jaffna district, were surveyed for larvae in brackish water habitats (Figure 1).
Black plastic ovitraps prepared as described previously [14] were used to collect Aedes mosquito eggs and larvae from ten randomly selected houses in Thirunelvely and Batticaloa. Four ovitraps were placed in each house. Each ovitrap, with a capacity of 250 ml, contained 100 ml of tap water and 3×10 cm plywood paddle resting against the upper rim. Ovitrap collections were done every two weeks from August to December 2008. Collected eggs and larvae from Batticaloa and Thirunelvely were brought to the Zoology Laboratories at the Eastern and Jaffna Universities respectively and reared under laboratory conditions (28±2°C, 70–80% R.H.). Larvae along with eggs in each ovitrap were reared separately in yoghurt cups (3.5×6.5 cm) containing 50 ml of tap water. Powdered fish meal pellets were provided twice a day as larval food. The emerging adults were identified using standard keys [16]. Adults identified as Ae. aegypti and Ae. albopictus were pooled separately and used to establish self-mating colonies that were fed on rabbit blood. The resultant larval progenies were used for subsequent salinity tolerance tests.
First and early third instar larvae of Ae. aegypti and Ae. albopictus were exposed to different salinity levels of 0, 2, 4, 6, 8, 10, 12, 14, 16, 18 and 20 ppt in the two laboratories . Required salinities were obtained by adding tap water (<0.5 ppt salinity) to sea water and the salinities measured with a refractor-salinometer (Atago, Japan). Twenty larvae in 150 ml capacity plastic containers containing 100 ml of water of different salinities were maintained at room temperature (28±2°C) until their emergence to adults. Plastic lids were used to partially cover the containers to minimise evaporation. Larvae were fed twice daily with powered fish meal. Three replicate tests were run in parallel for each level of salinity. Although experiments were conducted at different times in the Eastern and Jaffna Universities, the laboratory conditions and procedures used for rearing and testing larvae were the same in the two laboratories. Numbers of adults emerging were determined and the results were recorded as the mean percentage survival of larvae to reach adulthood at each salinity level ± standard errors of the mean.
An ovitrap-based experiment was conducted at Thirunelvely between August and October, 2010 in the dry season to assess the egg laying preferences of Ae. aegypti and Ae. albopictus for brackish water of varying salinity. Eleven ovitraps with salinities of 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20 ppt were prepared by adding tap water to sea water [14] and placed on an aluminum tray. Three similar trays each containing 11 ovitraps were prepared and placed 50 m apart in open buildings with roof cover to prevent dilution in case of rain. Observations were made weekly for six weeks and during each inspection, the larvae in each ovitrap were collected and identified. The solutions in the ovitraps were replaced with fresh solutions after every round of larval collection.
A larval survey that examined brackish water collections in discarded containers such as tins, glass-bottles, plastic cups and bottles and discarded tyres as well as abandoned boats and disused wells was carried out along the Thannamunai coast from August to October, 2010 and the Jaffna city coast from February to March, 2011. In a parallel study, the salinities of randomly selected domestic wells located along a 6 km stretch of the main road from the Jaffna coast to Thirunelvely were determined and the water examined for Aedes larvae. Water samples, with or without Aedes larvae, were taken to the Zoology Laboratories of the Eastern and Jaffna Universities for measuring salinity and identifying larvae.
Data on the number of new dengue cases in different administrative divisions of Jaffna city were obtained from the Medical Office for Health, Ministry of Health, Jaffna for the seven month period from 1 October 2010 to 30 April 2011. Population data for the divisions of Jaffna city were obtained from the Office of the Divisional Secretariat of Nallur and Jaffna. Incidence was calculated as the number of dengue cases per 1000 persons for the seven month period.
The level of salinity producing 50% failure to emerge as adults (LC50) and its 95% confidence limits were determined by Probit analysis using Minitab statistical software (Minitab Inc, PA, USA) for each larval population together with LC50 ratio tests to additionally determine the significance of LC50 variations among the populations as described by Wheeler et al. [17]. Multiple regression analysis was performed on the oviposition data using the Minitab statistical software to determine the relationships between experimental variables. The numbers of larvae of each species collected weekly was considered as the dependent variable and salinity, time and mosquito species as independent variables. Mosquito species were considered dummy variables and ascribed values of 0 for Ae. aegypti and 1 for Ae. albopictus in the analysis. The multiple regression model therefore had a mixture of quantitative (time and salinity) and qualitative (mosquito species) predictors.
The salinity tolerance of Ae. aegypti and Ae. albopictus larvae originating from Thirunelvely and Batticaloa town in the districts of Jaffna and Batticaloa respectively in Sri Lanka (Figure 1) were initially determined in the laboratory. The results are presented graphically in Figure 2 with relevant parameters tabulated in Table 1. The results show that, in all instances where 95% confidence intervals could be determined for the LC50 values, the third instar larvae of both species were significantly more tolerant of salinity at p<0.05 , based on non-overlapping confidence intervals, than the corresponding first instar larvae at both locations (Table 1). These findings were confirmed by LC50 ratio tests [17] at the p<0.01 level of significance (Table S1). The non-overlapping 95% confidence intervals (Table 1), confirmed by LC50 ratio tests at p<0.01 (Table S1), also suggest that Aedes populations in Thirunelvely may be more tolerant of salinity than the corresponding ones from Batticaloa, with the caveat that the experimental comparison was done at different times in two separate laboratories, albeit under similar assay conditions. Ae. albopictus larvae also tended to have higher LC50 values than the corresponding Ae. aegypti larvae at both locations but the differences were not statistically significant by either the 95% confidence interval (Table 1) or LC50 ratio (Table S1) tests.
The capacity of the two Aedes species to lay eggs in field ovitraps containing water with salinity varying from 0 to 20 ppt was also determined. Observations were made weekly for six weeks in Thirunelvely. Ae. aegypti and Ae. albopictus laid significant numbers of eggs in ovitraps with salinity up to 18 ppt and 16 ppt respectively (Table 2). The analysis of variance showed that the multiple regression analysis model was highly significant (F = 13.9, df = 3, p<0.001) with partial regression coefficients demonstrating that this was due to the numbers of larvae collected per week increasing over time (coefficient = 0.17, T = 2.4, p = 0.016) and decreasing with increasing salinity in the ovitraps (coefficient = −0.77, T = −5.8, p<0.001). However the multiple regression analysis showed that the numbers of larvae collected per week were not significantly different between the two mosquito species (coefficient = −2.49, T = −1.5, p = 0.145).
Potential brackish water preimaginal development sites of Ae. aegypti and Ae. albopictus in the coastal peri-urban environment of Thannamunai, Batticaloa district were then investigated over a four month period. Among 83 discarded food and beverage containers with brackish water that were inspected 14 (17%), with salinity levels ranging from 2 to 14 ppt, were found to contain either Ae. aegypti or Ae. albopictus larvae (Table 3). The salinity of water samples from containers that did not possess Aedes larvae varied from 2 to16 ppt. Of 89 potential brackish water habitats subsequently investigated along the coast of Jaffna city, three disused boats and two abandoned wells (6%) were found to contain Ae. aegypti larvae in a salinity range 3 to15 ppt (Table 3, with illustrative photographs of habitats in Figure 3). The salinity of brackish water samples from habitats in Jaffna city that did not contain Aedes larvae varied from 1–18 ppt. Along the main road from the Jaffna coast to Thirunelvely, 102 frequently used domestic wells were also examined. Aedes larvae were not found in such wells, where the salinities ranged from 9 ppt near the sea to 0 ppt in Thirunelvely (Figure 4).
The incidence of dengue in the period 1 October 2010 to 30 April 2011 in Jaffna city, which included the larval survey period of February and March 2011, was relatively high in divisions close to the coastal brackish water sites where Ae. aegypti larvae were detected (Figure 5).
Several mosquito species with salinity-tolerant larvae are vectors of human arboviral and parasitic diseases in many parts of the world [18]. They include some members of the genus Aedes, e.g. Aedes togoi Theobald found in coastal marshes and Aedes taeniorhyncus Wiedemann in splash pools, whose osmoregulatory mechanisms have been well studied [19]. Ae. aegypti and Ae. albopictus in contrast have been widely regarded to undergo preimaginal development only in freshwater [3], [5]. However early laboratory studies showed that Ae. aegypti larvae can tolerate limited salinity changes through an osmoconformation mechanism involving the accumulation of amino acids and ions in the haemolymph [20]. Our findings show that Ae. aegypti and Ae. albopictus can oviposit and undergo preimaginal development in fresh (tap) and brackish waters under field conditions, but that oviposition diminishes with increasing salinity of the water. Previous laboratory studies of oviposition by established laboratory colonies of Ae. aegypti showed that it prefers 2.5 ppt salinity compared to distilled water for oviposition [21], [22]. These studies also suggested that Ae. aegypti can oviposit in up to 30 ppt salinity under laboratory conditions. However such findings may not be directly applicable to the natural situation where freshwater habitats would contain dissolved minerals that might influence oviposition. They are also not directly comparable to our data with field mosquito populations and tap water from an artesian well source that contains dissolved minerals. A study on the oviposition preference of a laboratory colony of Ae. albopictus found that tap water was preferred to 1 ppt and higher concentrations of NaCl [23]. Our data from Thirunelvely suggest that Ae. aegypti and Ae. albopictus are able to oviposit in a field situation over a relatively wide range of salinity of up to 18 ppt and 16 ppt respectively. This is consistent with the finding of their larvae in brackish water collections of 2 to 15 ppt salinity in Thannamunai and Jaffna city coasts. In the context of the findings in Thirunelvely, it may be relevant to compare the salinity tolerance of the larvae of mosquitoes emerging from fresh and brackish water ovitraps as this can demonstrate potential genetic differences in the field populations of mosquitoes. Water conditioned by previous culture with Ae. albopictus larvae enhances oviposition by Ae. albopictus [23], presumably due to the presence of chemicals that are sensed by receptors in the gravid female mosquito [22]. This phenomenon could explain the greater oviposition with time in the field ovitraps observed in our investigation. However an increase in mosquito abundance could have also caused or contributed to the increasing oviposition during the six week study period.
Ae. aegypti larvae derived from an established laboratory colony showed approximately 80% survival in deionized water, >90% survival at 3.5 ppt salinity, rapid decrease in survival at >8 ppt salinity with an approximate LC50 of 14 ppt, and 0% survival at 17.5 ppt salinity in a study by Clarke et al. in the United States [24]. This study further showed that preimaginal development times were prolonged and pupal mass reduced at high salinities [24]. Because of the use of deionized water as the control and for diluting sea water, a different criterion for survival, and the use of larvae from an established laboratory colony in that study, the results are not strictly comparable to our present findings. However, the LC50 values for the two Aedes species for both first and third instar larvae in Batticaloa tend to be lower and the LC50 values for Thirunelvely similar to the LC50 observed with the Ae. aegypti laboratory colony by Clarke et al. [24]. The greater salinity tolerance of third instar larvae compared to first instar larvae observed in our study may be due to structural and physiological changes related to those that reduce ion permeability in pupae [19].
The Jaffna peninsula in Sri Lanka, unlike Batticaloa, is largely composed of sedimentary limestone and has many lagoons and other inland saline water bodies. Groundwater from aquifers and wells is over-used for domestic and agricultural purposes and therefore groundwater salinization, as illustrated also in Figure 4 for domestic wells, is widespread [25]. This effect is compounded by a high population density of approximately 700 persons per km2 in the 1130 km2 peninsula. Although Thirunelvely is an area where the accessible ground water in wells is fresh (Figure 4), the small size of the Jaffna peninsula provides many areas with brackish water close enough to Thirunelvely to be within reach of Aedes populations. The possibly greater salinity tolerance of the Aedes populations in Thirunelvely compared to Batticaloa needs to be confirmed by parallel experiments carried out under identical conditions in the same laboratory but may reflect adaptive genetic differences in the two populations analogous to that seen in Aedes camptorhynchus Thomson, a salinity-tolerant vector of Ross River virus in Southwestern Australia [26]. Larvae of coastal marsh populations of Ae. camptorhynchus tolerate greater salinity (52 ppt, i.e. hypersalinity) than inland populations (30 ppt, i.e. approaching the average salinity of sea water), probably due to genetic changes in osmoregulatory mechanisms [26]. The gradual adaptation of laboratory colonies of Ae. taeniorhyncus to increasing salinity are also likely to be due to genetic changes [19]. Genetic adaptation of mosquitoes to tolerate salinity in preimaginal habitats is also exemplified by differential salinity tolerance among sibling species of Anopheles mosquitoes [18].
Ae. aegypti and Ae. albopictus, are widely adapted to urban and suburban environments in Sri Lanka [27], Brunei [9] and many other countries [1], [10]. Although Ae. albopictus is reportedly more exophillic than Ae. aegypti, both mosquitoes normally lay eggs within a 1 km radius of their feeding site, thereby ensuring continuing proximity to human hosts [28]. This facilitates the human-mosquito-human transmission cycle which is normal for urban dengue and chikungunya, and during epidemics of yellow fever in urban areas [1]. However sylvatic cycles continue to exist in some regions for yellow fever and chikungunya [1]. Our findings now show that Ae. aegypti and Ae. albopictus lay eggs and their larvae and pupae survive to emerge as adults in brackish water, and that brackish water habitats with larvae are present in the peri-urban environment in Sri Lanka. The brackish water larval sites identified in this study are located in popular beaches or coastal areas <1 km from densely populated housing, consistent with their potential role in dengue and chikungunya transmission. These findings are compatible with oviposition by the two Aedes species in brackish water observed in a field situation, and the extent of salinity tolerance shown by their larvae in laboratory studies. Larval indices like the premises index (PI, the percentage of houses with freshwater containers positive for larvae) are commonly used entomological parameters for assessing the potential for the transmission of dengue by Ae. aegypti and Ae. albopictus [5], [10]. Dengue outbreaks have occurred in areas with a PI of approximately 2% in Singapore [10] and Cuba [29]. The relevance of 6% and 17% larval positivity rates for brackish water collections in Sri Lanka, although a different measure from the PI, needs to be evaluated further in the context of their potential contribution to the transmission of dengue and chikungunya. The present findings are also consistent with the possibility that sites where larvae of Ae. aegypti were found in brackish water may be located in the vicinity of neighborhoods with a higher dengue incidence in the coastal areas of Jaffna city, Sri Lanka. Ae. albopictus larvae were also found in brackish water within discarded food and beverage containers, with salinities up to 8 ppt, along the urban coast of dengue-endemic Brunei (R.R., S.N.S., Idris, F., Yasin, K.M., unpublished data), suggesting that our findings in Sri Lanka may be applicable to many other countries. Detailed epidemiological studies are however needed to conclusively demonstrate that preimaginal development of the two Aedes species in brackish water contributes to the transmission of dengue, chikungunya and other arboviral diseases.
It is possible that adaptation of the two Aedes populations to salinity may be accompanied by alterations in their vectorial capacity and this merits further investigation. The viability of the preimaginal stages, preimaginal development times and fitness of the emergent adults in brackish water habitats in the environment also need to be determined. Our results further suggest that brackish and freshwater domestic wells that are in frequent use are not common habitats for the preimaginal development of the two Aedes species, possibly because of regular disturbance of the water surface and rapid water turnover as well as a paucity of decaying organic matter that provide oviposition cues [30].
Small tropical islands and Southeast Asian countries have a high coastline to land mass ratio and therefore proportionately more potential coastal brackish water sites where the two Aedes species may undergo preimaginal development. Because of the exclusive focus of larval control measures on freshwater habitats [5], [10], [27], it is possible that Aedes vectors undergoing preimaginal development in brackish water collections in artificial containers have unknowingly contributed to the persistence or emergence of dengue and chikungunya in Sri Lanka [7], [8], chikungunya in Reunion [31], and dengue, despite intensive control programs, in countries like Cuba [32], Singapore [10] and Brunei [9]. High and increasing population densities in coastal areas with attendant socio-economic changes [18] and failing refuse collection systems may exacerbate this situation in resource-poor countries. Salinity-tolerant Ae. albopictus will further increase the potential for transmission of dengue and chikungunya in temperate zone countries.
Climate (temperature, rainfall, humidity) change due to global warming can expand the geographical range of vector mosquitoes, extend the disease transmission season, shorten the gonotrophic cycle and reduce the time taken for ingested viruses to develop to infectivity in mosquitoes, thereby increasing the propagation rates of arboviral diseases by Ae. aegypti and Ae. albopictus [1], [2], [4], [33]–[35]. Furthermore, a rise in sea levels consequent to global warming can increase the extent of natural brackish surface water bodies in coastal areas [36], an effect that can be compounded by higher rates of withdrawal of water from freshwater aquifers in coastal areas by expanding populations [37]. A hypothesis, presented in detail elsewhere [18], suggests that rising sea levels can therefore increase the abundance of salinity-tolerant mosquito vectors and lead to the adaptation of normally freshwater vectors to brackish waters, thereby additionally enhancing transmission of mosquito-borne diseases in coastal areas. The predicted increase in the worldwide population density of coastal areas from 87 persons per km2 in the year 2000 to 134 persons per km2 in 2050 [38] is also likely to exacerbate the situation by increasing human-vector contact.
Our results show that Ae. aegypti and Ae. albopictus have successfully adapted to oviposit and undergo preimaginal development in brackish water collections in unused wells and discarded artificial containers of up to 15 ppt salinity in the peri-urban environment. Similar salinity levels occur in parts of natural brackish water bodies like lagoons, estuaries, coastal marshes and tidal pools, as well as ponds, lakes and wells near urbanized coasts. There is no evidence at present for the large scale adaptation of Ae. aegypti and Ae. albopictus to undergo preimaginal development in natural brackish or saline water bodies. Continuous application of vector control methods solely to freshwater preimaginal development habitats in the urban environment may select for genetic changes favoring the development of Ae. aegypti and Ae. albopictus in artificial collections of brackish water in coastal urban areas, which in turn, could conceivably also lead to their adaptation to natural brackish water habitats in the future. Such changes could have serious consequences for the health of millions of people in many parts of the world, through a higher incidence of dengue, chikungunya and urban yellow fever as well as other rarer arboviral diseases [1], [4], [13]. The presence of Ae. aegypti larvae in disused brackish water wells in Jaffna and the possibly greater salinity tolerance of both Ae. aegypti and Ae. albopictus larvae in the Jaffna peninsula, where there is more salinization of groundwater compared to Batticaloa [25], may herald the beginning of this adaptive process. We recently showed that Anopheles culicifacies Giles, the major vector of malaria in Sri Lanka and an established freshwater species [39], is able to undergo preimaginal development in brackish waters of salinity up to 4 ppt in Sri Lanka [40]. Additionally Anopheles subpictus Grassi species B with genetic similarity to Anopheles sundaicus Rodenwaldt was demonstrated to be an euryhaline species (ability to tolerate a range of salinities) in Sri Lanka with its larvae being isolated from freshwater and 30 ppt salinity water in a lagoon [41]. Our present findings suggest Ae. aegypti and Ae. albopictus may have also developed euryhaline-like features, and that further investigations are required to characterize possible genetic changes that may be responsible. National and international health authorities however need to recognize the potential impact on human health of brackish water-adapted Ae. aegypti, Ae. albopictus and other mosquito vectors that were traditionally considered to be freshwater species, and institute appropriate surveillance and control measures.
|
10.1371/journal.ppat.1003926 | Shigella Type III Secretion Protein MxiI Is Recognized by Naip2 to Induce Nlrc4 Inflammasome Activation Independently of Pkcδ | Recognition of intracellular pathogenic bacteria by members of the nucleotide-binding domain and leucine-rich repeat containing (NLR) family triggers immune responses against bacterial infection. A major response induced by several Gram-negative bacteria is the activation of caspase-1 via the Nlrc4 inflammasome. Upon activation, caspase-1 regulates the processing of proIL-1β and proIL-18 leading to the release of mature IL-1β and IL-18, and induction of pyroptosis. The activation of the Nlrc4 inflammasome requires the presence of an intact type III or IV secretion system that mediates the translocation of small amounts of flagellin or PrgJ-like rod proteins into the host cytosol to induce Nlrc4 activation. Using the Salmonella system, it was shown that Naip2 and Naip5 link flagellin and the rod protein PrgJ, respectively, to Nlrc4. Furthermore, phosphorylation of Nlrc4 at Ser533 by Pkcδ was found to be critical for the activation of the Nlrc4 inflammasome. Here, we show that Naip2 recognizes the Shigella T3SS inner rod protein MxiI and induces Nlrc4 inflammasome activation. The expression of MxiI in primary macrophages was sufficient to induce pyroptosis and IL-1β release, which were prevented in macrophages deficient in Nlrc4. In the presence of MxiI or Shigella infection, MxiI associated with Naip2, and Naip2 interacted with Nlrc4. siRNA-mediated knockdown of Naip2, but not Naip5, inhibited Shigella-induced caspase-1 activation, IL-1β maturation and Asc pyroptosome formation. Notably, the Pkcδ kinase was dispensable for caspase-1 activation and secretion of IL-1β induced by Shigella or Salmonella infection. These results indicate that activation of caspase-1 by Shigella is triggered by the rod protein MxiI that interacts with Naip2 to induce activation of the Nlrc4 inflammasome independently of the Pkcδ kinase.
| Shigella are bacterial pathogens that are the cause of bacillary dysentery. An important feature of Shigella is their ability to invade the cytoplasm of host epithelial cells and macrophages. A major component of host recognition of Shigella invasion is the activation of the inflammasome, a molecular platform that drives the activation of caspase-1 in macrophages. Although Shigella is known to induce the activation of the Nlrc4 inflammasome, the mechanism by which the bacterium activates Nlrc4 is largely unknown. We discovered that the Shigella T3SS inner rod protein MxiI induces Nlrc4 inflammasome activation through the interaction with host Naip2, which promoted the association of Naip2 with Nlrc4 in macrophages. Expression of MxiI induced caspase-1 activation, Asc oligomerization, pyroptosis and IL-1β release which required Naip2, but not Naip5. Significantly, caspase-1 activation induced by Shigella infection was unaffected by deficiency of the Pkcδ kinase. This study elucidates the microbial-host interactions that drive the activation of the Nlrc4 inflammasome in Shigella-infected macrophages.
| Recognition of intracellular pathogenic bacteria by members of the nucleotide-binding domain and leucine-rich repeat containing (NLR) family triggers immune responses against bacterial infection [1], [2]. A major response against several pathogenic Gram-negative bacteria, including Salmonella, Legionella, and Shigella is the activation of caspase-1 via Nlrc4 in macrophages [1], [3]. Upon bacterial stimulation, Nlrc4 mediates the formation of a multi-protein complex termed the inflammasome that induces the activation of caspase-1 leading to the proteolytic maturation of pro-IL-1β and pro-IL-18 as well as the induction of pyroptotic cell death in macrophages [4]–[6]. Many Gram-negative bacteria encode a type III secretion system (T3SS) with conserved structural features that promote virulence by injecting bacterial effector proteins directly into the cytosol of host cells [7], [8]. In macrophages infected with Salmonella, the cytosolic delivery of flagellin or the bacterial rod protein PrgJ through the T3SS is recognized by Nlrc4 leading to inflammasome activation [9]. Recently, Naips (NLR family, apoptosis inhibitory proteins) have been shown to act as adaptor molecules that connect flagellin or the bacterial rod protein PrgJ to Nlrc4 [10], [11]. Specifically, Naip5 and Naip6 associate with flagellin to promote Nlrc4 oligomerization and inflammasome activation, whereas Naip2 links PrgJ to Nlrc4 [10]–[12]. These findings suggest a model in which certain Naips specifically recognize flagellin or PrgJ to mediate Nlrc4 inflammasome activation. Recent studies, however, have revealed that the activation of Nlrc4 is more complex in that phosphorylation of Nlrc4 at Ser533 was found to be critical for the activation of the inflammasome [13]. Furthermore, it was suggested that Pkcδ is the major Nlrc4 kinase responsible for Nlrc4 phosphorylation and inflammasome activation [13].
Shigella are non-flagellated bacterial pathogens that contain highly evolved invasion systems that enable them to invade host cells and colonize the epithelium of the large intestine, which ultimately leads to a severe form of colitis called bacillary dysentery [14]. After uptake of Shigella by intestinal macrophages, the bacterium delivers a subset of effector proteins via the T3SS apparatus into the host cytosol [7], [8], [15]. The inner rod of the T3SS needle complex forms a conduit for protein transport through the periplasm which is assembled by the polymerization of PrgJ in Salmonella and its homologue MxiI in Shigella [16], [17]. Because of the homology of Salmonella PrgJ with Shigella MxiI, it can be predicted that Shigella induces activation of Nlrc4 via the sensing of MxiI by host macrophages. Consistent with this notion, the T3SS of Shigella is required to induce IL-1β secretion and pyroptosis via the Nlrc4 inflammasome [18]. Furthermore, ectopic expression of MxiI reduced the viability of macrophages and this was inhibited in the absence of Nlrc4 [9]. However, the mechanism by which Shigella MxiI induces activation of the Nlrc4 inflammasome remains unknown. In this study, we provide evidence that MxiI mediates the activation of the Nlrc4 inflammasome through interactions with Naip2. Furthermore, we demonstrate that Naip2, but not Naip5, is critical for the interaction of MxiI with Nlrc4 and the activation of the inflammasome in macrophages infected with Shigella. Finally, we show that Pkcδ is dispensable for Nlrc4 activation.
In the case of flagellated pathogenic bacteria, flagellin is a major and potent stimulator of the Nlrc4 inflammasome. In addition, Salmonella T3SS rod protein PrgJ is sensed by Nlrc4 to activate caspase-1. Because Shigella are unflagellated bacteria, we hypothesized that the Shigella T3SS rod protein MxiI, a homologue of Salmonella PrgJ, induces the activation of the Nlrc4 inflammasome. To test this hypothesis, we expressed MxiI in wild-type (WT) and Nlrc4-deficient bone marrow-derived macrophages (BMDM) using a MSCV-IRES-GFP retroviral vector and assessed cell viability by the numbers of viable green fluorescence protein (GFP)-positive cells. After overnight culture, the viability of WT macrophages was dramatically decreased by MxiI-GFP expression when compared to expression of GFP (Figure 1A). Importantly, the decrease in cell viability was inhibited in Nlrc4−/− macrophages (Figure 1A). Consistently, expression of MxiI-GFP, but not GFP, induced the release of IL-1β in WT macrophages, which was abolished in macrophages lacking Nlrc4 (Figure 1B). These results indicate that expression of MxiI induces the activation of the Nlrc4 inflammasome.
We next tested whether the rod protein MxiI interacts with Naip2 or Naip5 in macrophages. Because expression of MxiI in macrophages causes cell death (Figure 1A), we used macrophages from caspase-1-deficient mice to assess the interaction of MxiI with Naip proteins by immunoprecipitation. In these experiments, we expressed T7-tagged MxiI in the presence of HA-tagged Naip2, HA-tagged Naip5 or control plasmid. Immunoprecipitation analysis showed that MxiI associated with Naip2, but much less with Naip5 as revealed by immunoblotting with anti-HA antibody (Figure 2A). Next, we investigated the interaction between Nlrc4 and Naip2 in Shigella-infected macrophages. To assess this, we expressed T7-tagged Nlrc4 and HA-tagged Naip2 or Naip5, or control empty vector in uninfected or caspase-1-deficient macrophages infected with WT or an isogenic Shigella strain deficient in the T3SS (S325). Immunoprecipitation analysis revealed that Naip2 interacts with Nlrc4 in macrophages infected with WT Shigella (Figure 2B). However, Naip2 did not associate with Nlrc4 in uninfected macrophages or macrophages infected with the mutant bacterium lacking a functional T3SS that are unable to release MxiI into the host cytosol (Figure 2B). Furthermore, infection with Shigella preferentially promoted the interaction of Nlrc4 with Naip2 relative to Naip5 (Figure 2B). MxiI is secreted into the culture medium by Shigella which relies on the presence of a functional T3SS [19]–[21]. Therefore, MxiI is presumably leaked into the host cytosol via the T3SS to activate Nlrc4, as it was suggested for Salmonella PrgJ [22], [23]. Therefore, we next asked whether expression of MxiI promotes the association of Naip2 with endogenous Nlrc4 in uninfected macrophages. Immunoprecipitation experiments showed that expression of MxiI induced the interaction of Naip2 with endogenous Nlrc4 (Figure 2C). Collectively, these results indicate that MxiI interacts preferentially with Naip2 and promotes the interaction between Naip2 and Nlrc4.
We next performed additional studies to verify that Shigella infection promoted the activation of Nlrc4 via Naip2. To confirm the preferential effect of Naip2 on Nlrc4 activation, we performed reconstitution experiments by expressing Nlrc4, Asc, caspase-1, pro-IL-1β and Naip2 or Naip5 in 293T cells. One day after transfection, cells were infected with WT or T3SS-deficient Shigella for 3 hrs and inflammasome activation was analyzed by immunoblotting with an antibody specific for mature IL-1β p17. In the absence of exogenous Naip2 or Naip5, infection with WT Shigella enhanced the processing of pro-IL-1β into IL-1β p17 (Figure S1A). The formation of IL-1β p17 was further enhanced by Naip2, but inhibited by Naip5 in Shigella-infected cells (Figure S1A). In this reconstitution system, the enhancement of IL-1β p17 formation by Naip2 in cells infected with WT Shigella required Nlrc4, Asc and caspase-1 (Figure S1B).
Shigella infection stimulates Nlrc4- and Asc-dependent inflammasome activation in macrophages [18]. However, Shigella was also shown to induce macrophage cell death via Nlrp3 after 2–6 hrs of infection at a bacteria/macrophage ratio of 50∶1 [24]. To verify these seemingly contradictory results, we reassessed the role of Asc, Nlrc4 and Nlrp3 in Shigella-induced caspase-1 activation. In these experiments, LPS-primed BMDM were infected with the Shigella WT or S325 (T3SS-deficient mutant) at a bacteria/macrophage ratio of 10∶1 for 30 min. As expected, WT, but not mutant Shigella, induced processing of procaspase-1 into the p20 subunit of caspase-1 (Figure S2A). The inability of the mutant bacterium to activate caspase-1 could not be explained by reduced uptake by macrophages (Figure S3). Importantly, caspase-1 activation, IL-1β release, and pyroptosis required Nlrc4 and Asc, but not Nlrp3 (Figure S2A–C). Because previous studies showed that Asc was not required for pyroptosis induced by Shigella in BMDM differentiated for 5 days [18], we assessed cell death induced by Shigella in BMDM differentiated for 3, 4 and 5 days in culture (Figure S4). Consistent with previous studies [18], Asc was not required for pyroptosis in macrophages differentiated for 5 days (Figure S4). In macrophages differentiated for 3 or 4 days, however, cell death induced by Shigella was enhanced in WT macrophages and impaired in Asc-deficient macrophages (Figure S2C and S4) which is in line with the results presented in Figure S2C.
Next, we investigated the role of Naip2 and Naip5 in caspase-1 activation induced by Shigella. We used siRNA-mediated knockdown to reduce the expression of Naip2 and Naip5 in macrophages (Figure 3A). Notably, caspase-1 activation induced by Shigella was attenuated by inhibiting the expression of Naip2, but not Naip5 (Figure 3B). Importantly, the ability of individual siRNA to inhibit caspase-1 activation correlated with reduction of Naip2 expression (Figure 3A, B). In addition, knockdown of Naip2, but not Naip5, reduced the release of IL-1β and IL-18 induced by Shigella infection at 1 or 2 hrs post-infection (Figure 3B, C). In control experiments, knockdown of Naip2 did not affect the production of IL-6 or CXCL2 in macrophages infected with WT or S325 mutant Shigella (Figure 3C). These results suggest that Shigella induces Nlrc4-dependent inflammasome activation via Naip2 in macrophages.
The Asc pyroptosome is a molecular platform that is thought to be important for the recruitment and activation of caspase-1 [25]–[27]. Infection of macrophages with WT, but not T3SS-deficient, Shigella induced the formation of the Asc pyroptosome which was detected in the cell cytoplasm by staining with an antibody that recognizes Asc (Figure 4A, B). The Asc pyroptosome induced by Shigella infection co-localized with FLICA staining that labels activated caspase-1 (Figure 4A). Importantly, knockdown of Naip2 by siRNA reduced Asc pyroptosome formation whereas Naip5 did not (Figure 4C, D). To provide direct biochemical evidence that the Asc pyroptosome is formed, we cross-linked the insoluble Asc protein complexes from Shigella or Salmonella infected macrophages and subjected them to immunoblotting with anti-Asc antibody. Immunoblotting analysis revealed that infection with WT Shigella or Salmonella induces prominent Asc dimer formation in WT, but not Asc-deficient macrophages (Figure 5A, upper panel). The induction of Asc dimers correlated with IL-1β release in culture supernatants (Figure 5A, lower panel). In contrast, Shigella deficient in T3SS and the fliA-deficient Salmonella mutant were impaired in the induction of Asc dimer formation (Figure 5A). Notably, expression of MxiI was sufficient to induce the formation of Asc dimers in caspase-1-deficient macrophages in the absence of Shigella infection (Figure 5B). Furthermore, knockdown of Naip2 by siRNA, but not Naip5, inhibited Asc dimer formation (Figure 5C). These results indicate that Shigella MxiI and Naip2 are important in Asc pyroptosome formation which is associated with inflammasome activation.
Recent studies reported that Nlrc4 phosphorylation by Pkcδ is critical for inflammasome activation induced by Salmonella infection [13]. Thus, we assessed whether inflammasome activation caused by Shigella infection also requires Pkcδ. In these experiments, LPS-primed BMDM from WT and Pkcδ-deficient mice were infected with WT or S325 (T3SS-deficient mutant) Shigella, and IL-1β release was evaluated at different time points and bacterial/macrophage ratios after infection. As expected, expression of Pkcδ was induced by LPS stimulation in WT, but not Pkcδ-deficient macrophages (Figure 6A). Importantly, Pkcδ was not required for IL-1β secretion induced by Shigella or Salmonella (Figure 6B–D). In fact, Pkcδ deficiency enhanced IL-1β secretion in response to Shigella and Salmonella infection (Figure 6B–D). Furthermore, Pkcδ-deficient macrophages produced higher amounts of IL-1α and CCL5, but not CXCL2 than WT macrophages in response to infection (Figure 6D). The increased production of cytokines in Pkcδ-deficient macrophages was not associated with enhanced NF-κB or MAPK activation after Shigella infection (Figure S5). Notably, induction of apoptosis in Shigella-infected macrophages was inhibited in macrophages deficient in Pkcδ (Figure S5). Furthermore, treatment with z-DEVD-fmk, a cell permeable caspase-3 inhibitor, increased the production of IL-1β in WT macrophages infected with Shigella (Figure S5), suggesting that increased production of IL-1β in Pkcδ-deficient macrophages is mediated, at least in part, by inhibition of apoptosis in Shigella-infected macrophages. Importantly, caspase-1 activation induced by Shigella or Salmonella was unimpaired in macrophages deficient in Pkcδ (Figure 6E), whereas it was abolished in macrophages deficient in Nlrc4 (Figure 6E). These results indicate that Pkcδ is not essential for inflammasome activation induced by Shigella or Salmonella infection.
The intracellular sensing of flagellin is the major trigger for the activation of the Nlrc4 inflammasome in macrophages infected with Salmonella [4]. Because Shigella is non-flagellated, the current studies were aimed at understanding the mechanism by which Shigella induces the activation of Nlrc4 in macrophages. We show here that Shigella induces the activation of the Nlrc4 inflammasome through MxiI, an inner rod protein of the T3SS. MxiI associated with Naip2 and was sufficient to induce Nlrc4-dependent IL-1β secretion and the interaction with Nlrc4. Importantly, inhibition of Naip2 expression impaired the activation of the Nlrc4 inflammasome and IL-1β/IL-18 release in Shigella-infected macrophages. Because IL-1β secretion induced by Shigella was not abolished by Naip2 knockdown, it is possible that Shigella also activates another inflammasome pathway that is minor and only unmasked by the inhibition of the Naip2-Nlrc4 pathway. Alternatively, it is possible that the partial inhibition of IL-1β secretion reflects residual Naip2 protein expression in macrophages.
Our work is consistent with a model in which the T3SS inner rod proteins including PrgJ in Salmonella and MxiI in Shigella are recognized by Naip2 and this interaction leads to the recruitment and activation of Nlrc4. Consistent with this model, we show that expression of MxiI promotes the association of Naip2 with Nlrc4 and induces the oligomerization of Asc in macrophages. Furthermore, WT, but not T3SS-deficient Shigella, enhances the association of Naip2 and Nlrc4 in macrophages. The failure of mutant Shigella to induce the interaction between Naip2 and Nlrc4 is presumably explained by the inability of the T3SS mutant to release MxiI into the host cytosol. A measure of inflammasome activation is the formation of Asc oligomers [25]–[27]. Importantly, Asc oligomerization induced by MxiI was observed in caspase-1-deficient macrophages, indicating that this critical event is not a secondary event of caspase-1 activation. MxiI is composed of 97 amino acids and is predicted to be a soluble protein using publically available tools (http://www.psort.org/psortb). It has been shown that MxiI is secreted into the culture medium by Shigella in a T3SS dependent manner [19]–[21]. Thus, as it was suggested for Salmonella PrgJ [9], [22], we propose that small amounts of MxiI are leaked into the host cytosol via the T3SS during Shigella infection to induce the activation of Nlcr4.
Recent studies showed that Nlrc4 phosphorylation was induced by Salmonella and was found to be critical for inflammasome activation [13]. Furthermore, it was proposed that Pkcδ was the major kinase responsible for phosphorylation of Nlrc4 [13]. In contrast to the latter finding, we found that IL-β secretion and caspase-1 activation induced by Shigella and Salmonella infection were not impaired in Pkcδ-deficient macrophages. Notably, the production of several cytokines including IL-1β was enhanced in infected Pkcδ-deficient macrophages. A possible mechanism to account for the enhanced production of cytokines in Pkcδ-deficient macrophages is the observation that Pkcδ regulates phagosomal production of ROS [28] which is known to inhibit pro-inflammatory responses including cytokine production [29]. However, we did not observe enhanced NF-κB or MAPK activation in Pkcδ-deficient macrophages infected with Shigella. Pkcδ has been shown to regulate the induction of apoptosis [30]–[32]. Consistently, apoptosis induced by Shigella infection was impaired in Pkcδ-deficient macrophages and treatment with a caspase-3 inhibitor enhanced IL-1β secretion in WT macrophages. These results suggest that the increased production of cytokines observed in Pkcδ-deficient macrophages might be due, at least in part, to suppression of apoptosis in infected macrophages. Regardless of the mechanism involved, our results clearly show that caspase-1 activation induced by Shigella or Salmonella infection is not impaired in Pkcδ-deficient macrophages. We do not have a clear explanation for the difference in results between our studies and previous results by Qu et al. These authors showed that in addition to Pkcδ, Pak2 was capable of phosphorylating Nlrc4 at the critical Ser533, although the results suggested that Pak2 was a minor Nlrc4-phosphorylating kinase [13]. Thus, it is conceivable that the difference in results could be explained by kinase redundancy and subtle variation in the expression of Nlrc4-phosphorylating kinases in different macrophage preparations. Regardless of the explanation, findings within this investigation clearly show that Pkcδ is dispensable for Nlrc4 activation. Thus, our results challenge the notion that Pkcδ is critical for inflammasome activation and indicate that further work is needed to understand the mechanism and role of Nlrc4 phosphorylation in inflammasome activation.
Shigella MxiI associates with Naip2 to induce the interaction of Naip2 with Nlrc4, which presumably leads to Nlrc4 oligomerization and inflammasome activation. In the Salmonella system, cytosolic flagellin binds to Naip5 and induces the association of Naip5 with Nlrc4 [10]–[12]. Reconstitution experiments with purified flagellin, Naip5 and Nlrc4 revealed that these components are sufficient to induce the formation of a disk-like complex composed of 11 or 12 proteins including Nlrc4 and Naip5, although the exact ratio of Naip5 and Nlrc4 in the complex remains unclear [12]. Based on the latter observations, we suggest that Shigella MxiI induces the oligomerization of Nlrc4 via their interaction with Naip2. Consistent with this model, we found that MxiI induced the interaction of Naip2 with Nlrc4 and the oligomerization of Asc. Furthermore, Naip2, but not Naip5, was critical for caspase-1 activation, pyroptosome formation, Asc oligomerization and IL-1β secretion. Collectively, these results support a model in which distinct Naip family members act as sensors of flagellin and T3SS inner rod proteins and oligomerized Nlrc4 provides a platform for the recruitment and activation of caspase-1. While Naip2 knockdown reduced inflammasome activation, Naip5 knockdown had the opposite effect in response to Shigella infection. Although further work is needed to understand the role of Naip5, one possibility is that there is competition between Naip2 and Naip5 protein complexes and inhibition of Naip5 enhances the Naip2-Nlrc4 inflammasome pathway. Nlrc4 and caspase-1 contain CARD domains and they could interact directly via homotypic CARD-CARD interactions. However, the adaptor Asc is essential for the activation of caspase-1 in response to Salmonella and Shigella [18], [33]. These results suggest that Asc is somehow required for the interaction between Nlrc4 and caspase-1 or that Asc is critical for another step which is important for inflammasome activation.
All animal experiments were conducted according to the U.S.A. Public Health Service Policy on Humane Care and Use of Laboratory Animals. Animals were maintained in an AAALAC approved facility and all animal studies followed protocol 09716-2 that was approved by the Animal Care and Use Committee of the University of Michigan (Ann Arbor, MI).
Mice deficient in Nlrc4, Nlrp3, Asc and caspase-1/11 have been previously described [4], [34], [35]. All mice were crossed at least 5 times on a C57BL/6 background. Bone marrow samples from Prkcd−/− mice in C57BL/6 background were provided by Hee-Jeong Im Sampen (Rush University Medical Center, Chicago, IL).
Shigella flexneri strain YSH6000 [36] was used as the WT strain, and S325 (mxiA::Tn5) [37] was used as the T3SS–deficient control. The WT S. enterica serovar Typhimurium SR-11 χ3181 and the isogenic fliA::Tn10 were provided by H. Matsui (Kitasato Institute for Life Science, Tokyo, Japan) [38]. ΔfliA Salmonella mutant is impaired in the expression of flagellin [18]. cDNAs encoding mouse Naip2, Naip5, Nlrc4, Asc, caspase-1, and bacterial MxiI were amplified by PCR and cloned into the pCMV based mammalian expression vector or the MSCV-IRES-GFP retroviral expression vector (Addgene). Human pro-IL-1β clone (RDB6666) was provided by RIKEN BRC which is participating in the National Bio-Resource Project of the MEXT, Japan.
BMDMs were prepared from the femurs and tibias of mice and cultured for 3–7 days in 10% FCS IMDM (Gibco) supplemented with 30% L-cell supernatant, non-essential amino acids, sodium pyruvate and antibiotics (Penicillin/Streptomycin). 293T cells were cultured on Dulbecco's Modified Eagle's medium (Sigma) containing 10% FCS and antibiotics (Penicillin/Streptomycin). The rabbit anti mouse caspase-1 p20 and anti-mouse Nlrc4 antibodies were produced in our laboratory by immunizing rabbits with mouse caspase-1 (p20 subunit) and mouse Nlrc4 (amino acids 1–152) recombinant proteins [39]. Anti–IL-1β p17 (#2021) and anti-Pkcδ (#2058) antibodies were from Cell Signaling. Mouse monoclonal anti-β-actin antibody was from Sigma. HRP-conjugated goat anti–rabbit (Jackson Laboratories) or anti–mouse IgG (Sigma) or anti-rat (Jackson Laboratories), or AP-conjugated goat anti-rabbit (Santa Cruz Biotechnology Inc.) or anti-mouse IgG (Santa Cruz Biotechnology Inc.) antibodies were used as secondary antibodies for immunoblotting.
Macrophages were seeded in 24-well plates at a density of 3×105 cells per well. Cells were stimulated with or without 0.1 µg/ml LPS (from E. coli O55:B5, Sigma) for 6 h and then infected with Shigella or Salmonella. Bacterial strains were pre-cultured overnight in Mueller-Hinton broth (Difco) at 30°C, then were inoculated into brain heart infusion broth (Difco) and incubated for 2 h at 37°C prior to infection. The cells were infected with Shigella at a bacteria/macrophage ratio of 10∶1, or with Salmonella at a bacteria/macrophage ratio of 1∶1 unless otherwise stated. The plates were centrifuged at 700 g for 5 min to synchronize the infection, and gentamicin (100 µg/ml) and kanamycin (60 µg/ml) were added after 20 min. At the indicated times after infection, cytokines were measured in culture supernatants by enzyme-linked immunoabsorbent assay (ELISA) kits (R and D Systems). RNA was isolated with E.Z.N.A. TM total RNA kit (Omega Biotek) according to the manufacturer's instructions. RNA was reverse transcribed using the High Capacity RNA-to cDNA kit (Applied Biosystem) and cDNA was then used for RT-PCR. For immunofluorescence studies, the infected cells were fixed and immunostained, and then analyzed with a confocal laser-scanning microscope (LSM510; Carl Zeiss) or fluorescence microscopy (Olympus). Carboxyfluorescein FLICA (Immunochemistry Technologies, LLC) was added 1 hr before bacterial infection. Apoptosis was measured by the AnnexinV (Roche) and TUNEL (Promega) assays using fluorometric protocols according to the manufacture's recommendations. For the caspase-3 inhibitor studies, the cells were treated with 200 µM z-DEVD-fmk (Calbiochem) for 1 h before bacterial infection.
293T cells were seeded in 6-well plates at a density of 5×105 cells per well and incubated overnight. Then, the cells were transfected with or without 1 µg T7-tagged Nlrc4, 1 µg T7-tagged Asc, 0.4 µg HA-tagged caspase-1, and 0.4 µg FLAG-tagged proIL-1β [40], and 1 µg HA-tagged Naip2 or Naip5, using FuGENE 6 (Roche). Cells were infected one day after infection. Intensities of casp1 p20 or IL-1β p17 bands were quantified by densitometry, the values normalized to the β-actin protein levels and results were analyzed with ImageJ software.
The Shigella MxiI gene was cloned into the MSCV-IRES-GFP retrovirus vector, which contains an IRES-GFP element to track retroviral infection. WT or Nlrc4−/− BMDMs were immortalized using the J2 virus to increase nucleofection efficiency [41]. Then, cells were nucleofected with MSCV-IRES-GFP or MSCV-IRES-GFP encoding Shigella MxiI using an Amaxa nucleofector system (Nucleofector kit V and the D-032 program). After 20 hrs, cell survival in the GFP-positive cell population was analyzed by fluorescence microscopy. The LDH activity in the culture supernatants of infected cells was measured using the CytoTox 96 assay kit (Promega) according to the manufacturer's protocol. Assays were performed in triplicate for each independent experiment.
The invasion efficiency of Shigella strains was evaluated using a gentamicin/kanamycin protection assay. Briefly, cells were infected for 20 min and then incubated for 20 min at 37°C in medium containing gentamicin (100 µg/ml) and kanamycin (60 µg/ml) to kill extracellular bacteria. The infected cells were then washed in PBS, lysed in 0.5% TritonX-100/PBS, and serial dilutions were plated on LB agar plates to determine the number of intracellular bacteria.
DNA and siRNAs specific for Naip2 and Naip5 were transfected into macrophages using an Amaxa nucleofector system (Y-001 program for primary macrophages or D-032 program for cell lines) according to the manufacturers' instructions. siRNA pools for mouse Naip2 (17948; D-044151-01-04) and Naip5 (17951; D-044141-01-4) and non-targeting siRNAs were purchased from Dharmacon or synthesized by Sigma and targeting the sequences CTTACACTGAATCACAAGA (naip2) or GTGCCTTTTTAGTCCTTGT (naip5). Primer sets for RT-PCR were naip2-forward (AGGCTATGAGCATCTACCACA), naip2-reverse (AAGACATCAATCCACAGCAAA), naip5-forward (TGCCAAACCTACAAGAGCTGA), naip5-reverse (CAAGCGTTTAGACTGGGGATG), actin-forward (CATGTACGTTGCTATCCAGGC) and actin-reverse (CTCCTTAATGTCACGCACGAT). To compare caspase-1 p20 levels in immunoblotting experiments, the bands were quantified by densitometry, analyzed with ImageJ software, and normalized to the β-actin protein levels.
Cell ware lysed in IP buffer [CelLytic M Cell Lysis Reagent (Sigma), 0.1 mM PMSF, and a complete protease inhibitor cocktail-EDTA (Roche) and clarified lysates were mixed with anti-T7 antibody–conjugated agarose beads (Novagen) or anti-HA conjugated sepharose beads (Covance) for 1 hr at 4°C with gentle rotation in IP buffer. Beads were washed with PBS, mixed with SDS-sample buffer and subjected to immunoblot analysis.
Cells were fixed with 4% paraformaldehyde and 0.1% NP40, washed and stained with anti- Asc antibody and FITC-conjugated anti–rat antibody (Sigma) as a secondary antibody. Imaging analysis was performed using fluorescence microscopy (Olympus), and percentage of cells containing Asc pyroptosomes was determined by counting at least 300 cells in 5 separate fields.
The Asc dimerization assay was previously described [25]–[27]. Briefly, cells were lysed (20 mM HEPES-KOH, pH 7.5, 150 mM KCl, 1% NP-40, 0.1 mM PMSF, and Complete protease inhibitor cocktail (Roche)) and forced onto a 21-gauge needle 10 times. The cell lysates were centrifuged at 6000 rpm for 10 min at 4°C to isolate the insoluble fraction in the pellet. The pellets were washed twice with PBS, resuspended in 500 µl of PBS and cross-linked with fresh 2 mM disuccinimidyl suberate (DTT, Sigma) for 30 min. The cross-linked pellets were isolated by centrifugation at 13000 rpm for 10 min and resuspended in 20 µl of SDS sample buffer for immunoblotting with anti-mouse Asc antibody.
Mouse cytokines in culture supernatants were measured by ELISA kits (R&D Systems). Assays were performed in triplicate for each independent experiment.
Statistical analyses were performed using the Mann–Whitney U test. Differences were considered significant at p<0.05.
|
10.1371/journal.pntd.0004679 | Cholera Incidence and Mortality in Sub-Saharan African Sites during Multi-country Surveillance | Cholera burden in Africa remains unknown, often because of weak national surveillance systems. We analyzed data from the African Cholera Surveillance Network (www.africhol.org).
During June 2011–December 2013, we conducted enhanced surveillance in seven zones and four outbreak sites in Togo, the Democratic Republic of Congo (DRC), Guinea, Uganda, Mozambique and Cote d’Ivoire. All health facilities treating cholera cases were included. Cholera incidences were calculated using culture-confirmed cholera cases and culture-confirmed cholera cases corrected for lack of culture testing usually due to overwhelmed health systems and imperfect test sensitivity. Of 13,377 reported suspected cases, 34% occurred in Conakry, Guinea, 47% in Goma, DRC, and 19% in the remaining sites. From 0–40% of suspected cases were aged under five years and from 0.3–86% had rice water stools. Within surveillance zones, 0–37% of suspected cases had confirmed cholera compared to 27–38% during outbreaks. Annual confirmed incidence per 10,000 population was <0.5 in surveillance zones, except Goma where it was 4.6. Goma and Conakry had corrected incidences of 20.2 and 5.8 respectively, while the other zones a median of 0.3. During outbreaks, corrected incidence varied from 2.6 to 13.0. Case fatality ratios ranged from 0–10% (median, 1%) by country.
Across different African epidemiological contexts, substantial variation occurred in cholera incidence, age distribution, clinical presentation, culture confirmation, and testing frequency. These results can help guide preventive activities, including vaccine use.
| Cholera burden in Africa remains unknown, often because of weak national surveillance systems. Reporting is non-exhaustive for various reasons, such as individual and community fears of stigmatization and economic loss. Furthermore, only 3% to 5% of all cases are laboratory confirmed. A variety of case definitions are used across countries, which could lead to cholera over or under-reporting. Our study presents the first data from prospective multi-country cholera surveillance in Africa, and the only such data based on culture confirmation and that includes a description of clinical presentation. We show how confirmed cholera cases varied over time by setting, and identified three epidemiological patterns that can guide decision-making processes. We documented that reliance on suspected cases–as is usually done in national surveillance–rather than confirmed cases can over-estimate substantially cholera incidence. Finally, our surveillance strategy of using case-based reporting and a standard comprehensive case reporting form provided more information on at-risk populations and geographical hot spots than is currently available in the literature; this is turn should facilitate development of efficient preventive strategies.
| Although cholera has disappeared as a public-health problem in developed countries, it remains a major concern in sub-Sahara Africa [1,2]. From 2007 to 2012, at least 20 African countries reported more than 100,000 cases of cholera (World Health Organization (WHO) weekly epidemiological records, 2007–2012). However, this surveillance has weaknesses. Reporting is non-exhaustive for various reasons such as individual and community fears of stigmatization and economic loss. Reporting from district to national levels may be delayed or incomplete. According to WHO, only 3% to 5% of all cases are laboratory confirmed [3]. A variety of case definitions are used across countries, which could lead to cholera over or under-reporting. Finally, few countries have implemented case-based surveillance, with information at national level provided in the form of weekly summaries limited to cumulative case numbers and deaths [1].
Since the Haiti epidemic during 2010, public and political attention on cholera has increased. Recently, WHO has prequalified a two-dose oral cholera vaccine (OCV) that is less expensive and less cumbersome to deliver than its predecessor. This, and the creation by WHO of a cholera vaccine stockpile for epidemic and potentially endemic cholera prevention, have stimulated interest in more timely, accurate, and comprehensive disease burden data from affected countries.
The African Cholera Surveillance Network (Africhol) was launched in 2009. Originally implemented in eight of the most affected sub-Saharan African countries, it has since expanded to three additional countries. Its primary aim is to better define cholera burden, geographic distribution, seasonal patterns, and risk groups to inform prevention strategies, including immunization. We present here incidence results and the associated case fatality ratio from eleven geographical areas located in the six Africhol countries having the strongest performing surveillance systems.
Starting in 2011, we implemented population-based cholera surveillance in all cholera treatment facilities in a given geographic zone chosen in collaboration with ministries of health (MoHs). Criteria for zone selection included: yearly occurrence of outbreaks or sporadic cholera cases; existence of dedicated diarrhea or cholera treatment facilities; and laboratory capacity for cholera confirmation by stool culture. In these zones, all health facilities providing treatment for severe cholera cases were included in surveillance. We also conducted a prospective surveillance in several outbreak sites outside of surveillance zones when these were reported to the MoH and when they had adequate laboratory facilities available. This was conducted the time of the epidemic.
Patients were followed in all the cholera treatment facilities of a given surveillance area. In areas without known ongoing cholera, a suspected cholera case was defined as a patient aged two years or more that developed severe dehydration or died from acute watery diarrhoea. In areas with known cholera, a suspected case was defined as a patient aged two years or more that developed acute watery diarrhoea, with or without vomiting. A confirmed case was defined as a suspect cholera having a stool culture positive for Vibrio cholerae.
Eight enhanced surveillance zones located in areas of known recent cholera occurrence were included in the analysis. Their location and starting dates were as follows: 1) Togo: five districts of Lome and Golfe district, Jun 2011; 2) Togo: Lake district in the Maritime region, Jun 2011; 3) Democratic Republic of Congo (DRC): Goma and Karisimbi districts, Aug 2011; 4) Guinea: five districts of Conakry, Jul 2011; 5) Uganda: Manafwa, Mbale, and Butaleja districts, Dec 2011; 6) Mozambique: Beira city, Aug 2011; 7) Cote d’Ivoire: one district of Abidjan, Koumassi–Port Bouet–Vridi district (KPBV), Jun 2012. While data collection is currently ongoing, here we include only surveillance data collected through Dec 31st, 2013. In addition to surveillance zones, we included data collected during outbreaks in Kasese district, Uganda (Oct 2011–Dec 2012); Pemba city, Mozambique (Jan 2013–Dec 2013); Adiake prefecture, Cote d’Ivoire (May–Oct 2012); and three districts of Kinshasa (Maluku, Kingabwa, and Massina districts), DRC (Jul 2011–Feb 2012). Within specifically defined study zones, we included all health care facilities known to treat cholera cases, including long-term facilities as well as newly established cholera treatment centers (in Africa, these centers frequently are opened only in response to an outbreak). While all cholera cases were supposed to have been referred to a designated cholera treatment center, it is likely that private health centers conducted unauthorized evaluation and treatment. Included centers were the following: 1) Conakry, Guinea. The infectious disease and paediatric departments of Donka hospital. The additional cholera treatment center (CTC) in the Ratoma neighbourhood opened during the 2012 epidemic was also included; 2) Lome, Togo. The infectious disease and paediatric departments of the Centre Hospitalier Universitaire, Be Hospital, and other district health centres in which a temporary cholera treatment center was opened; 3) Lake District, Togo. The infectious disease and paediatric departments of Aneho Hospital and health centres with temporary treatment centers; 4) Goma-Karisimbi district, DRC: The cholera treatment centers located in the General Provincial Hospital, the Buhimba cholera treatment and the Kiziba temporary cholera treatment unit; 5) Maluku-Kingabwa-Massina district, Kinshasa, DRC: cholera treatment centers of Kingabwa and Malaku and the cholera treatment unit of Massina; 6) Abidjan, Koumassi-Port Bouet, Vridi District, Cote d’Ivoire. The infectious disease and paediatric departments of Port Bouet and Koumassi Hospitals and the temporary cholera treatment center at the Vridi Health Centre; 7) Adiake prefecture, Cote d’Ivoire: Adiake general hospital and temporary treatment centers; 8) Mbale-Manafwa-Buteleja district, Uganda: Nabiganda health center, Namatela health center and Busiu health center; 9) Kasese district, Uganda: Bwera hospital, Kayangi health center, Kagando hospital, Kinyamaseke health center and Kitholhu health center and other temporary treatment centers; 10) Beira, Mozambique: Ponta-Gea health center, Macurrungo health center, Munhava health center, Macurrungo and the central hospital of Beira; 11) Pemba city, Mozambique: the temporary cholera treatment center of Pemba city.
In the enhanced surveillance zones and outbreak sites, the MoH teams collected data at health centers level using the same standardized data collection forms, which included sex, age, location, date of symptoms, culture results but also clinical information such as watery diarrhea, rice water stool, vomiting, dehydration. We identified all deaths among patients admitted to a cholera treatment facility. We did not include deaths occurring in the community or after treatment center discharge. In parallel, the MoH continued to register the overall number of suspected cases in their routine surveillance system using line lists with a limited number of variables (date of onset, district, age and sex). We used district–level population estimates for 2011 or 2012 that corresponded to the geographic area under surveillance. The 2011 and 2012 population estimates were derived from the last census data (Uganda, 2002; DRC, 1983; Togo, 2009; Guinea, 1996; Cote d’Ivoire, 1998; Mozambique, 2007), updated each year by district health officers based on estimated national annual population growth rates.
National public health laboratories in each country performed culture confirmation of suspected cases. Cholera polymerase chain reaction (PCR) testing was not available in any of the included countries. We aimed to collect whole stool or rectal swabs from all suspected cases. In practice, the proportion of cases with a collected stool varied according to context. During large outbreaks when laboratory capacity could become overwhelmed, local staff were advised to collect the first ten cases per day only. Samples were transported in Cary-Blair transport medium to the country national reference laboratories. There, they were enriched in alkaline peptone water and plated on thiosulfate-citrate-bile-salt-sucrose (TCBS) agar. Characteristic yellow colonies were sub-cultured in non-selective medium. Resulting colonies were tested for oxidase and, if positive, considered confirmed and serogrouped. External quality control was performed by the National Institute of Communicable Diseases in South Africa using PCR.
We adopted the definition of rainy season from the World Bank climate portal (sdwebx.worldbank.org/climateportal; accessed 2013) as follows: Uganda, Mar–Jun and Sept–Nov; Goma, DRC, Jan–May and Sept–Dec; Kinshasa, DRC, Jan–May and Oct–December; Mozambique, Oct–Mar; Cote d’Ivoire, May–Jun and Oct–Nov; Guinea (Maritime region), May–Nov; Togo (Maritime region), Apr–Jul and Sept–Nov.
Suspected and confirmed cholera cases were summed by age group, sex, occurrence during the rainy season and clinical symptoms. We calculated the crude and corrected incidence rates for confirmed cases. Correction was done as follows: 1) for lack of culture testing, we extrapolated the proportion of culture positive results among suspect cases tested by culture to all notified suspect cases in each geographical area; 2) because culture has a sensitivity of 66% (compared to combined results from culture, dipstick, direct fluorescent antibody, multiplex-PCR and Vibrio cholerae O1 El Tor specific lytic phage on plaque assay as gold standard) for imperfect reported culture testing, we extrapolated the number of cultures that would have been positive if culture had a sensitivity of 100% [3]. For point 2, we conducted a literature search and identified few studies that reported culture sensitivity relative to another gold standard, as culture itself has been the gold standard for years. Consequently the study by Alam et al. was used as an approximation, recognizing that the included data may not be definitive. For calculation of case fatality ratios (CFR), we included in the denominator patients admitted to a cholera treatment center with cholera symptoms and as the numerator all deaths that were identified at the treatment center Comparisons between groups were performed using Pearson’s chi-square test. Graphs were produced with R open-access software. Statistical analyses were performed using STATA software (version 12.1, College Station, Texas 77845 USA).
Africhol provided technical and financial resources to national MoHs to support cholera surveillance. Cholera is part of the national public health surveillance through the integrated disease surveillance and response system supported by WHO. The Africhol protocol was approved and implemented by the MoH of each country. The Togolese government further elected to submit the protocol for approval to a local Togolese institutional review board (IRB). The remaining countries did not seek IRB approval as they considered that they were conducting epidemic disease surveillance and response. covered by national public health laws as an integral part of the public health mandate of the MoH and associated executing agencies.
From June 2011 to December 2013, 13,377 suspect cholera case were notified: 47% (6343) occurred in surveillance zones in Goma, DRC and 34% (4585) in Conakry, Guinea (Table 1). We tested 26% (3536) of all suspected cases by culture, a figure that increased to 49% when excluding zones in Goma and Conakry, which both experienced large outbreaks in August 2012 and which respectively had testing only 7.4% and 0.5% of cases during this period (Fig 1 and Fig 2).
In the surveillance zones, a median of 31% of cases were culture positive ranging from 37% in Conakry, Guinea to 0% in Beira, Mozambique (Table 1).
With the exception of Adiake prefecture in Cote d’Ivoire, suspected cases were equally distributed by sex (Table 2).
However, confirmed cases were more likely to be male (Table 3). The proportion of suspected cases aged under five years ranged from zero percent in surveillance zones in Abidjan, Cote d’Ivoire to 40% in Beira, Mozambique (Table 2); for confirmed cases, the proportion aged under five years peaked at 29% in Goma, DRC. From 45–99% of suspected and 70–100% of confirmed cases occurred during the rainy season (Tables 2 and 3).
The monthly distribution of cases in Goma-Karisimbi districts (DRC), Mbale-Manafwa-Butaleja districts (Uganda), Lome and Golfe districts (Togo), Kasese district (Uganda) and Maluku-Kingabwa-Massina districts (Kinshasa, DRC) showed that cases with Vibrio cholerae identified by culture can be observed before the rainy season starts (Figs 1 and 2).
The mean proportion of persons presenting with watery diarrhea at each site was 91% (SD 7%) and 82% (SD 16%) had vomiting. The percentage presenting with rice water stool varied from <1% to 86% and with dehydration from 33% to 99% (Table 4).
We identified three epidemiological patterns (Figs 1 and 2). In surveillance zones in Goma (DRC), confirmed cases were seen continuously throughout the surveillance period. In zones in Lome (Togo), Mbale (Uganda) and Conakry (Guinea), there were sporadic confirmed cases plus additional outbreaks at irregular intervals. Lastly, in Beira, Mozambique and Abidjan, Cote d’Ivoire, there was a history of recurrent cholera epidemics in the period leading up to Africhol implementation but as of the end of 2013, no confirmed cases had been identified for 30 months and 17 months, respectively.
Annual confirmed incidence of cholera presenting to a treatment facility per 10,000 population was <0.5 in surveillance zones, except in Goma where it was 4.6. Goma and Conakry had corrected incidences of 20.2 and 5.8 respectively, while the remaining surveillance zones had a median corrected incidence of 0.3. During outbreaks, the annualized confirmed incidence of cholera presenting to a treatment facility ranged from 0.3–3.3 and corrected incidence from 2.6 to 13.0 per 10,000 population (Table 5). The ratio of the mean annual corrected incidence of confirmed cholera to the incidence of suspected cholera varied from 0.1 in Abidjan to 0.6 in Conakry while it was of 0.5 (SD 0.1) in outbreak sites.
Of 5980 suspected cases identified in a treatment facility with a documented outcome, 69 died. The median CFR was 1.1% [IQR: 0.7–4.3]. The CFR varied from zero percent in Abidjan, Cote d’Ivoire to 10% in Lake district, Togo (Table 6). We found no statistical differences in the CFR between confirmed and non-confirmed cases. However we observed that deceased patients were less likely to have received culture testing than those alive at discharge (35.3% vs. 55.6%, chi-square p. value = 0.001).
In the Africhol surveillance zones, we found an overall annual corrected incidence of confirmed cholera presenting to a treatment facility of 0.3 cases per 10,000 population, which increased to 20 cases per 10,000 during large epidemics. Strong spatial and temporal clustering occurred, with most cases from surveillance zones in Conakry, Guinea and Goma, DRC. Within our study many suspected cases were not cholera confirmed by culture. Furthermore the CRF measured at clinic level remained low in our surveillance sites. From the surveillance data collected in our sites, we were able to identify three epidemiological patterns of cholera: confirmed cases throughout the year such as Goma (DRC); sporadic cases plus additional outbreaks at irregular intervals such as in Lome (Togo), Mbale (Uganda), and Conakry (Guinea); and history of recurrent cholera epidemics but no cases during the surveillance period, such as Beira (Mozambique) or Abidjan (Cote d’Ivoire). Whatever the location, we found that most cholera cases occurred during the rainy season.
Our incidence estimates for confirmed cases showed similar fluctuations by place and time as those reported previously for suspected cases but are substantially lower than estimates modeled from WHO mortality strata [4–14]. In most national cholera surveillance systems, etiologic confirmation occurs only for the first suspected cases, before outbreak declaration. Subsequently, any person with acute watery diarrhea usually would be reported as a cholera case, even though some of these will have other etiologies. Consequently, syndromic surveillance–as reported by most previous studies–likely overestimates cholera incidence.
Moreover, the proportion of culture confirmed cases varied widely by site emphasizing the utility of laboratory based studies. At the extreme, in Beira, Mozambique, where a history of large outbreaks likely led providers to have a high index of suspicion for cholera, all sampled suspected cases remained negative for V. cholera [11].
The wide variation we found may have resulted from differences in health care seeking behavior, health care access, type and extent of available health structures, health work training, and adherence to case definitions. For instance, treatment centers in Goma, DRC provided care for patients with any diarrheal disease regardless of etiology, did not charge fees, and treated persons of all ages. In other Africhol sites, cholera treatment centers offering free treatment were established only when authorities declared the outbreak. These issues also may have led to the differences in health care access behaviors and therefore to clinical presentation across sites. Other factors may lead to underestimation of incidence. For example, not all patients will present for care at a medical facility and data collection and reporting may be incomplete. However, our system was not designed to assess these issues.
While our incidence rates were lower than those from early reports, CFRs for confirmed cholera cases were consistent with those for suspect cases attending health facilities [5,11]. The low identified CFRs emphasize the great strides some cholera endemic countries have made in identifying outbreaks rapidly and improving clinical management. They might also reflect the sensitization of populations in high-risk areas to the importance of seeking timely medical care. Our CFR estimates were limited by our inability to assess deaths in the community which contribute to potential underestimation. Lastly, both our CFRs and overall incidence rates were limited by lack of active community-based surveillance, an objective for which our work was not funded. It is likely that this problem was particularly large for deaths: for example, a study from Kenya found that most deaths occurred among persons who had not sought treatment [15]. Future geographically focused studies might address this issue. In theory, health utilization surveys and capture-recapture analysis could help with estimation of surveillance system sensitivity. However, in epidemic cholera prone settings in Africa, health care utilization surveys are seldom appropriate given the lack of human resources relative to the immediate priority of outbreak control. Capture-recapture analyses similarly are not feasible, given the fluid nature of a surveillance system in which cholera treatment centers are established and dismantled relative to cholera case counts.
We identified three epidemiological patterns of cholera in our sites: those with confirmed cases throughout the year such as Goma (DRC); those with sporadic cases plus additional outbreaks at irregular intervals such as in Lome (Togo), Mbale (Uganda), and Conakry (Guinea); and those with a history of recurrent cholera epidemics but no cases during the surveillance period, such as Beira (Mozambique) or Abidjan (Cote d’Ivoire). The presence of sporadic cases without ensuing outbreaks may occur from occasional introduction of infected persons into a low risk community, e.g., a community with recent cholera and a high degree of population immunity or a community with good water and sanitation infrastructure. By contrast, sustained occurrence of confirmed cases may result from ongoing environmental source contamination from which a continuously renewed susceptible, non-immune population is infected; this may have occurred in Goma, which has experienced several waves of immigration due to regional conflicts.
We found that most cholera cases occurred during the rainy season. However the presence of cases before the rain start suggests that the rainy season may play a role of outbreak amplificatory. Previous studies have found similar results [16]. Substantial precipitation can cause flooding and subsequent mixing of drinking water (pond, well, lake, river) with sewage in areas with poor sanitation [17]. Alternatively, the rainy season may trigger human movement, such as the seasonal migration of fishermen along the West African coast or in interior lakes [16,18–20].
Our study had several limitations other than those mentioned above. We report data from only eleven geographical sites located in six countries and this may not be generalizable to other African settings. Our correction of incidence based on the lack of testing was applied uniformly across the surveillance period without taking into account seasonal variations. We used a single value to correct for culture sensitivity although culture results may vary by setting based on factors such as laboratory technician skills and stool collection and transportation methods. Finally, CFRs were difficult to assess for confirmed cholera cases because of lack of testing.
In the African cholera context, oral cholera vaccine may provide an important short- and medium-term prevention and control measure in addition to case management and long-term efforts to improve water, sanitation and hygiene (WaSH Despite the utility of mass OCV campaigns have been already demonstrated in some African areas, it remains difficult to determine the best strategy to use and if a relatively circumscribed immunization campaign can prevent an epidemic on the scale of Zimbabwe or Haiti [21–25]. Short duration and geographically focal outbreaks as described in our results will make reactive OCV use challenging, as it was the case in Mozambique [11]. Even in settings with large outbreaks such as Goma or Conakry, cases may occur over a brief period in relatively small geographic areas, such as districts. Preventive immunization may be indeed more appropriate to reduce cholera in target communities, with a potential secondary benefit of reducing transmission outside the target zone… We might also learn from Neisseria meningitidis (Nm) meningitis in the meningitis belt [26]. As with cholera, Nm outbreaks were often highly focal, of short duration, difficult to predict, and occurred in areas with limited laboratory facilities. The strategy for years was reactive campaigns following notification of an epidemic. However, vaccine frequently arrived after the epidemic peak and thus its overall efficacy questioned. This situation changed with the introduction of a low-cost Nm serogroup A conjugate vaccine (MenAfriVac) through national preventive immunization programs via mass campaigns into persons 1 to 29 years of age [27]. The analogy between OCV and MenAfriVac is also based on the need for national and international commitment for an evidence-based prevention strategy, availability of low-cost vaccine produced in sufficient quantity, and the availability of adequate financial and human resources.
While limited to health care facilities, our study presents some of the only prospectively obtained incidence data currently available for Africa. Our findings suggest that confirmed cholera burden is substantially lower than that reported from previous studies based on suspected cholera cases, and that incidence varies substantially over time and place. Efficient use of resources, such as vaccines, could be enhanced by better definition of cholera hot-spots, community behaviors that contribute to cholera spread, and high risk populations, particularly those likely to contribute to seasonal cholera spread.
Because of the frequent occurrence of non-cholera causes of diarrhea in cholera endemic zones, development of public health strategies would benefit from reinforcement of local laboratory capacities for diagnosing Vibrio cholerae, something that also would benefit from development of better low-cost diagnostic methods. Environmental reservoirs should be identified and mitigation strategies developed. Determination of other diarrheal disease etiologies across all age groups will help determine the utility of etiology specific interventions. OCV interventions must be conducted, monitored and evaluated to better assess their cost-effectiveness and their health impact among at-risk populations in African contexts. Finally, there is a role for evaluation of low-cost water and sanitation improvements within an integrated strategy for cholera prevention and control.
|
10.1371/journal.ppat.1006853 | Dynamic remodeling of lipids coincides with dengue virus replication in the midgut of Aedes aegypti mosquitoes | We describe the first comprehensive analysis of the midgut metabolome of Aedes aegypti, the primary mosquito vector for arboviruses such as dengue, Zika, chikungunya and yellow fever viruses. Transmission of these viruses depends on their ability to infect, replicate and disseminate from several tissues in the mosquito vector. The metabolic environments within these tissues play crucial roles in these processes. Since these viruses are enveloped, viral replication, assembly and release occur on cellular membranes primed through the manipulation of host metabolism. Interference with this virus infection-induced metabolic environment is detrimental to viral replication in human and mosquito cell culture models. Here we present the first insight into the metabolic environment induced during arbovirus replication in Aedes aegypti. Using high-resolution mass spectrometry, we have analyzed the temporal metabolic perturbations that occur following dengue virus infection of the midgut tissue. This is the primary site of infection and replication, preceding systemic viral dissemination and transmission. We identified metabolites that exhibited a dynamic-profile across early-, mid- and late-infection time points. We observed a marked increase in the lipid content. An increase in glycerophospholipids, sphingolipids and fatty acyls was coincident with the kinetics of viral replication. Elevation of glycerolipid levels suggested a diversion of resources during infection from energy storage to synthetic pathways. Elevated levels of acyl-carnitines were observed, signaling disruptions in mitochondrial function and possible diversion of energy production. A central hub in the sphingolipid pathway that influenced dihydroceramide to ceramide ratios was identified as critical for the virus life cycle. This study also resulted in the first reconstruction of the sphingolipid pathway in Aedes aegypti. Given conservation in the replication mechanisms of several flaviviruses transmitted by this vector, our results highlight biochemical choke points that could be targeted to disrupt transmission of multiple pathogens by these mosquitoes.
| The Aedes aegypti mosquito transmits arboviruses that cause dengue, Zika, chikungunya and yellow fever. These viruses are endemic in tropical and subtropical regions of the world placing 2.5 billion people at risk of infection. Transmission is critically dependent upon the replication of these viruses in both human and mosquito hosts. Successful viral replication is greatly influenced by the biochemical environment of the host cell or tissue and flaviviruses rearrange this environment to benefit their needs. Host-cell derived metabolites such as lipids, sugars and amino acids are utilized to produce progeny virions, help evade the host immune system and enable successful completion of the life cycle. In this study, we applied high-resolution mass spectrometry to understand the alteration of the biochemical landscape of the mosquito during infection by dengue virus. We focused on the mosquito midgut, as this is the initial site of infection. We identified several metabolites that exhibited dynamic profiles during the course of viral infection and replication. By pinpointing biochemical “choke points” required for viral replication, we can devise strategies that will stall virus replication in the mosquito and prevent its transmission to humans.
| The transmission cycle of dengue viruses (DENV) require a human host and mosquito vector. Mosquitoes acquire DENV via feeding on the blood of an infected human. The blood meal is deposited in the midgut of the mosquito and infection is first established in the midgut epithelium [1]. While digestion of the blood meal is complete within 48 hours [2], viral replication in the midgut tissue reaches its peak only at 7–8 days post-blood meal (pbm) ingestion [1]. Subsequently, the virus disseminates from the midgut and infects other tissues including the fat body and salivary glands. Approximately 10–14 days pbm the salivary glands become infected and the virus can be transmitted in the saliva to a human when the mosquito acquires another blood meal. Since the successful transmission of this virus depends greatly upon its ability to replicate efficiently in several mosquito tissues, the local biochemical and physical environment of each tissue plays a critical role in virus propagation.
In both human and mosquito cells, lipids play an integral role in the life cycle of DENV [3–7]. Host-derived membranes are incorporated into a lipid envelope that surrounds the capsid protein and genomic RNA of DENV particles [8]. This membranous structure facilitates virus release from infected cells by budding into the endoplasmic reticulum and re-entry into new cells through fusion of virus-host membranes [7, 9, 10]. Additionally, electron tomography has revealed that significant rearrangements of host cell membrane architecture occur upon DENV infection [4, 7, 11–13]. Virus-induced membrane structures are required as platforms for virus replication and assembly and protect replicating genomes from antiviral defense mechanisms of the host. They have been identified in DENV-infected human and mosquito cells and the midgut epithelium and salivary glands of Culex mosquitoes infected with West Nile virus [14]. This intracellular membrane reorganization imposes a significant metabolic cost to the host cell. It requires activation of biosynthetic processes, and the trafficking and degradation of lipids and other related molecules. Our previous studies investigated the perturbation of lipid homeostasis in C6/36 mosquito cells following infection with DENV [5]. We identified lipids involved in maintaining the stability, permeability and curvature of membranes, as well as bioactive lipid molecules that were significantly changed upon infection. Specifically, a burst of glycerophospholipids was observed coincident with viral replication kinetics. This burst of lipids was attributed to the activity of fatty acid synthase (FAS), a key enzyme in glycerophospholipid biosynthesis. Inhibition of this enzyme was detrimental for DENV replication in both human and mosquito cells [5, 6]. Collectively, these data demonstrate that lipids play critical roles in DENV infection in cell culture models.
Although lipid biochemistry has been studied in mosquitoes for several years, very little is known about the relationship between virus and mosquito host and the alteration of intracellular lipids that underpins infection capacity. In this study, we used high-resolution mass spectrometry to explore metabolic changes in the midgut of Ae. aegypti exposed to DENV-containing blood meals as this tissue represents the crucial site of initial viral replication. Using a time course study, we compared metabolic profiles of infected and uninfected midguts at early-, mid- (peak viral replication) and late time points post-infection. Our results demonstrate significant fluctuations in molecules that function as membrane building blocks, bioactive messengers, energy storage molecules and intermediates in lipid biosynthesis and lipolysis pathways. Presumably these changes represent both manipulation of cellular resources for viral replication as well as the cellular response to infection. They may result from either de novo biosynthesis or the consumption/conversion of metabolites. Additionally, import/export may contribute to the perturbation of metabolite pools. Import pathways are specifically important in the mosquito for molecules such as cholesterol and its derivatives since they cannot be synthesized de novo [15, 16]. Many unidentified metabolites (not found in currently available databases that include animal, plant, fungal and bacterial metabolites and synthetic compounds) were also significantly perturbed during virus infection and may represent mosquito-specific metabolites that have not yet been annotated. The flaviviruses transmitted by this vector share similar replication mechanisms; identifying metabolic bottlenecks that condition vector competence and transmission of these pathogens could be exploited for novel transmission-blocking interventions for control of globally important diseases.
To profile the alterations in the metabolic environment of the mosquito midgut over time post-blood feeding and during the course of DENV infection in the mosquito vector, Ae. aegypti Chetumal strain mosquitoes were fed an infectious blood meal containing DENV type 2 strain Jamaica-1409 (JAM-1409; blood titer 1x107 PFU/mL) or exposed to a noninfectious blood meal as a negative control (S1 Fig). Midguts were dissected from the mosquitoes on days 3, 7 and 11 pbm representing early viral infection in the midgut (day 3), high replication activity in the midgut with early dissemination to the salivary glands (day 7), and high replication activity in salivary glands and other tissues with virus being cleared from the midgut (day 11). Immediately after dissection, individual midguts were tested for DENV RNA by qRT-PCR (Fig 1A). To ensure that the metabolic profile of the midgut tissues accurately represented the infected environment, every individual midgut was evaluated by qRT-PCR for DENV RNA and only infected tissues were pooled. On day 3 pbm, viral genomes were detected in 55% of the dissected midguts, and on days 7 and 11 pbm, detection increased to 73% and 76% respectively. Plaque assays of whole mosquito homogenates harvested on day 11 pbm (n = 30) indicated that 60% of mosquitoes showed detectable infectious virus averaging 8.5x103 PFU/mL (Fig 1B). These were similar infection rates to those reported previously by Salazar et al, 2007 [1]. Two independent pools of ~100 tissues were included for each treatment group and time point representing a total of 200 biological samples. The metabolic profiles observed in the infected tissues represent true biological deviations corresponding to the infected state. It should be noted however, that estimation of variance and statistical power are compromised by the small sample sizes since the maximum limit of samples acquirable in order to maintain metabolite integrity at a given time point is limited.
Mosquito metabolites in both polar and non-polar phase extracts were analyzed by LC-MS in both positive and negative ionization modes to obtain the most comprehensive coverage possible. Experimental data from each extract in each mode were analyzed separately by the workflow described in materials and methods and shown in S1 Fig. Only features (molecular identities defined by a unique mass to charge ratio, m/z and retention time) detected in both pools (representing 200 biological replicates) were considered ‘present’ in the samples. For the midguts, a total of 6,103 molecular features from all modes, treatments and time-points pbm were detected. These features were subjected to statistical analysis to compare the differences in metabolite abundance between DENV-infected and uninfected midguts at each time point pbm. The features that showed statistically significant differences in abundance (|log2 fold change|) ≥ 1 and p-value < 0.05; total of 936 features) were identified using LIPID MAPS, HMDB, and Metlin databases. Approximately 39% of the significantly different features (363 molecules) were identifiable to an accuracy of < 6 parts per million (ppm) (Table 1 and S1 Table). About 93% of identified and significantly different features (341 molecules) have recognizable biological relevance. The metabolites detected in the non-polar phase from both negative and positive ionization modes comprised glycerophospholipids (GPs), sphingolipids (SPs), glycerolipids (GLs), sterols, fatty acyls, acyl-carnitines, prenols, polyketides, amino acids and peptides and other organic compounds. The metabolites detected in polar phases from both detection modes also contained the above molecular species with the exception of GLs and contained nucleosides instead. All nomenclatures and categories for lipids are based on the LIPID MAPS comprehensive classification system [17].
To explore the metabolic environment in the mosquito midgut during DENV infection, we compared the metabolite repertoire of DENV-infected midguts to uninfected controls. A summary of the data is shown in Fig 2 and S1 Table. Of 6,103 features detected, 936 features (15%) have differential abundance in DENV-infected midguts compared to uninfected controls in at least one time point (Fig 2A and 2B). The Venn diagram shows the numbers of molecular features with differential abundance in the midgut (|log2 fold change| ≥ 1 and adjusted p-value < 0.05) upon DENV infection on days 3, 7, and 11 pbm. The profiles of 5,167 features were unaltered at any day post-infection (Fig 2A and 2B). Among the features with altered levels, 274, 283 and 114 features showed differential abundance only on day 3, 7 or 11 pbm, respectively, whereas 98 features showed differential abundance in all 3 days (Fig 2B).
Identified metabolites: About 39% of features (363 out of 936) with differential abundance were identifiable through use of three metabolite databases and the putative identifications were categorized into different metabolite classes. The metabolic relationships between the lipid classes are mapped in Fig 2C. Fatty acyl-CoA can be synthesized de novo. It is a precursor that can be modified and incorporated into more complex lipid molecules such as glycerophospholipids, sphingolipids, glycerolipids and sterols. On the other hand, fatty acyl-CoA can be degraded through β-oxidation when cells require energy. The overall trend of lipid molecular levels that changed upon DENV infection is summarized in Fig 2D. A majority of the metabolites were unchanged (had intensity changes that are less than 2-fold upon infection or p ≥ 0.05), while a majority of those that showed greater than 2-fold changes in intensity had higher abundances upon DENV infection (Fig 2D and S1 Table, identified tab). Only 61 out of 363 features (~17%) decreased in abundance during DENV infection.
Unidentified metabolites: Interestingly, a large number (~61%) of metabolites that were differentially altered in abundance during DENV infection were unidentified (S1 Table, unidentified tab). A majority of these unidentified metabolites had elevated abundance in the DENV-infected samples compared to uninfected controls at all three time points post-infection. These metabolites represent a potential resource since they highlight molecules that are not found in currently available databases that include animal, plant, bacterial and fungal metabolites as well as synthetic compounds. Therefore, they could be compounds unique to Diptera, mosquitoes or Aedes species that are yet to be annotated. Given their importance to DENV infection, they should be exploited in the future as possible transmission-blocking control points.
GPs are major components of cellular membranes. GPs are composed of a polar head group, a phosphate group and two fatty acyl chains. The specific composition of GPs in a membrane influences membrane fluidity, leakiness, nutrient exchange, assembly and function of signaling protein complexes and vesicular traffic [18–21]. In insects, the GP composition of membranes also affects tolerance to changing environmental conditions [22]. Pathogenic or endosymbiont infections also drive changes in the intracellular environment and could influence the regulation of GP content in membranes [5, 23–26]. As a point of reference, the metabolic pathways of the major classes of GPs are shown in Fig 3A. Our study identified 565 species of GPs in Ae. aegypti midguts with 87 species that were altered upon DENV-infection. The alteration landscape and distribution of the different GPs are shown in Fig 3B and 3C.
Minimal enrichment of GP species and levels can be observed on day 3 pbm, during the initial establishment of viral infection in the midgut (Fig 3B). This enrichment was significantly enhanced on days 7 and 11 pbm. These observations were coincident with known viral infection dynamics (peak viral replication in the midgut on day 7 and viral dissemination to other tissues and clearance from the midgut by day 11) [1]. The signaling GPs, PIs, were the most diverse (Fig 3C) closely followed by PGs, important as surfactants and precursors of cardiolipins in mitochondrial membranes. PCs, PEs and PSs, which are primary components of most cellular membranes were also increased during infection especially at the peak of viral replication (day 7 pbm). Interestingly, levels of most of the lyso- or short chain GPs were decreased during infection.
GLs, including mono-, di- and triacylglycerols (MAG, DAG and TAG), are critical effectors of energy metabolism in insects and mammals. Fatty acids absorbed from a digested blood meal or synthesized de novo from a sugar meal can be converted into DAG and TAG that can be transported to the fat body for energy storage or the ovaries for vitellogenesis [27–31] (Fig 4). DAG is also an important intermediate in GP synthesis [18, 32] and a critical second messenger regulating cell proliferation, survival, mitochondrial physiology, gene expression and apoptosis [33, 34]. Seventy species of GLs were identified in our study, of which eight GLs were significantly changed in abundance upon infection. Most of the MAG, DAG and TAG levels were higher during early time points (day 3 and 7 pbm) in DENV-infected midguts (Fig 4). This might be a result of transportation of these lipids from storage tissues to support the demand for lipids during infection [35, 36]. Interestingly, the level of cytidinediphosphate (CDP-DAG (37:1)) were elevated only during the later phases of infection (days 7 and 11 pbm). CDP-DAG is a precursor for GP synthesis (Fig 3A). As a result, increased CDP-DAG (37:1) might support the high demand for GP synthesis on days 7 and 11 pbm during virus infection.
SPs are bioactive molecules that play important roles in the structural composition of cellular membranes and numerous cell-signaling pathways and are critical in microbial pathogenesis [22, 37–40]. De novo synthesis of SPs occurs in the endoplasmic reticulum through the condensation of L-serine and palmitoyl-CoA to form ceramide (Cer) via several intermediates [41, 42]. Cer is a precursor for the biosynthesis of several complex SPs such as sphingomyelin (SM) and glycosylceramides or the production of fatty acids and sphingosines through its hydrolysis [43–45]. The SP pathway has been highlighted in many studies in mammalian cells [37] and in mosquito cells [5] as a lipid metabolic pathway altered by flavivirus infection. In this study, using the identified SPs from both infected and uninfected mosquito tissues we reconstructed a significant proportion of the SP metabolic pathway for Ae. aegypti (Fig 5). Several members of the pathway were identified at all three time points in both biological replicates of the infected and uninfected tissues. We observed accumulations of several SPs such as sphinganine, sphinganine-1-PC, sphingosine, Cer and hexosylceramide that were elevated in DENV-infected midguts on days 3, 7 and/or 11 pbm. GP-Cers, especially PE-Cers, are sphingomyelin analogues. They function as the principal membrane sphingolipids in Drosophila, which lacks sphingomyelin synthase [46]. In mosquito midguts, we found decreased abundance of PE-Cer (d32:2(2OH)) upon DENV infection on day 3 pbm but an increase in the levels of PE-Cer(d33:2(2OH)), PE-Cer(d36:1), PE-Cer(d36:3(2OH)) and PE-Cer(d36:2(2OH)) on day 7 and/or 11 pbm. Interestingly, sphingomyelin was found in Ae. aegypti mosquito midguts and in cells in culture (Fig 5 and S2–S4 Figs); however, the levels of sphingomyelin were not changed upon DENV infection at any of the time points tested. Glycosylated ceramides (hexosylceramide and hexosylceramidesulfate) play critical roles in modulating cellular signaling and gene expression resulting in changes in processes such as proliferation, apoptosis, autophagy and endocytosis [47]. The abundance of the neutral glycosphingolipid, FMC-6(d40:1(2-OH)), was increased on all days and HexCer(d40:1(2OH) sulfate was increased on day 3 following DENV infection (Fig 5).
In this study, we identified a large number of SPs with significantly altered abundance following DENV infection of mosquito midguts. Important biological roles of the SP pathway are determined by conversion from DHCer to Cer, the central precursor of diverse SPs. The balance between Cer and DHCer concentrations plays a critical role in physiologically regulating properties such as membrane architecture, fluidity and function [48]. Therefore, we investigated if alteration of Cer or DHCer levels and the Cer/DHCer ratio had an effect on DENV infection, using Ae. aegypti-derived (Aag2) cultured cells, which are readily amenable to loss of function studies. Cells were treated with 4-hydroxyphenyl retinamide (4HPR) or long double stranded RNA (dsRNA) to inhibit activity of or knock down expression of sphingolipid Δ-4 desaturase (DEGS, VectorBase: AAEL013047, NCBI:LOC5577178), the enzyme that catalyzes conversion of DHCer to Cer, the final step in the de novo biosynthesis of Cer (Fig 5), and thus should greatly influence the Cer/DHCer ratio [49]. 4HPR is a synthetic retinoid that activates the Cer de novo synthesis pathway prior to DHCer synthesis, but inhibits the activity of DEGS [50] (Fig 6A). Significant reduction of virus titer and genome replication were observed upon 4HPR treatment of Aag2 cells at non-cytotoxic concentrations (S2A–S2C Fig). We validated the metabolic impact of 4HPR on sphingolipid metabolism using multiple reaction monitoring (MRM) LC-MS/MS. We observed accumulation of Cer(d18:1/16:0), Cer(d18:1/18:0), Cer(d18:1/24:1), and Cer(d18:1/26:1) and DHCer(d18:0/18:0) in cells following 4HPR treatment (S2D Fig, lower). However, their accumulation did not alter the Cer/DHCer ratios (S2D Fig, upper). None of the levels of long chain sphingoid bases Cer and DHCer with 16-carbon backbones were affected (S2E Fig). We also did not see changes in sphingosine (d18:1), sphingosine-1-P (d18:1-P), sphinganine (d18:0) or SM levels following 4HPR treatment which is expected as the enzymes that produces these molecules are not the primary targets of 4HPR (S2F and S2G Fig). The effect of 4HPR on DENV replication therefore might not be due to a direct effect on the Cer/DHCer ratios, but due to off target effects such as ROS, changes in bcl-2, p53 mRNA expression and apoptosis [51]. A similar observation has been made in mammalian cells [50].
To validate that the reduction of DENV replication was caused by the loss of SP metabolites or Cer/DHCer imbalance and not due to off-target effects of the inhibitor, the expression of DEGS in Aag2 cells was transiently knocked down by RNA interference, using long dsRNA (Fig 6B–6E). In DEGS knockdown (DEGS-KD) cells, DENV titer and genome replication were significantly reduced compared to the GFP dsRNA (GFP-KD) negative control, but were similar to the DENV dsRNA (DENV-KD) positive control (Fig 6B and 6C). SP metabolic profiling of the DEGS-KD cells compared to GFP-KD cells revealed striking reduction of Cer(d18:1/18:0), DHCer(d18:0/20:0) and DHCer(d18:0/22:0) after DEGS-KD (Fig 6F and 6G, lower panel). Interestingly, DEGS-KD also greatly altered the Cer/DHCer ratio for molecules with 18- and 16-carbon long chain sphingoid bases, especially those with very long fatty acyl chain lengths (20:0–24:0) (Fig 6F and 6G, upper panel). These results indicated that perturbation of the Cer-DHCer balance reduced both DENV genome replication and infectious virus formation. As expected we did not see changes in sphingosine (d18:1), sphingosine-1-P (d18:1-P), sphinganine (d18:0) or SM levels during dsRNA treatment (S3 Fig). When DENV-infected and uninfected Aag2 cells were compared for alterations in the SP pathway, we observed a significant increase in the abundance of Cer(d18:1/20:0) and Cer(d18:1/24:1) and DHCer(d18:0/22:0) in infected cells, but a decrease in Cer(d16:1/16:1) in infected cells (S4A and S4B Fig, lower). These changes did not, however, affect the Cer/DHCer ratio for very long fatty acyl chain containing molecules (S4A and S4B Fig, upper). We did not see changes in sphingosine (d18:1), sphingosine-1-P (d18:1-P), sphinganine (d18:0) or SM levels during dsRNA treatment (S4C and S4D Fig).
Fatty acyls can be acquired from the midgut digestion and absorption of a blood meal or can be de novo synthesized from central carbon metabolism precursors [27, 52]. Fatty acids that are covalently linked to coenzyme A, fatty acyl-CoAs, can be incorporated into complex lipids such as GPs, SPs and GLs that have structural roles in membranes. In addition, free fatty acids or fatty acyls that are linked to other functional groups can have critical roles in signaling, energy homeostasis and the immune response in insects and mammals (e.g.: fatty amides and eicosanoids) [53–55]. In this study, at least nine subclasses of fatty acyls were detected and many of these had higher abundances in DENV-infected midguts compared to uninfected controls in at least one time point pbm (S5 Fig and S1 Table). The molecules observed were fatty amides, hydroxy fatty acids, free fatty acids, eicosanoids and leukotriene (i.e. immunomodulators), fatty-amines, glycosides, dicarboxylic acids, and keto-fatty acids. Fig 7 shows molecules with significantly increased or decreased levels at least at one time point following DENV infection and have known functions in immunomodulation (Fig 7A) and N-acylamides and other fatty acyls whose accumulation is a marker for malfunction of fatty acid oxidation (Fig 7B). Prostaglandins, leukotrienes, thromboxanes and lipoxins are eicosanoids that are synthesized in various tissues. Eicosanoids are inflammatory mediator molecules have been found to play several important roles in insect immunity [55]. In this study, three prostaglandins (prostaglandin a2, prostaglandin d2, and PGD2-dihydroxypropanylamine) and one thromboxane (dehydrodinor-TXB2) showed increased levels during DENV infection and one leukotriene (leukotriene e4) decreased during DENV infection. Anti-inflammatory molecules (resolvin d5 and epoxy-DHA) increased during DENV infection [56, 57]. Metabolites of linoleic acid that may regulate prostaglandin synthesis (HEPE, HpOTrE and TriHOME) also increased upon DENV infection [58]. N-acylamides play important roles in endocannabinoid signaling systems in mammals and have been detected and reported as potentially important signaling molecules in Drosophila [59]. Elevated levels of these molecules are also observed in the urine and blood of human patients with fatty acid oxidation disorders [60]. In our study, we found that five N-acylamide molecules (e.g. N-arachidonoylglycine, myristoylglycine, N-stearoylarginine, N-decanoylglycine, and N-undecanoylglycine) had increased levels and N-heptanoylglycine displayed a decreased level in midguts during DENV infection (Fig 7B).
One of the most prominent observations in this study is the change in abundance of acyl-carnitines in midgut tissue after DENV infection (Fig 8). These molecules are intermediates that shuttle fatty acyl-CoA from the cytoplasm into the mitochondria for β-oxidation and subsequent energy (ATP) production (Fig 8). A total of 54 acyl-carnitine molecules were detected in the midgut metabolome. Following DENV infection, 26 acyl-carnitines had a significant increase in abundance and only one had decreased abundance (Fig 8A). Interestingly, of those that increased in abundance, 25 out of 26 molecules contained medium chain-length fatty acyls of 4–12 carbons. The accumulation of acyl-carnitines with medium-length carbon chains was reported to be a result of incomplete β-oxidation [61].
Sterol lipids are a group of cyclic organic compounds composed of 17 carbon atoms arranged in a four-ring structure. In insects, cholesterol is a component of cellular membranes and a precursor of the ecdysone hormone [15, 62–65]. However, insects cannot synthesize cholesterol de novo but must absorb it from dietary sources and/or microflora [15, 16, 66]. During infection of mammalian cells, cholesterol in the flavivirus envelope seems to be critical for viral entry and fusion [67, 68]. It has also been shown that flaviviruses can manipulate cellular cholesterol homeostasis to facilitate genome replication [69]. However, the role of cholesterol in flavivirus replication in the mosquito vector has not been determined. In this study, we detected 111 sterol lipid molecules. Many of these have no reported function in insects. Since sterols in the mosquito are acquired from dietary sources (in this case, raisins, sucrose, sheep’s blood and cell culture medium), we did not exclude sterol lipids identified as plant or other animal from our analyses. Of the 111 molecules, 25 showed different levels of abundance during DENV infection compared to controls (S1 Table). Twenty-one molecules increased during infection while only four molecules decreased during infection. A majority of the changes occurred on day 3 (10 molecules) and day 7 (14 molecules) pbm. Only one molecule showed significant changes (decreased) on day 14 pbm.
The dynamic metabolic environment of an organism is an indicator of physiological changes that occur following exposure to varying environmental conditions. When arboviruses such as DENV infect a mosquito vector, they must strike a delicate balance between metabolic commensalism and competition to achieve persistent replication in mosquito tissues to allow transmission to a new host. Here, we have explored the metabolic landscape of Ae. aegypti, the primary vector of DENV, ZIKV, YFV and CHIKV, during infection with DENV. Our study focused on the midgut, the site of initial viral replication prior to viral dissemination to other tissues and transmission to the vertebrate host [1]. The observed biochemical changes highlight specific metabolites that may be markers of infection. They may be required for viral replication or produced as part of the defensive response of the vector to the invading pathogen. We have consistently observed a link between DENV infection and an increased abundance of GPs; our previous work on DENV-infected human cells showed that fatty acid synthase (FAS), a key enzyme required for fatty acid synthesis (essential pre-cursors of GP synthesis), is activated by the expression of virus-encoded nonstructural protein NS3, and relocates to sites of viral RNA replication through interaction with NS3 [6]. Inhibition of FAS activity using C75 caused a reduction in viral replication and we observed this in both cultured human and mosquito cells [5, 6]. We profiled the metabolic changes upon C75 inhibition of FAS in mosquito cells and demonstrated that the GP repertoire was directly reduced coincident with the reduction in viral replication. Therefore, in the cell culture models an abundance of GPs is required for viral replication. The current study supports these observations and reveals that DENV infection also significantly alters the abundance of GPs and other lipid-related molecules in the midgut of the mosquito vector during infection (summarized in Fig 9). While our previous studies showed that de novo synthesis of lipids plays an important role in DENV replication, other processes such as lipolysis, lipid conversion/consumption or import/export may also critically shape the metabolite pool observed in these infected tissues [70, 71].
One of our most striking observation was the marked elevation of GP levels coincident with the kinetics of viral replication in the midgut [1]. Digestion of the blood meal, is completed within ~48 hours post-ingestion, and provides a major source of amino acids and fatty acids used for the biosynthesis of macromolecules [74, 75]. These macromolecules are required for vitellogenesis (yolk formation) and energy storage [29, 76–78]. In DENV-infected mosquitoes, the rapid increase in GP content is observed at day 7 pbm when viral replication reaches its peak in the midgut. While it is currently unknown if biosynthesis alone contributes to this burst in GP content the increase is consistent with requirements for DENV replication within human and mosquito cells in culture [5, 6]. Specifically, DENV infection induces a massive expansion of intracellular membranes required for the assembly and function of viral replication complexes and assembly and release of enveloped progeny virions [4, 7]. This membrane expansion is coincident with increased GP biosynthesis through the co-opted functions of FAS, a key enzyme in the pathway. Electron microscopic studies have revealed similar membrane expansions associated with the midgut of WNV-infected Culex mosquitoes but equivalent studies have not been performed in Ae. aegypti infected with DENV [14].
Reconstruction of the SP pathway in Ae. aegypti revealed significant conservation with that of Homo sapiens, allowing us to identify a majority of orthologous metabolites in the pathway. Importantly, most metabolites were identified in both pools of 100 midguts, at all time points and in both infected and uninfected tissues confirming a significant presence in the midgut metabolome. Additionally, we were also able to identify orthologous effector enzymes for a majority of the SP metabolic pathways in our study using annotations in the AaegEL5 assembly of the Ae. aegypti genome on VectorBase [79]. SPs are important components of lipid rafts in membranes that in mammalian systems, together with cholesterol support the assembly and function of various protein signaling complexes [80]. Metabolites of SPs are well-studied signaling molecules [41]. These lipids have been studied in Drosophila [23, 81, 82] and it has been demonstrated that these lipids are critical for development, reproduction and maintenance of tissue integrity in the fruit fly [83]. SPs have also been identified in Anopheles stephensi by spatial mapping [84] and in Aedes [5] and Culex cells in culture [85]. However, the specific functionality of these lipids has not been extensively studied either in vitro or in vivo in mosquitoes. All viruses and most bacteria are incapable of SP synthesis and rely on the host to provide these lipids. The specific species of SPs that are increased in abundance during infection could influence the balance between commensalism, competition and/or host damage [37, 86].
In our data, the most striking observation is that several SP metabolites that are precursors or derivatives of Cer showed increased abundance in DENV-infected midguts compared to controls. In Ae. aegypti-derived cells we observed a significant increase in the abundance of select Cer and DHCer species and a reduction in other Cer species during infection. Importantly, altering the ratio of Cer to DHCer (through inhibition of DEGS) significantly reduced viral genome replication and infectious virus release, highlighting the importance of this hub for infection. In our previous studies in C6/36 Ae. albopictus mosquito cells, we also observed elevated levels of Cer and SM during DENV infection [5]. An important observation from the three systems studied thus far (Ae. aegypti mosquito midgut tissue, Ae. aegypti-derived Aag2 cells and Ae. albopictus-derived C6/36 cells) is that the SP synthetic pathway is significantly activated during infection, with the Cer concentration being a focal point. In contrast, work by Aktepe et. al., on mammalian cells showed that DENV replication was enhanced upon the inhibition of Cer synthesis [86]. Therefore, while a critical focal point during infection, the SP synthetic pathway requires further investigation to determine how the Cer hub and other intermediates in the SP metabolic pathway might support or limit infection.
During mitochondrial β-oxidation activated fatty acids (FA-CoAs) in the cytoplasm are esterified with carnitine to produce acyl-carnitines that are then shuttled to the mitochondrial matrix for energy production [87–89]. In this study, we observed a temporal accumulation of medium-length carbon chain-containing acyl-carnitines in DENV-infected midguts compared to uninfected midguts. The accumulation of medium-length carbon chain acyl-carnitines was reported to be indicative of incomplete β-oxidation [90]. Given that DENV replication requires FA-CoAs, for lipid biosynthesis and subsequent membrane expansion, it is possible that acyl-carnitine accumulation results from a manipulation of β-oxidation (an increase or decrease) in order to provide for viral replicative needs [5, 71, 86].
There are two possible hypotheses to explain the accumulation of medium-chain length acyl-carnitines during DENV infection (Fig 10). In hypothesis I, accumulation of acyl-carnitines might be caused by inhibition or blockage of β-oxidation in the mitochondria reducing the production of ATP (Fig 10I). This phenomenon has been observed in cells exposed to hypoxic conditions [91–93]. Similar effects were observed during DENV infection of human cells; in infected HepG2 cells, mitochondrial membrane damage appeared to lead to a reduction in ATP levels [94]. To maintain energy homeostasis, energy consumption is diverted towards increasing glucose uptake and glycolysis [91]. This response has also been observed in DENV-infected primary human cells [95]. Due to the stalling of acyl-carnitine transport and a blockage of β-oxidation, FA-CoA lipid partitioning may occur to divert FA-CoAs towards a synthesis of complex lipids and membrane expansion at the expense of fatty acid oxidation [92, 95, 96].
Alternatively, in hypothesis II, an accumulation of medium-length carbon chain acyl-carnitines could occur because of mitochondrial overload (Fig 10II). In this scenario, due to increased energetic demands during infection, a large proportion of FA-CoAs enter mitochondria but are only partially broken down by β-oxidation, which serves as a bottle neck for this reaction. This phenomenon was observed in human skeletal muscle insulin resistance [61]. The ATP that is produced by β-oxidation is sufficient to supply the energetic needs for producing more virions, and acetyl-CoA (2 carbon product) can be recycled for de novo synthesis of new longer-chain FA-CoA molecules directly at the site of virus replication [6]. This hypothesis is supported by Heaton et al. during DENV infection of Huh7.5 human liver cells, where they observed an increase in autophagosome processing of cellular lipid droplets and TAGs to produce free fatty acids and a stimulation of mitochondrial β-oxidation [71]. The molecular mechanisms of how mitochondrial energetics are diverted to benefit DENV replication in the mosquito midgut are still unclear and the observations in this study present the first insight into the possibilities.
FA-CoAs (activated fatty acids) form a primary hub that integrates multiple lipid metabolic pathways, some of which were found to be perturbed in this study during DENV infection (Fig 9). These molecules are key to the synthesis of more complex lipids, which are important for DENV replication in both human and mosquito cells [4–7]. FA-CoAs are also important for activating several signaling pathways in the cell [97, 98]. Interestingly, there is evidence that Wolbachia, the endosymbiotic bacterium used to control DENV infection in Ae. aegypti, might drive this metabolic balance in the opposite direction (FA-CoA to free fatty acid) suggesting opposing metabolic driving forces resulting in a competition between virus and endosymbiont [73]. Recent studies in Wolbachia-infected, Ae. albopictus (Aa23) cells have supported this hypothesis and indicated that several lipid groups such as SP, PC, and DAG previously shown to be elevated in DENV-infected C6/36 cells were depleted upon Wolbachia infection of the cells [5, 23, 24].
The hypothesis that metabolic challenge rather than immune pathways form the basis for pathogen-blocking capabilities of Wolbachia is further supported by observations by Caragata et al, 2013 [99] who showed that cholesterol might also be a limiting factor that drives metabolic competition between virus and endosymbiont. Specifically, increasing cholesterol availability via an enriched diet increases virus replication in the insect and reduces the impact of Wolbachia-mediated pathogen suppression. Mosquitoes are incapable of de novo synthesis of sterols [16, 100], and must sequester these molecules from the blood meal or from microbiota that are capable of synthesizing these lipids. In the presence of Wolbachia, it is possible that cholesterol “stealing” from the vertebrate blood meal may make cholesterol unavailable for DENV infection/replication. In mammalian systems, it has been shown that cholesterol is critical for DENV replication [67, 68], and therefore, it is interesting to contemplate what might happen in the mosquito where these lipids are limited. Alternately, other species of lipids could substitute for the function(s) of cholesterol during DENV replication in the mosquitoes. In this study, we detected over hundred cholesterol molecules with one fourth of the molecules showing alterations during early phases of DENV replication in the midgut. The specific role these molecules might play in the virus-vector interactions yet remain to be explored.
These in vivo studies are the obvious next step to previous studies that have evaluated metabolic changes induced by DENV infection in cell culture models. Interestingly, in both human [6] and mosquito cell cultures models [5] we observed a significant increase in GP expression during DENV infection. We determined that this burst of GP abundance is due to the activity of FAS that seems to be stimulated and re-localized to sites of viral RNA replication by DENV nonstructural protein 3 (NS3) [5]. We anticipated that this might also be true in vivo. Accordingly, we saw a significant increase in GPs in Ae. aegypti midguts that directly coincides with the peak of viral replication (similar to that observed in both human hepatoma and C6/36 cells) indicating that it could be a key modulator of DENV infection. In mosquitoes, in vitro (C6/36 cells) we did not see significant PG/PI involvement the during infection. However, we do see activity in mosquito midguts following infection. The concentrations of PG lipids are reportedly lower in eukaryotic membranes but higher in bacterial membranes, and therefore could be elevated in the midguts by contributions from the microbiome [101]. However, to accurately compare the specific GP landscape with respect to true concentrations of head groups, extensive targeted analyses will need to be carried out in the future. Lysophospholipids (LysoGPs) also seemed more heavily perturbed in C6/36 cells compared to mosquito midguts with evidence that phospholipase A2 may have increased activity during DENV infection [5]. A striking observation in this study is that among the GPs that were significantly altered in abundance during infection, lysoGPs were prominently decreased in abundance in infected midguts while they were increased in abundance in infected C6/36 cells. These lipids are known to be stress-related as well as to function as precursors and signaling molecules that also play important roles in influencing membrane architecture [102, 103]. These different response patterns in vitro and in vivo indicate a more complex situation in the whole animal and is evidence that in vitro observations are not always translatable to the in vivo environment.
Sphingolipids (SPs) also showed differences between the cell culture and in vivo systems. While in C6/36 cells some of the major SPs like SM and Cer were perturbed during infection, in the midgut tissues, we saw both precursors and downstream products of Cer metabolism as well as complex SPs being perturbed. Interestingly, while several SMs were detected, we did not see any significant changes in SM abundance during infection. The observation of SPs in Aag2 cells was rather uninformative, with only a few Cer and DHCer molecules showing significant differences in abundance during infection. However, alteration of the Cer to DHCer ratio in these cells by RNAi knockdown significantly reduced DENV genome replication and infectious virus release. These observations suggested that there may be a battle between the virus and the host to regulate the Cer hub that results in keeping most of the SPs in balance. Actively changing this hub can significantly impact the viral life cycle. A caveat is that when comparing metabolic responses of these different in vivo and in vitro models systems to DENV infection, it is important to take into account the inherent differences between the systems that might impact this response. For example, C6/36 cells, are a cell line derived from Ae. albopictus larvae (a different species to Ae. aegypti), and are defective in the RNAi pathway [104]. The Aag2 cell line, while derived from Ae. aegypti embryos are persistently infected with cell fusing agent virus (CFAV) [105]. Our in vivo studies were carried out in adult female Ae. aegypti midgut tissue, representing a much more complex and biologically relevant cellular environment.
We present the first comprehensive analysis of the metabolome of Ae. aegypti, the primary mosquito vector of dengue, Zika, chikungunya and yellow fever viruses, during DENV infection. Alterations in the metabolome are quantitative indicators of the outcome of arbovirus-host interactions that facilitate virus replication or may result from host responses to infection. Our study provides evidence that DENV infection specifically alters the lipid repertoire at the initial site of viral replication, the midgut of the mosquito. Since this study was carried out using a highly susceptible mosquito strain with a laboratory-adapted virus, it highlights metabolic pathways that may be critical for achieving optimal levels of replication in the midgut required for successful dissemination and ultimately transmission. Given that vector competence for infection and transmission is dependent on the specific virus-vector pairing, it will be critical to evaluate the metabolic environment under conditions where the virus has a less robust infection rate and/or the vector presents infection and transmission barriers. This is particularly important in field settings where vector competence is also influenced by environmental conditions, co-infecting pathogens and the microbiome of the vector. In addition, there is a growing body of evidence that metabolic competition between arboviruses and endosymbionts such as Wolbachia may form the basis of endosymbiont-mediated control of viral transmission. Therefore, exploring vector metabolism is a powerful and integrative means to identify metabolic control points that could be exploited to interfere with pathogen transmission or to implement vector control.
DENV serotype-2 strain Jamaica 1409 (JAM-1409) was obtained from the Centers for Disease Controls and Prevention (CDC), Fort Collins, CO, USA [106]. The virus was passaged in C6/36 cells cultured in L15 medium supplemented with 3% fetal bovine serum (FBS), 50 μg/mL penicillin-streptomycin, and 2 mM L-glutamine by infecting the cells at a multiplicity of infection (MOI) of 0.01. Fresh medium was replaced at 7 days post-infection. At 12–14 days post-infection, virus-containing supernatant and infected cells were harvested to prepare the infectious blood meal.
Ae. aegypti strain Chetumal was originally collected from Yucatan Peninsula, Mexico [107]. Adult mosquitoes were fed on raisins and water, with uninfected blood meals to stimulate oogenesis, and were maintained at 28°C, 80% relative humidity with 12–12 h light-dark periods. Male mosquitoes (20–25 males) were placed in one-pint cartons with 200–250 female mosquitoes to maintain the colony.
Mosquitoes were orally exposed to a DENV-infectious blood meal using an artificial membrane feeder as described previously [108] but raisins and water were only removed at 24 and 4 hours prior to blood feeding. For the uninfected control group, we mixed uninfected C6/36 cells suspended in cell culture medium with blood meal to maintain as similar a metabolic input to the DENV-infected group as possible. Mosquitoes were allowed to feed for 45–60 mins. Blood-engorged mosquitoes were reared up to 11 days and fed on sucrose and water.
Two technical replicates of each pool of midguts were analyzed by a LTQ Orbitrap XL instrument (Thermo Scientific, Waltham, MA). It was coupled to an Agilent 1100 series LC (Agilent Technologies, Santa Clara, CA) equipped with a refrigerated well plate auto sampler and binary pumping device.
For polar metabolite analysis, reverse phase liquid chromatography was used to analyze the samples. An Atlantis T3 column (Waters Corp., Milford, MA) with 1.0 x 150 mm, 5.0 μm dimensions was used for the separation. Solvent A consisted of water + 0.1% formic acid. Solvent B consisted of acetonitrile + 0.1% formic acid. The flow rate was 140 μL/minute. A volume of 5 μL was loaded onto the column. The gradient was as follows: time 0 minutes, 0% B; time 1 minutes, 0% B; time 41 minutes, 95% B; time 46 minutes, 95% B; time 50 minutes, 0% B; time 60 minutes 0% B. The LC-MS analysis was run twice, using positive and negative polarity electrospray ionization (ESI). Data were acquired using data dependent scanning mode. FTMS resolution of 60,000 with a mass range of 50–1100 was used for full scan analysis and the FTMS was used for MS/MS data acquisition with a resolution of 7500 and collision induced dissociation (CID) mode.
For non-polar metabolite analysis, reverse phase liquid chromatography was also used to analyze the samples. An Xterra C18 column (Waters Corp., Milford, MA) with 2.1 x 150 mm, 5.0 μm dimensions was used for the separation. Solvent A consisted of water + 10mM ammonium acetate + 0.1% formic acid. Solvent B consisted of acetonitrile/isopropyl alcohol (50/ 50 v/v) + 10mM ammonium acetate + 0.1% formic acid. The flow rate was 300 μL/minute. A volume of 10 μL was loaded onto the column. The gradient was as follows: time 0 minutes, 35% B; time 10 minutes, 80% B; time 20 minutes, 100% B; time 32 minutes, 100% B; time 35 minutes, 35% B; time 40 minutes 35% B. The LC-MS analysis was run twice, using positive and negative polarity in ESI. For MS/MS experiments, data were acquired using a data dependent scanning mode. FTMS resolution of 60,000 with a mass range of 100–1200 was used for full scan analysis and the FTMS was used for MS/MS data acquisition with a resolution of 7500 and higher-energy collisional dissociation (HCD) mode. The top five most intense ions were acquired with a minimum signal of 500, isolation width of 2, normalized collision energy of 35 eV, default charge state of 1, activation Q of 0.250, and an activation time of 30.0. The acquired data were evaluated with Thermo XCalibur software (version 2.1.0).
Raw data were converted to.mzXML in centroid form using msConvert [111]. Downstream analyses were conducted in the open source R program [112]. The xcms package was used with the centWave [112–114] algorithm and a Gaussian fit for peak-picking, and the OBI-Warp method for retention time correction and alignment [115]. Parameters used are given in the supplement material. Features that had retention times outside of limits determined by the gradient of solutions used were not included in further analysis (2–33 minutes for nonpolar modes and 0–47 minutes for polar modes). Intensities for technical replicates were averaged to provide a single value for each biological replicate and then normalized using median ratio scaling described by Wang, et al. [116]. One was added to intensity values prior to transformation to log2 because of the presence of zeroes. Statistical analysis was conducted in the R package limma [117], which fits linear models to the data using an empirical Bayes approach, allowing for information across all features to be used to develop inference about individual features [118]. P-values were based on a moderated t-statistic with a Bayesian adjusted denominator and adjusted for false discovery rate [119]. Each mode (negative / positive + polar / nonpolar) was processed and analyzed separately. Features were designated as significantly different when the absolute log2 fold change was greater than or equal to one, and the adjusted p-value was less than 0.05. Where results from all modes are included in one table or graph, the significance is based on the analysis in single modes alone. Putative molecular names of features were identified by searching mass per charge ratio (m/z) against the Human Metabolome Database (HMDB), LIPID Metabolites and Pathways Strategy (LIPID MAPS) and Metabolite and Tandem MS Database (Metlin) [118]. [M+H]+, [M+Na]+ and [M+NH4]+ adducts were accounted for in the neutral mass calculation in the positive ionization mode nonpolar phase data and [M+H]+, [M+Na]+ for the polar phase data. [M-H]- was the only adduct accounted for in the calculation of negative ionization mode data. The identifiable molecules with mass accuracy of <6 ppm error was further classified into metabolic classes. Features with MS/MS data were further validated by searching against the NIST MS/MS (2014) (Agilent Technologies) and LipidBlast (2012) [120] libraries using MS PepSearch (2013) [121] and manually inspected using NIST MS Search (v2.0) [122].
Long double-stranded RNA (dsRNA) was reversed transcribed and amplified from Ae aegypti mosquito total RNA. Briefly, total RNA was extracted from a whole mosquito homogenized in TRIzol reagent (Ambion). Primers were designed to amplify a region of ~ 500 bp of the gene of interest and a T7 promotor sequence was added to the 5’ end of both forward and reverse primers (S2 Table). Reverse transcription reaction (RT) and polymerase chain reaction (PCR) were performed in two-steps using SuperScript III Reverse Transcriptase (Invitrogen) and Taq polymerase (NEB) respectively. PCR products were purified using the GeneJET PCR Purification kit (Thermo Scientific) and subjected to in vitro transcription using the MEGAscript T7 kit (Invitrogen) according to the manufacturer’s protocol. The reactions were incubated at 37°C for 12 h. The transcribed products were then heated to 75°C for 5 minutes and cooled down at room temperature for 4 h to allow dsRNA to anneal. The products were then treated with DNase to get rid of the DNA template and purified by phenol-chloroform followed by ethanol precipitation. The purified dsRNAs were stored at -80°C for future use.
Aedes aegypti (Aag2) cells were cultured in Schneider’s insect medium (Sigma-Aldrich) supplemented with 2 mM L-glutamine, 1% non-essential amino acids, and 10% FBS. To perform dsRNA knockdown in Aag2 cells, the cells were seeded in a 48-well plate at 50,000 cells/well. Twenty four hours later, the cells were transfected with 260 ng of dsRNA mixed with TransIT-2020 Reagent (Mirus) following the manufacturer’s protocol. New medium with 2% FBS was replaced at 6 hours post transfection. Cell viability assays were performed at 2 days post transfection using the CellTiter-Glo Luminescent Cell Viability Assay (Promega).
Cells were infected with DENV at 48 h post dsRNA transfection. To prepare virus inoculum, DENV-containing cell culture supernatant was mixed with 1x PBS supplemented with 0.5mM MgCl2, 1.2mM CaCl2 and 1% FBS to make a final concentration of virus at a MOI of 0.3. Medium from Aag2 cells was removed and replaced with 0.3 mL of DENV inoculum per well. The cells were allowed to absorb viruses at room temperature for 1 h. Then, the inoculum was removed and replaced with fresh Minimum Essential Medium (MEM; Gibco) supplemented with 1% non-essential amino acids, 2mM L-glutamine and 2% FBS. Quantification of infectious DENV was carried out by plaque assay on BHK cells [6]. Absolute quantification of intracellular DENV genome replication was performed by qRT-PCR as previously described in ‘quantitative RT-PCR detection of infected mosquitoes’ section above. DENV RNA standard was produced from in vitro transcribed DENV genome as previously described in [123].
Cells were collected in 200 μl TRIzol reagent at two days post dsRNA transfection. RNA extraction was performed following the manufacture’s protocol. Total RNA (500 ng) was subjected to RT using random primers (Invitrogen). The cDNA was quantified by quantitative PCR (qPCR) using PowerUp SYBR Green Master Mix (Applied biosystems). Primer sequences designed to quantify DEGS and actin (internal control) genes are shown in S3 Table. ΔΔCt method was used for calculating percent gene expression of the target gene compared to the GFP knockdown control.
Aag2 cells pre-treated with 3.75 μM of N-(4-hydroxyphenyl) retinamide (4HPR; Sigma-Aldrich) dissolved in DMSO or DMSO only as a vehicle control. The final concentration of DMSO was at 1% in Schneider media supplemented with 2% FBS. At 24 h post treatment, cells were infected with DENV at MOI of 0.3. Following absorption, cells were overlaid with MEM containing 2% FBS and fresh 4HPR or DMSO. Virus-containing supernatant was harvested at 24 hpi and titrated by plaque assay. Cell viability assays were performed with 4HPR or DMSO treatment without DENV infection using the CellTiter-Glo Luminescent Cell Viability Assay (Promega).
SPs were extracted from Aag2 cells following the protocol published by [124]. Briefly, cells were trypsinized and washed twice in PBS. Equal cell numbers between treated and control samples were collected. Cer/Sph mixture I Internal standard (Avanti Polar Lipids) was premixed with chloroform at a final concentration of 2 nM. Metabolic extraction was performed as described. Metabolites were dried using nitrogen gas.
An Agilent 1200 Rapid Resolution liquid chromatography (LC) system coupled to an Agilent 6460 series QQQ mass spectrometer (MS) was used to analyze SPs in each sample according to Merrill et al., 2005 with some modifications [124]. A Waters Xbridge C18 2.1mm x 100mm, 3.5 μm column was used for all LC separations (Waters Corp. Milford, MA). The buffers were (A) methanol/water/formic acid (74/25/1 v/v) + 5mM ammonium formate and (B) methanol/formic acid (99/1 v/v) + 5mM ammonium formate for all analyses. All extracted, dried samples were reconstituted in 200 μL of 80/20 buffer A/B just prior to analysis and 10 μL was injected for each analysis. All data were analyzed with Agilent MassHunter Quantitative Analysis (Version B.06.00). The detailed analysis of different SPs is shown in supplemental materials and methods section (S4–S6 Tables).
|
10.1371/journal.pgen.1003206 | A Retrotransposon Insertion in the 5′ Regulatory Domain of Ptf1a Results in Ectopic Gene Expression and Multiple Congenital Defects in Danforth's Short Tail Mouse | Danforth's short tail mutant (Sd) mouse, first described in 1930, is a classic spontaneous mutant exhibiting defects of the axial skeleton, hindgut, and urogenital system. We used meiotic mapping in 1,497 segregants to localize the mutation to a 42.8-kb intergenic segment on chromosome 2. Resequencing of this region identified an 8.5-kb early retrotransposon (ETn) insertion within the highly conserved regulatory sequences upstream of Pancreas Specific Transcription Factor, 1a (Ptf1a). This mutation resulted in up to tenfold increased expression of Ptf1a as compared to wild-type embryos at E9.5 but no detectable changes in the expression levels of other neighboring genes. At E9.5, Sd mutants exhibit ectopic Ptf1a expression in embryonic progenitors of every organ that will manifest a developmental defect: the notochord, the hindgut, and the mesonephric ducts. Moreover, at E 8.5, Sd mutant mice exhibit ectopic Ptf1a expression in the lateral plate mesoderm, tail bud mesenchyme, and in the notochord, preceding the onset of visible defects such as notochord degeneration. The Sd heterozygote phenotype was not ameliorated by Ptf1a haploinsufficiency, further suggesting that the developmental defects result from ectopic expression of Ptf1a. These data identify disruption of the spatio-temporal pattern of Ptf1a expression as the unifying mechanism underlying the multiple congenital defects in Danforth's short tail mouse. This striking example of an enhancer mutation resulting in profound developmental defects suggests that disruption of conserved regulatory elements may also contribute to human malformation syndromes.
| Birth defects are a major cause of childhood morbidity and mortality. We studied the Danforth's short tail mouse, a classic mouse model of birth defects involving the skeleton, gut, and urinary system. We precisely localized the mutation responsible for these birth defects to a 42.8-kb segment on chromosome 2 and identified the mutation as an 8.5-kb transposon that disrupts highly conserved regulatory sequences upstream of the Pancreas Specific Transcription Factor, 1a (Ptf1a). The insertion disrupts a Ptf1a regulatory domain that is highly conserved across evolution and results in spatiotemporal defects in Ptf1a expression: we detected increased expression, temporally premature expression, and (most important for elucidating the mutant phenotype) the ectopic expression of Ptf1a in the notochord, hindgut, and mesonephros—the three sites that will give rise to organ defects in Danforth's short tail mouse. Our data also provide a striking example of how a noncoding, regulatory mutation can produce transient spatio-temporal dsyregulation of gene expression and result in profound developmental defects, highlighting the critical role of noncoding elements for coordinated gene expression in the vertebrate genome. Finally, these data provide novel insight into the role of Ptf1a in embryogenesis and lay the groundwork for elucidation of novel mechanisms underlying birth defects in humans.
| Congenital malformations occur in about 3% of all live births and are a major cause of childhood morbidity and mortality [1], [2]. These defects can affect multiple systems, defining syndromes. Caudal regression syndrome (or caudal dysgenesis) is a major developmental syndrome characterized by malformation of the neural tube, caudal spine, the hindgut and lower limbs [3], [4]. Caudal regression can also occur with bilateral renal agenesis, and this form is usually fatal because it produces secondary pulmonary hypoplasia [3], [4]. Caudal regression most often occurs as sporadic disease; mutations in HLXB9 cause Currarino syndrome (sacral agenesis, OMIM 176450), accounting for a minority of cases [5]. On the milder side of the clinical spectrum, isolated defects, such as renal hypoplasia or unilateral renal agenesis, are common but often remain subclinical [6], [7]. Kidney malformations are highly genetically heterogeneous. Mutations in the PAX2 and HNF1B genes may account for up to 15–20% of pediatric renal hypoplasia [8]–[10]. However, the molecular basis for the majority of severe syndromes, such as bilateral renal agenesis and caudal regression are not well understood.
The Danforth's short tail mutant mouse is a classic Mendelian model of caudal malformations [11]. First described in 1930, this spontaneous mutation (symbol Sd) produces combined defects of the axial skeleton, urogenital system and distal gut [11], [12]. The homozygous mutant mice have fully penetrant abnormalities including truncation of the caudal vertebral column- resulting in a short or absent tail- as well as bilateral renal agenesis, colonic aganglionosis and absence of anorectal opening [12]. The Sd/Sd mice die shortly after birth due to these multiple organ malformations. The heterozygous (Sd/+) mutant mice exhibit short tails with complete penetrance and 30–40% incidence of unilateral renal agenesis; the solitary kidneys in the Sd/+ mice are devoid of major structural defects and consequently, these mice have a normal lifespan [12], [13]. The Sd homozygote and heterozygote mutants thus represent excellent models of human caudal regression syndrome and isolated unilateral renal agenesis, respectively.
The earliest defect detected in Sd/Sd mice is the progressive disintegration of the notochord and the floor plate starting at E9.5, causing patterning defects in both neural tube and somites, leading to vertebral defects; the ureteric bud, derived from the mesonephric duct, either fails to reach or fails to induce the metanephric blastema, resulting in renal hypoplasia or aplasia; abnormalities in the development of the hindgut and cloaca result in the absence of an anorectal opening and could also be the cause of aganglionosis of the rectal pouch [12]–[16]. Studies of chimeric embryos have shown that the Sd cells are selectively lost from the notochord and ventral hindgut endoderm starting at E9.5, implicating a cell-autonomous defect in these tissues [17]; in rare instances that metanephric kidneys develop, however, Sd cells are robustly incorporated into chimeric kidneys, suggesting that the urogenital defect may be cell-nonautonomous or due to specific impairment of signaling for the mesonephric to metanephric transition [17]. More recent studies have shown that genetic ablation of the notochord with Diphtheria toxin recapitulates the axial defects observed in the Sd mutants, but notochord-ablated mice exhibited only kidney fusion and no noticeable defects in nephrogenesis, suggesting that an additional mechanism accounts for the renal agenesis in the Sd/Sd mouse [18]. Thus, identification of the genetic basis of the Sd mutation will provide insight into mechanisms of axial skeletal development and reconcile potentially contradictory findings about the origin of visceral defects in this mutant strain.
Previous studies have assigned the Sd locus to Chr. 2A3 but the mutation was not known [19]. Here we refined the Sd locus to a 42.8 kb interval and identified the Sd mutation as an insertion in the 5′ regulatory domain of Pancreas Specific Transcription Factor 1a gene (Ptf1a). This mutation results in ectopic expression of Ptf1a in the notochord, mesonephros and gut providing an explanation for the complete spectrum of abnormalities seen in Danforth's short tail mutants.
The Sd mutation arose in 1930 on an outbred stock prior to generation of classical inbred strains. Mice carrying the Sd mutations were transferred in 1950 to Jackson laboratories in a linkage testing stock called E1, which was sequentially outcrossed to C57BL/6J, C3H/He and CBA and then maintained as closed colony until 1970 (Jackson laboratories). This colony, segregating the Sd mutation, was subsequently sibling mated at Jackson laboratories and named RSV/LeJ. We obtained RSV/LeJ-Sd/+ mice and confirmed that Sd/Sd, Sd/+ and +/+ mice are obtained in an expected Mendelian ratio. RSV/LeJ-Sd/Sd mice were readily recognizable at birth because they manifested all the organ malformations defects originally described [12]–[14]. RSV/LeJ-Sd/+ exhibited short tails with no other outwardly visible defects; on dissection, unilateral renal agenesis was detected in 10/28 (35%) of RSV/LeJ-Sd/+ mice, consistent with previous reports [12]–[14]. RSV/LeJ- +/+ exhibited long tails and were anatomically indistinguishable from other inbred strains.
Previous studies had mapped the Sd locus to a 5 cM interval on Chr. 2 [19]. To confirm and refine this interval, we generated 174 backcross (BC) and 120 F2 intercross progeny (using C57BL/6J as the counterstrain). Analysis of linkage with 10 informative markers confirmed the mapping of the Sd locus to this region, yielding a peak lod score of 39 (p<6×10−41, Figure 1 and Figure S1). Three informative recombinants refined the locus to 1.3 Mb interval containing nine transcriptional units, including five genes with orthologs in other mammalian species (Pip4k2a, Armc3, Mrsb2, Ptf1a, Otud1, Figure 1A and 1B).
However, sequence analysis of exons and flanking regions as well as copy number analysis of all nine positional candidates did not identify any coding variants or intragenic copy number changes that distinguished RSV/LeJ-Sd/Sd mice from the background strain. Because the mapping was based on multiple critical recombinants in affected mice, this excluded the possibility of mismapping due to incomplete penetrance. Taken together these data indicated that the Sd mutation occurs in a noncoding region within this interval.
We therefore undertook further refinement of the Sd locus by meiotic mapping in additional F2 intercross mice. We did not identify informative markers between B6 and RSV/LeJ-Sd/Sd, or RSV/LeJ- +/+ across this minimal interval, suggesting that these strains share the same ancestral haplotype in this region. However, DBA/2J harbored informative SNPs and was therefore used as the counterstrain for fine mapping. We generated 1203 F2 intercross mice and genotyped 29 informative markers across the minimal 1.3 Mb interval. We identified 14 F2 progeny with recombinants within this interval, with 4, 7 and 3 mice exhibiting the wildtype, heterozygote and homozygote phenotypes, respectively. Among these, three critical recombinants localized the Sd locus to a 42.8 Kb intergenic region between rs13476366 and rs27124414, proximal to Ptf1a (Figure 1C, 1D). These mapping results were highly reliable because two of these critical recombinants occurred in affected mice exhibiting the homozygous and heterozygous mutant phenotype (Figure 1C).
We performed Sanger sequencing of the 42.8 Kb segment spanning the minimal meiotic interval in an RSV/LeJ-Sd/Sd and an RSV/LeJ-+/+ mouse, achieving 100% coverage of the interval with an average base call accuracy of 99%. We identified an 8.53 Kb insertion within this region in Sd homozygotes (located at nucleotide position 19,355,026 on Chr. 2, genome build 37.2). The insertion was present in homozygosity in RSV/LeJ-Sd/Sd mice, in heterozygosity in RSV/LeJ-Sd/+ mice, and was absent in the background RSV/LeJ-+/+ mice and in the two counterstrains used for the mapping study (DBA/2J, C57BL/6J) (Figure 1E). We also genotyped this insertion in a random sample of 142 mice from the F2 mapping progeny and found that it perfectly segregated with the heterozygote and homozygote Sd phenotypes. There were no other variants distinguishing RSV/LeJ-Sd/Sd from RSV/LeJ-+/+ within the 42.8 kb minimal recombinant intervals, and the RSV/LeJ-Sd/Sd sequence otherwise shows identity with the C57BL/6J reference sequence (Figure 1F).
Sequence comparison with Repbase data indicated that the insertion is an endogenous retroviral element, with the closest alignment with the early retrotransposon (ETn) subtype (Figure 2A and 2B, Table S1). The insertion results in duplication of 6 bp at its flanking sites, without loss of reference sequence, which is typical of ETn sequences (Figure 2C) [20], [21]. Many copies of this mobile element are interspersed across the mouse genome, and transposition of ETn sequences into genic regions is responsible for many spontaneous mutant mouse phenotypes [20]–[23]. We further determined the haplotype on which this mutation occurred by genotyping 37 SNPs in the 95.5 kb interval surrounding Ptf1a (2.6 kb spacing, Table 1). This analysis demonstrated that in the RSV/LeJ-Sd/Sd and RSV/LeJ-+/+ strains shared the same ancestral haplotype at the Ptf1a locus and therefore RSV/LeJ-+/+ provides a valid reference haplotype for RSV/LeJ-Sd/Sd. This same haplotype is also present in ten other inbred strains (C57BL/6J, CBA/J, BALB/cJ, C3H/HeJ, C57BL/6NJ, LP/J, 129P2/OlaHsd, 129S1/SvImJ, 129S5SvEv, and NZO/HlLtJ, Figure 3). These data are consistent with the lack of polymorphism between C57BL/6J, RSV/LeJ-Sd/Sd and RSV/LeJ-+/+ based on sequencing of the 42.8 kb minimal recombinant interval. Genotyping and Sanger Sequencing verified that the ETn insertion was absent in these ten strains as well as fifteen other strains with different haplotypes at the Ptf1a locus (Figure 1F). These data demonstrate that the ETn insertion is not an old polymorphism but is a new variant that arose on ancestral haplotype still commonly represented among laboratory inbred strains. Altogether, the precise mapping of the Sd locus to a 42.8 kb interval, the detection of an ETn insertion in the Sd mice and its absence in the RVS/LeJ background strain and twenty-five additional inbred strains, including strains with the same ancestral haplotype, demonstrated that we successfully identified the mutation responsible for the Danforth mouse phenotype.
The ETn is inserted 12.2 kb upstream of the Pancreas Specific Transcription Factor 1a (Ptf1a) start site (Figure 2A). Ptf1a encodes a subunit of a trimeric Pancreas Specific Transcription Factor Complex (PTF1) which regulates cerebellar, retinal, pancreatic and spinal cord development [24]. The 5′ regulatory region of Ptf1a contains highly conserved elements, including a 2.3 kb autoregulatory enhancer domain located 13.4 kb upstream of the Ptf1a start site (Figure 2B), which normally directs Ptf1a expression to the dorsal spinal cord [25], [26]. Deletion of this 2.3 kb enhancer results in mislocalization of reporter constructs to the ventral spinal cord [26]. The Sd insertion occurs downstream of this highly conserved tissue enhancer, and displaces this element 8.5 kb upstream from its conserved position in the Ptf1a regulatory region (Figure 2A and 2B). The next closest gene Mrsb2, is located 64 kb proximal to the insertion, and has very low sequence conservation around its 3′ untranslated region. The data suggested that the ETn insertion is most likely to affect Ptf1a expression.
Analysis with GENSCAN predicted a few low probability open reading frames within the transposon sequence, but quantitative PCR analysis of whole embryo found no evidence that these sequences are transcribed, consistent with the known lack of transcriptional activity of ETn elements (data not shown) [20]. We examined expression of Ptf1a as well as other genes that were located 500,000 bp upstream or downstream of the Sd mutation. Quantitative PCR expression analysis of whole embryos at E9.5 and E10.5 revealed a consistent four-to ten fold increased expression of Ptf1a in Sd mutants (p<9×10−14 for comparison of Sd/Sd vs. +/+ mice, Figure 2D). There were no detectable differences in expression of neighboring genes with mammalian orthologs at these two time points (Pip4k2a, Armc3, OtuD1 and Mrsb2, Figure 2E). The other predicted genes in the region did not have mammalian orthologs and did not have detectable expression at these time points. These data rule out an effect on other genes located 500 kb upstream or downstream of Ptf1a.
To follow-up these findings, we performed in situ hybridization for Ptf1a in developing embryos. Consistent with a previous study [27], Ptf1a was robustly expressed throughout most of the length of the dorsal neural tube by E10.5 in wild type and was also seen in the same location in Sd/+ and Sd/Sd embryos (Figure 4A–4F). In addition to this normal expression pattern, Sd/+ and Sd/Sd embryos showed ectopic expression of Ptf1a in the hindgut, around the cloaca and in the hindgut diverticulum extending into the tail (Figure 4H, 4I, 4K, 4L). Consistent with E10.5 qPCR data, they're appeared to be a higher level of expression in the Sd/Sd embryos compared to Sd/+ (Figure 4B, 4C, 4E, 4F), and also a higher level of endogenous expression in Sd/+ compared to wild type (Figure 4A, 4B, 4D, 4E).
A day earlier at E9.5, extensive ectopic Ptf1a expression was evident in Sd/+ and Sd/Sd embryos in the tailbud mesenchyme (Figure 5B, 5C, 5E, 5F, 5N, 5O), the notochord and hindgut (Figure 5K, 5L, 5N, 5O), and throughout the length of the developing mesonephros and mesonephric duct (Figure 5B, 5C, 5K, 5L). Ectopic expression was evident even at E8.5 in the lateral plate mesoderm and tail bud mesenchyme of Sd/+ and Sd/Sd embryos (Figure 6A, 6B, 6C, 6E, 6F) and the notochord of Sd/Sd embryos (Figure 6C, 6F). In summary, we detected ectopic Ptf1a expression in every organ that will manifest a developmental defect in Sd/+ and Sd/Sd mice - the notochord, the mesonephros and the hindgut.
These data suggested that dysregulated timing and localization of Ptf1a expression is responsible for the Danforth phenotype. Nonetheless, because increased Ptf1a expression was also detected in mutant embryos, we attempted to distinguish hypermorphic from neomorphic effects of the Sd allele by crossbreeding Sd/+ mice to Ptf1a haploinsufficient mice (Ptf1a-cre mice, reference [28]). If the Danforth phenotype was solely due to increased expression of Ptf1a, then reduction of Ptf1a gene dosage should attenuate the organ malformations. While the Sd mutants had significantly reduced kidney and tail lengths, there were no phenotypic differences between Sd/+, and Sd/Ptf1a-cre mice at weaning (Table 1). Histologic analysis of the kidneys also did not reveal morphological defects, consistent with prior reports of normal histology in Sd heterozygotes [12]–[16]. The absence of rescue by Ptf1a haploinsufficiency suggests that inactivation of one wild-type allele could not compensate for the increased expression from the Sd allele or more likely, that the neomorphic effects of the Sd allele predominate in the pathogenesis of malformations in the Danforth mice.
Only a few PTF1a targets are known. Recent data suggest that PTF1a regulates Mnx1, Nkx6-1, Bmp7, Dll1 and Onecut1 expression [29]. However, we did not detect any differences in expression of these genes between Sd/+, Sd/Sd and +/+ mice at E 9.5 (Figure S2), suggesting that increased Ptf1a expression is not sufficient to activate these particular targets in this tissue context and this time point.
Danforth first described his spontaneous mutant strain over 80 years ago [11]. Since that time, this strain has been studied as a classical model of developmental defects of the spinal cord, hindgut and the urogenital tract [12]–[19]. Although the Sd locus was assigned to Chr. 2A3 in 1980, the underlying mutation had not been identified. We used meiotic mapping in 1203 F2 mice to precisely map the mutation to a 42.8 kb intergenic region. This segment contained an ETn insertion upstream of Ptf1a in the mutant strain, which was absent in the RSV/LeJ background strain and twenty-five wild-derived or classical inbred strains. SNP analysis of the Ptf1a locus suggests that the mutation arose on an ancestral haplotype that is shared by at least eleven laboratory strains, including RSV/LeJ and three strains used in the derivation of this laboratory strain (C57BL/6J, CBA/J and C3H/HeJ). Our findings are independently supported by accompanying papers (by Vlangos et al. and Semba et al.) who used a alternatives approaches to identify the same mutation. Taken together, these data establish that we have identified the genetic basis of the Sd mutation. The presence of a noncoding mutation explains the difficulties in identifying the genetic lesion in the Danforth strain since its initial description in 1930. We observed perfect co-segregation of all of the skeletal, urogenital and gastroenterological phenotypes in all affected F2 and BC mice generated in this study, indicating that the noncoding mutation is at the origin of the entire spectrum of defects. The ETn insertion occurs within the regulatory sequences upstream of Ptf1a, and is associated with dysregulated dosage, timing and localization of expression. These data implicate dysregulated Ptf1a expression as the cause of developmental defects in the Danforth mouse and indicate that Sd should be recognized as a neomorphic Ptf1a allele.
Ptf1a is a member of the trimeric transcriptional complex PTF1, and loss of function mutations result in recessive cerebellar and pancreatic hypoplasia in humans and mice [24]–[26], [30]–[34]. Ptf1a has highly restricted temporo-spatial expression during embryogenesis [24]–[26], [30]–[33]. Transient Ptf1a expression between E10 to E13 initiates the genetic programs required for specification of dorsal horn neurons, the cerebellum and the pancreas; Ptf1a expression subsequently declines to undetectable levels, with postnatal expression present only in pancreatic acinar cells [24], [31], [32], [35]–[42]. This tightly regulated temporal and spatial expression pattern is controlled by highly conserved regulatory elements upstream and downstream of Ptf1a but the consequences of regulatory mutations have not been reported. The Sd mutation occurs in the vicinity of a highly conserved enhancer sequence (positioned from 13.4 to 15.6 kb upstream of the Ptf1a start site), which contains multiple autoregulatory domains required for restricting expression to the dorsal spinal cord and maintaining expression in the adult acinar cells (Figure 2B) [25], [26]. Reporter constructs lacking this enhancer are ectopically expressed in the ventral spinal cord [26] and although not specifically reported, are also evident in the hindgut at E10.5 (see reference [26], Figure 1B), consistent with the mislocalized expression pattern in Sd/+ and Sd/Sd mutants (Figure 3, Figure 4). The retrotransposon insertion may disrupt this enhancer or a neighboring negative cis-acting element, or may act as a broadly-acting positive regulator, resulting in ectopic expression of Ptf1a in Sd mutants.
We show that each tissue that has been implicated as being primarily affected in Sd mutant mice – the notochord, the mesonephric duct and the hindgut – shows ectopic Ptf1a expression at a critical stage in its early development, indicating that dysregulated Ptf1a expression is at the origin of the developmental defects arising from these compartments. Notably, ectopic expression of Ptf1a is present in the notochord at E8.5, prior to or coincident with the earliest notochord defects and prior to the start of notochord disintegration [12]–[16]. Our data also suggest that the mouse embryo is very sensitive to Ptf1a gene dosage as Sd/Sd mice, which have a higher level of ectopic expression, consistently manifest more severe axial defects and near complete penetrance of bilateral renal agenesis resulting in death, whereas Sd/+ animals have less severe defects compatible with survival.
PTF1a requires two cofactors to form an active transcriptional complex [34], [37]. If its canonical partners were present, its misexpression may activate its direct downstream targets, ectopically initiating neural or pancreatic developmental programs [43]. Alternatively, PTF1a may form transcriptional complexes inappropriately in regions where it is ectopically expressed, interfering with normal developmental processes and impairing notochord, urogenital and hindgut development. At present, the downstream targets of Ptf1a have not been comprehensively identified. We examined five known Ptf1a targets but did not detect increased expression in Sd mutant mice, indicating that unbiased genome-wide approaches will be required to discover dysregulated developmental programs downstream of ectopic Ptf1a expression.
The present findings provide a unifying mechanism for the multiple developmental defects in the Danforth mouse and reconcile several prior observations. For example, our data confirm a previous study that used genetic interaction to infer that Sd is a gain-in-function mutation [44]. The present data also explain why the Danforth mutation was not fully recapitulated by ablation of the notochord [18] and could not be explained based on dysregulation of any single embryonic compartment [17], as ectopic Ptf1a expression likely misdirects developmental programs independently in each affected tissue. Now that we identified the initiating genetic lesion, future studies can determine the precise molecular cascade leading to the developmental defects in each compartment.
To date, the vast majority of developmental defects reported in mice and humans are produced by coding mutations that result in loss of function of the encoded protein; developmental defects arising from disruption of conserved regulatory elements have been less frequently described [45]. Comparative genomics studies of bony vertebrates have identified highly conserved sequences that are enriched around genes that have tissue-specific enhancer activity, acting as developmental regulators [45]–[47]. The Sd mutation provides a striking example of a cis-regulatory mutation that produces profound developmental defects that are quite distinct from phenotypes resulting from simple loss/gain of function mutations. The Danforth mouse can thus serve as an excellent model for dissecting the role of enhancer elements on the temporo-spatial regulation of gene expression in vertebrate development. These data also suggest that mutations in conserved regulatory elements may contribute to human malformation syndromes. For example, we recently studied 522 children with kidney malformations and identified 72 rare copy number disorders that disrupt coding segments, accounting for up to 16.6% of cases [48]. However, we also identified many rare or unique intergenic CNVs in this population, suggesting that disruption of noncoding elements may also play a pathogenic role in this phenotype [48]. Given the high sequence conservation in the Ptf1a 5′ regulatory region, this segment is a good candidate for mutational screening in larger patient populations with caudal regression, axial or urogenital defects.
The Association for Assessment and Accreditation of Laboratory Animal Care guidelines were followed for all animal procedures, and all procedures were approved by the Institutional Animal Care and Use Committee of Columbia University. All inbred strains including the strain carrying the Sd mutant allele (RSV/LeJ-Sd/+) mice were purchased from The Jackson Laboratory. To refine the Sd locus, we generated 174 backcross and 120 intercross mice between RSV/LeJ-Sd/+ mice and C57BL/6J mice and also a second mapping F2 intercross cohort of 1203 mice by intercrossing (RSV/LeJ-Sd/+ x DBA/2J) F1 mice. Mice were phenotyped at birth by visual inspection. Affection status was assigned based on presence of the short tail in homozygotes and heterozygotes. While heterozygotes have no other outwardly visible phenotypes, the homozygote mutants are readily recognizable based on the presence of caudal agenesis, and on dissection have major gut and urogenital malformations as previously described [11], [13], [14], [16]–[19]. The Ptf1atm1(cre)Wri or Ptf1a–cre mice (carrying a Cre recombinase replacing the Ptf1a protein-coding region, resulting in a Ptf1a null allele) were obtained from Wright and colleagues [28]. To examine the effect of Ptf1a haploinsufficiency on the heterozygote phenotypes, we generated F1 hybrids between Ptf1a–cre/+mice and RSV/LeJ-Sd/+ mice. Phenotypes were determined at weaning (postnatal day 22), by measurement of tail length, kidney length and kidney weight. H&E staining of kidneys were also performed. Genomic DNA was isolated using the Genomic DNA Isolation kit (Lamda Biotech).
Marker loci for inbred strains were obtained from the Mouse Phenome Database (http://phenome.jax.org/); we filtered these for SNPs that were polymorphic between C57Bl/6J, DBA/2J and RSV/LeJ. The BC group was genotyped using microsatellite markers across the Sd locus. Multipoint lod score was calculated using the R/QTL package, utilizing the discrete trait analytic model. The F2 cohort was genotyped with 29 informative SNPs distributed across the 1.3 cM region delimited by rs27129240 and rs29504224 (Sequenom Mass Array system, Columbia University Genotyping facility). To define the haplotype at the Ptf1a locus, we searched SNPs in the Mouse Phenome Database (MPD) and genotyped 37 SNPs that differentiated common haplotypes between 17 inbred strains sequenced by the Sanger Center [49]. Haplotype analysis was performed by Sanger sequencing of RSV/LeJ-+/+, RSV/LeJ-Sd/Sd, C57BL/6J and DBA/2J The genotypes in C57BL/6J and DBA/2J mice demonstrated a 100% concordance with genotypes from dbSNP, confirming accuracy of Sanger sequencing.
Gene annotation was performed using NCBI database (http://www.ncbi.nlm.nih.gov/) and UC Santa Cruz genome browser (http://genome.ucsc.edu/). Mutational screening of the positional candidate genes was performed by Sanger sequencing of exons and flanking introns, comparing RSV/LeJ-Sd/Sd mouse to a RSV/LeJ- +/+ mouse. Long range-PCR was performed in order to amplify the insertion in the Sd/Sd mutant (TaKaRa long range PCR). We performed Sanger sequencing of the 42.8 kb minimal recombinant interval in an RSV/LeJ -Sd/Sd and RSV/LeJ −+/+ mouse. We achieved 100% coverage of this interval with an average base call accuracy of 99% (Phred scores of 20); there were no gaps or ambiguity in the interval. Bidirectional sequencing was obtained for 76% of the region. The sequence of the insertion was analyzed by alignment with the Repbase data, the most commonly used database of repetitive DNA elements (http://www.girinst.org/repbase/index.html ref [50]). From The Jackson Laboratory, we obtained genomic DNA from 25 strains of mice which comprehensively represent all Mus musculus sub-species, including 4 wild-derived (PWD/PhJ, MOLF/EiJ, WSB/EiJ, CAST/EiJ) and 21 inbred strains (DBA/2J, DBA/1J, C57BL/6J, C57BL/6NJ, BALB/cJ, FVB/NJ, RIIIS/J, C3H/HeJ, AKR/J, NOD/LtJ, SJL/J, 129/SvimJ, 129/SvEv, 129P2/OlaHsd, CBA/J, CFW, SWR/J, BTBR T<+>tfJ, NZW/LacJ, KK/HlJ, and A/J).
Timed embryos were collected at embryonic day (E) 9.5, 10.5 and E12.5 (where E0.5 is the day of detection of the vaginal plug). Comparisons were made between littermates with differing genotypes. RNA was isolated from whole embryos with TRIzol (Invitrogen) followed by DNaseI treatment and clean-up using the RNeasy mini kit (Qiagen). cDNA was generated with the Omni-Script kit (Qiagen). cDNA levels were quantified in duplicate by qPCR using SYBR Green Mix on an IQ thermal cycler (Bio-Rad). The same internal control was included in each run to standardize each qPCR run, and β-actin was used as reference gene (using Pfaffl algorithm). Expression levels were standardized to the same E10.5 wild-type mouse embryo.
Embryos were collected at E8.5, E9.5 and E10.5 and yolk sacs were removed for PCR genotyping. Embryos were fixed in 4% paraformaldehyde (PFA) (Fisher), washed in PBT, dehydrated in methanol and processed for whole mount in situ hybridization (ISH) as described previously [51]. Samples at each developmental stage were processed together and images captured under identical settings to allow qualitative comparison of staining intensity between genotypes. The Ptf1a probe was generated by cloning the full-length Ptf1a cDNA from mouse embryonic pancreas into the TOPO blunt PCR cloning vector (Invitrogen). Briefly, embryos were bleached in hydrogen peroxide for an hour, treated with proteinase K (10 µg/ml) (Roche) and fixed in PFA. Further the embryos were incubated at 65°C in hybridization buffer for an hour and then digoxygenin labeled Ptf1a antisense RNA probe was added to hybridization buffer overnight. The following day the embryos were washed with solutions of decreasing stringency of saline sodium citrate salt solutions, washed in Tris-buffered saline with tween (TBST) and incubated in alkaline-phosphatase labeled anti-digoxygenin (Roche) in TBST overnight at 4°C. On the third day the embryos were washed in TBST followed by an overnight wash. On day four the embryos were developed using the BM purple (Roche) solution in dark. The embryos were washed in PBT after developing, post fixed with PFA and stored in PBT until photographed on a Nikon SMZ1500 dissecting microscope.
|
10.1371/journal.pgen.1002037 | PTG Depletion Removes Lafora Bodies and Rescues the Fatal Epilepsy of Lafora Disease | Lafora disease is the most common teenage-onset neurodegenerative disease, the main teenage-onset form of progressive myoclonus epilepsy (PME), and one of the severest epilepsies. Pathologically, a starch-like compound, polyglucosan, accumulates in neuronal cell bodies and overtakes neuronal small processes, mainly dendrites. Polyglucosan formation is catalyzed by glycogen synthase, which is activated through dephosphorylation by glycogen-associated protein phosphatase-1 (PP1). Here we remove PTG, one of the proteins that target PP1 to glycogen, from mice with Lafora disease. This results in near-complete disappearance of polyglucosans and in resolution of neurodegeneration and myoclonic epilepsy. This work discloses an entryway to treating this fatal epilepsy and potentially other glycogen storage diseases.
| Lafora disease (LD) is a fatal epilepsy that afflicts previously normal teenagers. It is caused by mutations in the EPM2A or EPM2B genes encoding the laforin carbohydrate-binding phosphatase and the malin E3 ubiquitin ligase. LD is the most common neurodegenerative epilepsy of adolescents. Affected children suffer an ordeal lasting 10 years, consisting of escalating seizures, constant body jerking, particularly frightening epileptic visual hallucinations, and later on dementia. They die of massive convulsion. Brain biopsies reveal accumulation of a starch-like compound, polyglucosan, overtaking dendrites and likely causing the disease, and neurodegeneration. Glycogen synthase (GS), the enzyme that forms normal glycogen, is also responsible for synthesizing these polyglucosans. We reasoned that reducing GS activity might prevent polyglucosan formation. Mice deficient of Epm2a replicate LD and are a standard model. Members of our group generated mice deficient of PTG, a protein involved in activating GS. By breeding LD mice with PTG-lacking mice, we generated LD mice lacking the GS-activating effect of PTG. This resulted in a cure. The double knockout mice have almost no polyglucosan, no neurodegeneration, and no seizures. Our work opens an avenue of treatment for this fatal epilepsy, which may also be applicable to other glycogen storage diseases.
| Lafora disease (LD) is caused by recessively inherited mutations in the EPM2A or EPM2B genes, encoding laforin (a carbohydrate binding phosphatase) and malin (an E3 ubiquitin ligase) [1], [2]. The disease begins around age 15 with myoclonus (jerk-like seizures) and generalized convulsive seizures, which initially respond to medications. Over the next five years seizures become intractable and the myoclonus near-constant, and epileptic hallucinations with highly frightening content appear. Extremely frequent myoclonic seizures (repetitive jerks) and epileptic absence attacks permeate consciousness and prevent formulation of complete thoughts. Dementia and a vegetative state in constant myoclonus follow. Death occurs around age 25 in status epilepticus. Pathology consists of the progressive formation of polyglucosans, which are insoluble glucose polysaccharides that precipitate and aggregate into concretized masses called Lafora bodies (LB), and in neurodegeneration. LB form in neuronal perikarya (i.e. in the cell body near the nucleus) and in neuronal short processes (mostly dendrites). LB in the neuronal processes are much smaller but they massively outnumber LB in the perikarya. Extraneurally, LB also form in heart, liver, and skeletal muscle, but cause no symptoms in these organs [3]–[6].
A normal glycogen molecule contains up to 55,000 glucose units, yet remains soluble because its glucose chains are short (13 units), each chain is a branch of another, and the whole molecule is a sphere, the surface of which is composed of the hydrophilic ends of chains [7]. This unique organization allows mammalian cells to store large amounts of carbohydrate energy in a soluble rapidly accessible form. Without branching, glucose polymers 13 units or longer are poorly soluble and tend to precipitate and crystallize [8]. Polyglucosans are malformed glycogen molecules. They have very long chains, insufficient branches, and a resultant lack of spherical organization. They are more similar to plant amylopectin or starch than to glycogen, and like these plant carbohydrates they are insoluble, precipitate, and accumulate [3], [5], [9].
Glycogen is synthesized through coordinated actions of glycogen synthase (GS) and glycogen branching enzyme, the former responsible for chain elongation, the latter for chain branching. Glycogen is digested by glycogen phosphorylase (GP) and glycogen debranching enzyme. PTG (protein targeting to glycogen) is an indirect activator of GS and an indirect inhibitor of both GP and glycogen phosphorylase kinase (GPK), the enzyme that activates GP. PTG performs this reciprocal activation of synthesis and inhibition of breakdown by binding the pleiotropic phosphatase PP1 through its C-terminus, binding glycogen, and through a common region in its N-terminus (amino acid sequence WDNNE) binding GS, GP, or GPK, thus targeting PP1 to each of the three enzymes. PP1 dephosphorylates each of the three enzymes, activating GS and inhibiting GP and GPK [10], [11].
There are two main hypotheses of polyglucosan formation, the first based on evidence from cell models that laforin interacts with malin and with PTG, and that the laforin-malin complex downregulates GS through malin-mediated ubiquitination and degradation of PTG. In this hypothesis, absence of laforin or malin would increase PTG, which would over-activate GS, leading to excessive extension of glycogen chains and conversion of glycogen to polyglucosan [12]–[14]. Although results from animal models have yet to confirm this idea [15]–[17], there is indeed a body of work implicating PTG.
The second hypothesis is based on the observation that laforin dephosphorylates glycogen and that in LD there is progressive hyperphosphorylation of glycogen, causing it to unfold and precipitate. GS remains bound to the precipitating glycogen, but glycogen branching enzyme, the enzyme responsible for branching, even under normal condition does not associate tightly [16]–[19]. In this hypothesis, elongation by GS of the chains of the precipitated glycogen, with no branching, would convert glycogen to polyglucosan. Both hypotheses predict that inhibiting GS would prevent polyglucosan formation, and if LB are causative of the PME, this might ameliorate or cure the epilepsy. One way to inhibit GS would be to interfere with its activation by PTG. In the present work we genetically remove PTG from mice with LD. We obtain dramatic reduction in LB, and resolution of neurodegeneration and the PME. This work has direct implications for therapeutic intervention in this fatal disease.
We initially considered removing the muscle/brain isoform of GS (GYS1) from LD mice by breeding GYS1-deficient mice with laforin-deficient mice. However, this is impractical because in 90% of cases GYS1-deficient mice cannot survive birth (although the 10% that do are subsequently healthy with normal lifespan and exercise tolerance) [20], [21]. Recently, DePaoli-Roach generated a mouse line deficient of PTG. In contrast to an earlier report that disruption of the PTG gene was embryonic lethal [22], the present mice are healthy and have normal lifespan [23]. Their glycogen is reduced by 30% in skeletal muscle and by 70% in brain.
Laforin-deficient mice (LKO) have been extensively characterized and exhibit LB formation, neurodegeneration, and PME [24]. The PME is not as severe as in humans. The mice develop progressively worsening myoclonus, but convulsive seizures are not seen [24]. Unlike human patients and despite the neurodegenerative changes and progressive myoclonus LKO mice do not have a shortened lifespan (unpublished observation). Metabolically, LKO mice have progressively increasing accumulation of glycogen in tissues, reaching approximately fivefold normal in brain and threefold in skeletal muscle by age nine to 12 months [16]. To remove PTG from the laforin-deficient mice, we bred LKO mice with PTG knockout mice and interbred their litters to produce PTG/laforin double knockout (DKO) animals. DKO mice are born at Mendelian frequency, have normal skin, body habitus and growth, exhibit no obvious behavioral abnormalities, and appear to have normal lifespan, our oldest presently healthy at 18 months of age.
As mentioned, nine to 12 month-old LKO mice have vast amounts of LB in brain and other organs, and neurodegeneration [24]. We studied brain and skeletal muscle from LKO and DKO mice and their wild-type (wt) littermates at 12 months and found massive reduction in LB in DKO mice (Figure 1 and Figure 2). In hippocampus, frontal cortex and cerebellum, the numbers of LB in neuronal processes in DKO were respectively 3%, 0.1%, and 0.5% of those in LKO animals. The numbers of perikaryal LB were diminished to 10% in hippocampus and 5% in frontal cortex. In cerebellum, perikaryal LB were not significantly reduced in number, although they were much smaller in size. In skeletal muscle, LB had completely disappeared, compared to their very large quantities in LKO animals (Figure 3). Wt animals, as expected, had no LB in either tissue. To determine whether the reductions in LB correlated with reductions in glycogen content, we measured total glycogen in whole brain and skeletal muscle and found that the increased glycogen content of LKO mice had normalized to wt levels in DKO animals (Figure 4).
Lost neurons are replaced by astrocytes. We assessed neuronal loss in DKO, LKO and wt animals at 12 months first by measuring gliosis, which we quantified by morphometric counts of glial fibrillary acidic protein (GFAP)-positive cells. In cerebellum, there were no differences between the genotypes. In hippocampus and frontal cortex, however, DKO mice had half the number of astrocytes as LKO animals, and the same number as wt, i.e., they have no measurable gliosis (Figure 5). We next assessed neurodegeneration directly. In their original study of neuropathology in LKO mice, Ganesh and colleagues noted absence of apoptosis and necrosis. Using electron microscopy (EM), they documented an unusual form of somatic degeneration characterized chiefly by shrinkage and retraction of plasma and nuclear membranes and darkening of the cytoplasm [24]. We performed EM studies in the present set of LKO, DKO and wt mice. Figure 6A–6C show representative wt cerebellar Purkinje neurons with characteristic full nuclei and cytoplasms and taut and circular plasma membranes. Numerous axon terminals are seen directly apposed to the membranes forming normal synapses lined one next to the other around the circumferences of the cells. Figure 6D–6F show typical LKO Purkinje cells. Nucleus and cytoplasm are shrunken. The plasma membrane is wrinkled and retracted with appearance of indistinct spaces between it and the axon boutons that would normally associate with it, effectively resulting in loss of synaptic contacts. Numerous LB in neuronal processes are present. Figure 6G–6I show representative DKO Purkinje cells. The cells are essentially back to normal with full nuclei and cytoplasms, circular plasma membranes, and generally a full complement of synapses around the cell body. However, the correction while near-perfect is not completely perfect. The plasma membrane is not quite as taut as in wt, and there are rare instances of synaptic contact loss.
Myoclonus is a single jerk of the body or of a body part. Mice, like humans, exhibit a certain amount of physiologic myoclonus, such as hypnagogic myoclonus [25], [26]. In LD patients, myoclonus is extremely frequent and in later stages near-constant and debilitating [4], [5], [25]. We counted myoclonus in 12 month-old wt, LKO, and DKO animals, blind to genotype. Myoclonus was defined as sudden rapid jerks of the head or of the dorsum of the animal. In the latter, the split-second myoclonus causes retropulsion of the animal, closely resembling the myoclonus we documented previously in canine LD [27]. LKO mice have fourfold increased myoclonus over wt. DKO were the same as wt (Figure 7). In their original description Ganesh and colleagues reported that in addition to myoclonus 80% of nine to 12 month-old LKO animals also exhibit myoclonic seizures (polymyoclonus), consisting of rapid repetitive head and body jerks lasting few seconds and associated with epileptic discharges on electrocorticography [24]. We observed polymyoclonus in 80% of the present 12 month-old LKO mice, in no wt mice, and in no DKO mice.
In this study we show for the first time that removal of PTG in an animal model of LD reduces LB formation, and eliminates neuronal loss and the myoclonic epilepsy. PTG is not the only protein that targets PP1 to glycogen and glycogen metabolizing enzymes. Others include R6, which like PTG is ubiquitously expressed, RGL/GM specific to striated muscle, and GL found in rodent liver [7], [28]. It is therefore not surprising that skeletal muscle and brain of PTG-deficient mice still make glycogen, 70% and 30% of normal respectively [23]. What is surprising is the complete absence of LB in skeletal muscle in DKO. It would have been expected that if there is 70% glycogen synthesis in the absence of PTG, there would be 70% LB formation in the absence of laforin and PTG. Possibly, LB formation requires a threshold amount of glycogen. Alternatively, the laforin-malin complex in skeletal muscle acts specifically through PTG. On the other hand, if PTG is the preferred mediator of laforin-malin, it is surprising that its elimination from brain results in incomplete disappearance of LB, despite deeper glycogen reduction in brain in PTG deficient mice than in muscle. Much work ahead is needed to resolve these paradoxes.
The cause of neurodegeneration in LD has received much attention in recent years. Presence of up to 28% protein in some LB [9], [29], and signs of neurodegeneration in LKO mice at two months of age when LB are still small [24], led to considerations as to whether the neurodegeneration is related not to polyglucosans but to protein aggregation, similar to Alzheimer's and other neurodegenerative diseases [24], [30]–[32]. In the present study, correction of the neurodegeneration through interference in glycogen metabolism suggests that the neurodegeneration is connected to the disturbance in glycogen metabolism. This is consistent with recent observations that neurons, unlike other cell types, are highly vulnerable to increases in glycogen and polyglucosan content, with upregulation of GS leading to cell death [13]. Presence of small LB in two month-old LKO mice indicates that polyglucosans were already formed and accumulating by that time, likely triggering cell death, even as they had not yet formed large LB. Proteins in LB could be glycogen-metabolizing and other proteins trapped amidst aggregating polyglucosans.
Recently, it was reported that laforin enhances macroautophagy and that macroautophagy is dysfunctional in LD [33], indicating that laforin might function not only to prevent polyglucosan formation but also in clearing polyglucosans when they do form. Our results show that preventing polyglucosan formation obviates other laforin functions and suffices to prevent LD in mouse.
A major question in LD is why this particular neurodegenerative disease exhibits extremely severe epilepsy. Polyglucosans and LB occur in one other neurological disease, Adult Polyglucosan Body Disease (APBD), caused by mutations in the glycogen branching enzyme gene [34]. APBD LB differ from LD LB in one respect. For reasons unknown, they form exclusively in axons, especially long axons traveling to and from peripheral structures (skin, muscle, etc.) and the central nervous system. Affected patients suffer from motor neuron disease, may have mild dementia, but have no epilepsy [34], [35]. LD LB, on the other hand, are not seen in long tract axons, but instead almost completely replace the cytoplasm of vast numbers of small neuronal processes, mainly dendrites [3], [5], [6]. One possibility for the intractable epilepsy in LD is the progressive disturbance of dendritic function, the chief determinant of a neuron's excitability state. Near-complete disappearance of dendritic LB in the present study may account for the correction of the PME in our DKO mice.
In this paper, we correct the pathology and eliminate the PME of LD through genetic depletion of one of the proteins that targets the PP1 phosphatase to glycogen and the glycogen metabolizing enzymes. The effect on glycogen is partial, i.e. glycogen is not altogether eliminated, only reduced, the reduction returning the elevated glycogen levels of LD to normal wt levels, correcting the cardinal features of the disease, and causing no apparent harm to the mice. The crystal structures of PP1 [36], GS [37], [38], GP [39], [40], and GPK [41] are known, as is the PTG interaction domain with GS, GP and GPK [10], [12]. Identification of inhibitors of this interaction through rational design or large-scale small molecule screens could result in a treatment for this fatal epilepsy.
In addition to LD, accumulation of normal or abnormal glycogen is a cause of disease in several glycogen storage diseases including APBD and its severe fatal infantile form Andersen's disease [42], and the common and debilitating glycogenosis Pompe disease (acid maltase deficiency) [43]. Our results in LD suggest that removal of PTG could also improve these diseases. In fact, GS itself was recently removed from a Pompe mouse model resulting in a cure of the disease in that model [44]. While complete elimination of GS in humans cannot be contemplated as this causes significant pathology including sudden cardiac death [45], the Pompe study and our present work suggest that classes of medications that partially reduce GS or that partially reduce GS and activate GP, e.g. through interference in the PTG – GS/GP/GPK interaction, could have therapeutic benefit in multiple glycogenoses.
All animal procedures were approved by the Toronto Centre for Phenogenomics Animal Care Committee.
Laforin-deficient mice were a gift of Dr. AV Delgado-Escueta and S Ganesh. Mice were sacrificed by cervical dislocation and tissues immediately fixed in 10% formalin. Periodic acid-Schiff-diastase (PAS-D) staining was as previously described [17]. PAS stains normal glycogen and polyglucosans. The short treatment with diastase (amylase) digests glycogen but not polyglucosans. Diastase resistant PAS stained structures are LB. For GFAP staining, deparaffinized 5 µm sections were incubated with a polyclonal GFAP antibody (Dako) for one hour. Sections were thoroughly rinsed, and antibody visualized using diaminobenzidine conjugated avidin biotin complex (Vector).
Images from PAS-D slides were acquired at a 400× magnification (Olympus) by a CCD camera (Roper Scientific). Perikaryal and granular (neuronal processes) LB were distinguished by size and location. Numerical density [46] of both perikaryal and granular LB was then determined using the formula:where N is the number of either the perikaryal LB or the number of granular LB per unit volume of tissue (number/mm3), NpLBa is the number of perikaryal LB per area, NgLBa is the number of granular LB per area, d is the average diameter of either the perikaryal or granular LB, t is the thickness of the section (5 µm), and h is the smallest recognizable LB (1 µm). A minimum of 500 fields/animal were analyzed using an image analysis program (Image Pro Plus, Media Cybernetics, Bethesda). Data were expressed as means ± SEM and significance calculated using an ANOVA analysis.
Images from GFAP stained slides were acquired at a 250× magnification using the same microscope and equipment as above. The total number of GFAP positive cells was divided by the total area and expressed as cells/mm2. Genotype was blinded to the reviewer. A minimum of 250 fields/animal were analyzed. Images were analyzed using an image analysis program (Image J, NIH, Bethesda). Data were expressed as means ± SEM and significance calculated using ANOVA.
Mice were sacrificed by cervical dislocation and tissues quickly frozen in liquid nitrogen. Tissues were ground with a mortar and pestle in liquid nitrogen. Aliquots of 30–50 mg of tissue were mixed with 30% potassium hydroxide (KOH) and boiled at 100°C with frequent mixing. Glycogen was then precipitated with a final concentration of 67% ethanol at −20°C, then pelleted. This process was repeated three times. The purified glycogen samples were then dried and suspended in sodium acetate buffer. Glycogen was digested with amyloglucosidase (Sigma) at 37°C. Released glucose was determined using a glucose assay kit (Sigma). The amount of glycogen was calculated and expressed as µmoles of glucose per gram of tissue.
Brains for electron microscopy were taken from mice first perfused through the left ventricle of the heart with 2.5% glutaraldehyde in 0.1 M phosphate buffer (pH 7.4). The tissue was minced into cubic 1 mm blocks and fixed for an additional two to four hours. Samples were then washed in buffer and post fixed in phosphate-buffered 2% OsO4 for one hour. They were then dehydrated in an ascending series of acetones prior to being infiltrated, embedded and polymerized at 60°C overnight in embed 812-Araldite. Ultrathin sections were then prepared and stained with uranyl acetate and lead citrate prior to examination and image acquisition in the EM (JEOL JEM 1011, Peabody, MA).
Mice were placed in individual Plexiglas chambers and videotaped for four hours. Myoclonus was counted during periods when the animal was not exploring. Myoclonus counts were obtained in periods of a minimum of 10 minutes per mouse. The entire record was reviewed for detection of polymyoclonus. Observer was blinded to genotype. Myoclonus data in Figure 7 is shown as means ± SEM and significance calculated using an unpaired student's t-test.
|
10.1371/journal.pgen.1000877 | Bias and Evolution of the Mutationally Accessible Phenotypic Space in a Developmental System | Genetic and developmental architecture may bias the mutationally available phenotypic spectrum. Although such asymmetries in the introduction of variation may influence possible evolutionary trajectories, we lack quantitative characterization of biases in mutationally inducible phenotypic variation, their genotype-dependence, and their underlying molecular and developmental causes. Here we quantify the mutationally accessible phenotypic spectrum of the vulval developmental system using mutation accumulation (MA) lines derived from four wild isolates of the nematodes Caenorhabditis elegans and C. briggsae. The results confirm that on average, spontaneous mutations degrade developmental precision, with MA lines showing a low, yet consistently increased, proportion of developmental defects and variants. This result indicates strong purifying selection acting to maintain an invariant vulval phenotype. Both developmental system and genotype significantly bias the spectrum of mutationally inducible phenotypic variants. First, irrespective of genotype, there is a developmental bias, such that certain phenotypic variants are commonly induced by MA, while others are very rarely or never induced. Second, we found that both the degree and spectrum of mutationally accessible phenotypic variation are genotype-dependent. Overall, C. briggsae MA lines exhibited a two-fold higher decline in precision than the C. elegans MA lines. Moreover, the propensity to generate specific developmental variants depended on the genetic background. We show that such genotype-specific developmental biases are likely due to cryptic quantitative variation in activities of underlying molecular cascades. This analysis allowed us to identify the mutationally most sensitive elements of the vulval developmental system, which may indicate axes of potential evolutionary variation. Consistent with this scenario, we found that evolutionary trends in the vulval system concern the phenotypic characters that are most easily affected by mutation. This study provides an empirical assessment of developmental bias and the evolution of mutationally accessible phenotypes and supports the notion that such bias may influence the directions of evolutionary change.
| Random mutation does not generate random phenotypic variation because genetic and developmental architecture may constrain and bias the mutationally inducible phenotypic spectrum. Understanding such biases in the introduction of phenotypic variation is thus essential to reveal which phenotypes can ultimately be explored and selected through evolution. Here we used lines which had accumulated spontaneous random mutation over 250 generations starting from four distinct wild isolates of the nematode species C. briggsae and C. elegans, to study how a developmental system—vulval cell fate patterning—responds to mutational perturbations. We show that developmental defects and variants increase upon mutation accumulation in lines derived from all four isolates. However, some mutationally induced phenotypic variants occur more frequently than others, and the degree and spectrum of developmental variation further differed between isolates. These results illustrate how the phenotypic spectrum induced by random mutation can be biased due to both developmental system features and variation in the genetic background. Moreover, the mutationally most sensitive phenotypic characters are the ones that show most evolutionary variation among closely related species. These observations show how random mutation translates into a biased, limited range of phenotypes—a phenomenon likely impacting possible trajectories of phenotypic evolution.
| A principal quest in biology is to disentangle the relative contribution and interplay of mutational versus selective forces in the evolutionary process [1]. While biological research is predominated by the search for adaptive explanation underlying phenotypic evolution, it is also of critical importance to study how the mutational process alone produces phenotypic variation. Such studies indicate which phenotypic space can actually be explored by mutation to generate variation for selection to act upon. The mutationally inducible phenotypic spectrum is thus the fundamentally limiting force constraining and biasing potential evolutionary trajectories of the phenotype.
Importantly, the mutational spectrum is multidimensional and quantitative in character, where certain regions of the phenotypic space may be easier to reach by mutation than others. In quantitative genetic terms, the mutational variance VM of the phenotype represents the amount of variation introduced into the population by mutation each generation and can be extended to a multidimensional phenotypic space, theoretically the M matrix of mutational variance-covariance between phenotypic traits [2]–[4]. The structure of the mutationally accessible space can be best determined through the use of mutation accumulation (MA) lines, where mutations are allowed to accumulate for many generations with minimal selection [5]. Although the importance of the multivariate mutational process is well-appreciated theoretically [6],[7], empirical data are limited and most studies have focused on complex, composite traits, particularly life-history traits [8]–[10]. To our knowledge, no study has attempted to characterize the multivariate mutational structure of a developmental system.
It is evident that the genotype-phenotype map, encompassing organismal development, is highly non-linear, so that random mutation does not result in random phenotypic variation. For example, mutation may induce plentiful phenotypic variation for one trait but none for another. In the extreme case there is an absolute bias, so that certain phenotypes are impossible to generate though mutationally induced developmental changes, i.e. there is a developmental constraint [11],[12]. The phenomenon of developmental bias can be thought of as milder, relative constraint, where random mutational (or environmental) effects translate preferably into certain phenotypes [13]–[15]. Differences in such bias may be primarily quantitative and can be expressed as different probabilities of generating a given phenotypic spectrum upon random perturbation.
There is circumstantial evidence that developmental bias is common [13], [15]–[19]. In addition, experimental evidence suggests that genetic and developmental architecture bias the production of phenotypic variation. For example, repeated instances of parallel evolution indicate that evolution may follow a limited range of pathways [e.g. 20,21]. However, identifying the relative contribution of mutational versus selective forces in these comparative analyses remains challenging. Recent tests using experimental evolution approaches provide direct evidence on how genetic architecture may bias molecular variation made available to selection [22].
Overall, very few studies [e.g. 23] have quantified the inducible spectrum of phenotypic variation to evaluate whether “intrinsic” tendencies may influence the direction of phenotypic evolution. In general, as pointed out by Yampolsky & Stoltzfus (2001) there is little research focusing on experimental characterization of the spectrum of spontaneous variation and the underlying causes of molecular and developmental causes of any observed biases, which would allow testing the hypothesis that biases in the introduction of variation have influenced evolutionary patterns of the examined traits.
The mutational architecture may itself evolve, i.e. the regions of phenotypic space reached by mutation differ among genotypes. In other words, developmental bias is genotype-dependent. The inducible phenotypic spectrum for a given genotype has been referred to as “phenotypic neighbourhood” [24] or “local bias” [17]. Such evolutionary variation in mutational properties may be characterized by comparative quantitative analyses of mutation accumulation (MA) lines started from multiple distinct genotypes. Such studies show that mutational parameters may vary substantially between taxa and/or between genotypes of a single species [25]–[27]. We previously showed that mutational damage accumulates about twice as fast in C. briggsae as in C. elegans for lifetime reproductive output (≈“fitness”) [25],[28], body size [29], and at dinucleotide microsatellites [30]. These results reveal evolution of quantitative biases in the production of phenotypic variation (which could be due to evolution of mutation rates), but the underlying developmental and molecular causes of biases in the examined traits are so far unknown. To quantify and evaluate the significance of developmental bias and its genotype-dependence, analogous studies need to be carried out in simple, tractable developmental systems.
C. elegans vulval cell fate patterning is a model system for the study of intercellular signalling events [31] and has also served to study developmental robustness, cryptic variation and evolution [32]–[35].
The C. elegans hermaphrodite vulva develops from a subset of ventral epidermal blast cells, the Pn.p cells. In wild-type animals, three neighbouring cells, P5.p, P6.p and P7.p adopt vulval cell fates in the sequence 2°−1°−2°. Furthermore, three additional Pn.p cells, P3.p, P4.p and P8.p, have the capacity to adopt a vulval cell fate, when one or more cells of P5.p to P7.p are missing [36]. The six cells, P3.p to P8.p, therefore constitute the vulval competence group. During the second and third larval stages, the vulval precursor cells adopt alternative cell fates governed by an intercellular signalling network of Ras, Notch and Wnt pathways (Figure 1). A correct fate pattern of three vulval precursor cells (2°−1°−2°) is required to form a functional vulva. Deviation from this pattern can cause a reduction in offspring number due to impaired egg laying capacity and may further prevent male mating [34].
Vulval cell fate patterning is conserved among Caenorhabditis species [37]–[39]: P5.p to P7.p adopt vulval fates with the pattern 2°−1°−2° while all other vulval precursor cells adopt non-vulval fates, either a 3° fate (the Pn.p cell divides once) or a 4° fate (the Pn.p cell fuses early to the epidermal syncytium hyp7 without division). Species, however, may differ in the frequency of 3° versus 4° fate adopted by P3.p, P4.p and P8.p [37] and in the replacement competence of these cells upon laser ablation [38].
We previously quantified the precision of vulval development of (isogenic) C. elegans and C. briggsae isolates in multiple experimental environments [34],[37]. The results suggest that vulval development is robust to environmental and stochastic perturbations: apparent vulval defects occur in approximately 1 out of 1000 animals [34]. In contrast, developmental defects and variants increased significantly in mutation accumulation lines derived from a single C. elegans isolate, N2 [40], thus degrading the precision of vulval cell fate patterning [34]. This result indicates that mutation accumulation represents a feasible approach to quantify largely unbiased, mutationally induced phenotypic variation of this developmental system.
In this study, we examined the variation in mutational responses of the vulval developmental system within and between related species. We used mutation accumulation (MA) lines derived from two C. briggsae (HK104 and PB800) and two C. elegans (N2 and PB306) wild isolates that had accumulated mutations over approximately 250 generations [25]. We focused on quantifying and characterizing the spectrum of vulval developmental variants induced by spontaneous random mutation to address the following questions:
1) Does developmental precision decay upon mutation for all four isolates, and if so, can the action of natural selection be inferred by comparison of the degree of precision among wild isolates?
2) Does the vulval developmental system show a bias in its mutational response, i.e. are certain developmental variants more likely to occur than others? Which phenotypic characters of the developmental systems show maximal mutability?
3) Do the degree and spectrum of mutationally induced developmental variation vary between genotypes, i.e. to what extent is developmental bias genotype-dependent? How does the degree of mutability of a given developmental phenotype relate to its actual evolutionary variation within and between species?
The canonical vulval cell fate pattern in C. elegans and C. briggsae ancestral controls is 3°−2°−1°−2°−3° (P4.p to P8.p), whereas the most anterior P3.p cell adopts either a 3° or a 4° fate (Figure 1). The MA lines showed a consistently increased proportion of diverse variants (Figure 2), although the canonical P(4–8).p pattern remained the most frequent. Based on the observed variation in MA lines, we distinguished 13 distinct non-canonical cell fate variants deviating from the canonical vulval pattern (Material and Methods; Figure 2 legend). For some tests, these 13 variants were placed into three classes of decreasing order of vulva pattern disruption (A, B, and C). All variants were expressed in proportion of animals adopting the corresponding pattern.
Correlations of line means between two categories of non-canonical variant patterns (Class A and B) and two categories of fitness-related traits (W, CVE,W) are reported in Table S4. Given the number of variant categories and examined isolates, these tests are not powerful, but several trends emerged from the pattern of correlations. First, the correlation between class A variants (disrupted 2°−1°−2° pattern, likely resulting in defects) and other variants with complete 2°−1°−2° (class B+C) was positive in all isolates. The strength of the correlation between defects and variants was dependent on the starting genotype but was not species-specific: the correlation was strong and significant in C. briggsae PB800 and C. elegans PB306, but much weaker in the other isolates of each species. Second, the correlation between fitness traits and variants with complete 2°−1°−2° pattern (class B+C), but not variants with disrupted 2°−1°−2° pattern (class A), was stronger in C. elegans than in C. briggsae. In particular, the correlation between variant classes B+C and the within-line variance in fitness was uniformly strong and positive in C. elegans (∼0.5) and much weaker in C. briggsae (not significantly different from zero). The correlation in the VEL N2 lines was less than in the CFB N2 lines (∼0.2; not significantly different from zero). Third, all correlations were uniformly weak in the HK104 isolate of C. briggsae, a result we have consistently observed in this isolate [29].
To compare the mutational variance (VM) for variant vulval phenotypes with the standing genetic variance (VG), we analyzed data on developmental precision obtained from 10 C. briggsae and 25 C. elegans isogenic wild isolates (Nindividuals = 8′460). Wild isolate data are presented in Table S6, showing the proportion of variants for classes A, B, and C. Point estimates of the variance in line means (V L̂) were very low (∼10−5) for class A variants (strongly disrupted vulval patterns, defects) and for the pool of class B + C variant categories, and jackknife 95% confidence limits included zero in both categories in all isolates. Further, when isolates for which multiple estimates of p were available were considered, the maximum likelihood estimates for the among-isolate (genetic) component of variance were zero for both categories in both species. Thus, vulval development was highly invariant in both C. elegans and C. briggsae wild isolates, and most of the variant patterns observed were limited to variants of class C (3° to 4° transformation of P4.p/P8.p), in C. briggsae.
Across all four sets of MA lines, the different vulval variant patterns were observed at unequal frequencies (Table 1). Vulval precursor cells adopting a non-vulval 3° fate (P3.p, P4.p and P8.p) showed overall more variability than the cells adopting a vulval cell fate (P5.p to P7.p). Specifically, we found that the developmental phenotype with the highest mutational variance is that already showing high variability in the ancestral controls, i.e. P3.p division frequency (3° versus 4° fate; variant #14; class D) (Table 1 and Figure 3; note change of scale for this variant). The second most common variants concern P4.p and P8.p division frequency (variant #12 and 13; class C). Behind comes a subset of the variant patterns that affect the vulval fates such as centering shifts (class B), hyperinduction (class A or B) or missing precursor cells (class B). Therefore, variants causing likely defects in vulval function (class A) were overall less frequent than variants in classes B or C. That different sub-traits of the vulval developmental system degrade at different rates is further confirmed by the mixed-model analysis of the rate of change in the trait mean frequency Rm (see below).
To detect evolutionary variation in the mutability of the vulval developmental system, we tested for an overall interaction between variant vulval phenotype and ancestral genotype in an analysis of variance framework. The mixed-model analysis of the rate of change in the trait mean frequency Rm confirmed a substantial main effect of trait (nominal P<0.0001) and the expected large main effect of species (nominal P<0.002) (Table S5). Thus, the rate of change in mean frequency during mutation accumulation depended on the variant trait and the species. The main effects of isolate (nominal P>0.8) and trait x isolate (nominal P>0.10) were not significant. However, note that several of the most extreme differences in mutational induction of specific vulval variants occurred between the isolates of the same species rather than between species (see below).
Below we report specific examples of genotypic biases in mutationally induced phenotypic variants. Note that because of low frequency of developmental variants and multiple comparisons, the significance level of given comparisons may be poor (the critical experiment-wide 5% significance level for thirteen individual comparisons is P<0.0038).
The clearest examples of intraspecific variation in the mutational pattern are the hyper- and hypo-induction variants in C. elegans: MA lines displayed more hyperinduction variants and less hypoinduction variants in the PB306 isolate compared to the reference isolate N2. One hypothetical scenario to explain the elevated propensity to generate hyperinduced variants upon mutation accumulation in PB306 might be an increased activity of inductive vulval signalling, already present in the ancestral (wild type) genotype. In this scenario, such a difference would rarely be phenotypically expressed in the ancestral genetic background, but become more prevalent in MA lines due to mutational perturbations. To test this hypothesis, we asked whether the activity of the main signalling cascade inducing vulval cell fates, the EGF/RAS/MAPK cascade, was higher in PB306 than in N2. We introgressed an integrated construct containing a transcriptional Ras reporter, egl-17::cfp [41], into the two isolates to examine Ras activity levels during the vulval patterning process from mid-L2 to early-L3 stage (see Materials and Methods). Consistent with the hypothesis, PB306 showed a significantly higher Ras pathway activity in the relevant vulval precursor cell, P6.p, during mid-L2 and early L3 stages compared to N2 (Figure 4). Thus, the difference in the mutational accessibility of hyperinduced variants between PB306 and N2 may result through variation in the activity of the Ras pathway, which is phenotypically silent (cryptic) under normal conditions.
The developmental system underlying Caenorhabditis vulva precursor cell fate patterning was consistently degraded in mutation accumulation (MA) lines derived from all four isolates. In contrast to previously examined traits, such as body size, a quantitative trait varying along a single axis [29],[42], the variation is here practically absent among and within ancestral controls and mutational challenges induce novel variants. Vulval patterning variants almost always had a very low penetrance in a given mutation accumulation line. Many MA lines showed multiple, distinct variants and we never found a line in which a specific variant pattern was fixed.
The observed mutational pattern of small-effect variants may either be explained by non-null mutations in structural genes or mutations in regulatory regions with effects too small to be retained in conventional genetic screens. The core genetic elements of the vulval signalling network amount to approximately 30 genes [31], covering an estimated 150 kb. A conservative estimate of the mutation rate is one mutation per genome per generation in C. elegans [43], so that tested MA lines exhibit an average of 250 mutations per genome (100 Mb). Assuming that about a third of the nucleotide sites are susceptible to mutations having some phenotypic effect, the probability of mutating such a site in this category of “identified vulva genes” is 0.125 for a given MA line. This is consistent with the frequency of defects that we observe; however, this estimate is highly speculative, in particular, because we have no information on the distribution of mutational effects at a given locus. Moreover, it appears likely that several of the mutationally induced vulval variants may have been triggered by mutations of genes not directly involved in the vulval signalling network. Diverse developmental mutations primarily affecting body size and shape have the potential to disrupt the spatial and temporal integrity of the vulval induction process [44], and we have observed that many of these mutations (e.g., dpy, lon, sma, unc) show diverse low-penetrance vulval variants and defects similar to the ones observed in MA lines.
One consequence of the induction of deviation from an invariant pattern is an increase in the within-line component of variance. We previously demonstrated that the environmental (within-line) component of variance (VE) consistently increases with mutation accumulation for W, total lifetime fecundity, and body volume in these same lines [45]. We interpreted this result as evidence that spontaneous mutations de-canalize the phenotype, but could not completely rule out the possibility that that result was an artefact of the way in which these data were scaled. In contrast, the increase in vulval variants and defects with MA is most straightforwardly interpreted as an increase in the environmental component of variance, i.e., de-canalization, and it cannot be attributed to scaling. Thus, mutation accumulation increases the sensitivity of the vulval developmental system to stochastic (micro-environmental) perturbations [46].
We calculated an estimate of the standing genetic variance (VG) for variant vulval phenotypes using data from 25 C. elegans and 10 C. briggsae wild isolates. At mutation-selection balance in a large population, the ratio of the mutational variance to the standing genetic variance provides an estimate of the strength of purifying selection of mutations affecting the trait, i.e., S≈VM/VG, where S is the average selection coefficient against a new mutation. Using the point estimate of of the wild isolates as a surrogate for VG and the point estimate of ΔV as a surrogate for VM, the average selection coefficient against mutations affecting Class A variants inferred from the ratio VM/VG ( = S) is on the order of 10% or larger (for C. briggsae the point estimate of S = 0.30; for C. elegans S = 0.16). Conversely, the ratio VG/VM can be interpreted as the “persistence time” of a new mutation, i.e., the expected number of generations the mutation segregates before it is lost [47]. Thus, as expected, new mutations that cause Class A variants segregate for only a very few generations before they are removed by selection (Class A variants in the system are clearly deleterious in laboratory conditions, because they prevent egg-laying and reduce progeny number [34]). By way of comparison to life history traits in the same species, selection coefficients inferred in this way for W, body volume, and lifespan are on the order of 1–5% [48],[49]. This result confirms that vulval development is under strong purifying selection to maintain an invariant phenotypic output. The observed selection thus very likely corresponds to the type of stabilizing selection, as defined by Schmalhausen [50], and canalizing selection [51].
Concerning other variant classes, comparison of the genetic variance among wild isolates and after spontaneous mutation accumulation with minimal selection provides indirect evidence of their elimination by selection in natural populations. Especially in class B, the frequency of developmental variants was very low in the four controls as well as in a large set of wild isolates of C. elegans and C. briggsae covering a much larger range of genetic variation than the MA lines [43],[52] (Table S6). Averaged over variants and species, the ratio VM/VG ( = S) of Class B variants is again on the order of 10%, very similar to Class A variants (for C. briggsae the point estimate of S = 0.12, for C. elegans S≈1). Among the class B variants, variants with vulva centering shifts or missing Pn.p cells (variants #6–9) form a complete vulva due to cell fate regulation among the vulva competence group (cells that can adopt a vulval fate through expression of the lin-39/Hox gene [31]). Importantly, this result strongly argues for strong selection against class B variants in natural populations although these variants do not disrupt functionality of the vulval organ and show no fitness effects in the laboratory [34]. By contrast, selection against class C variants appears much weaker (S on the order of 0.1%). Class C variants describe variation in non-vulval fates of P4.p and P8.p, which normally do not affect P(5–7).p vulval fates. When adopting the variant pattern (i.e. adoption of the 4° fate), P4.p and P8.p fuse to epidermal syncytium without division in the L2 stage [53], so that the cells lose their competence to respond to late inductive vulval signalling. Nevertheless, these cells may still be able to respond to Wnt or EGF signalling earlier before hypodermal fusion, and thus to replace one of the P(5–7).p cells in the case of co-occurrence of a class B variant.
In contrast to classic mutagenesis screens selecting for developmental mutants with high penetrance phenotypes, the screening of the phenotypic spectrum of MA lines is largely unbiased and representative of the phenotypic spectrum induced by spontaneous random mutation. We found that MA induced certain phenotypic variants much more readily than others, demonstrating biases in the mutational accessibility of phenotypic variants. The vulval trait with the highest mutational variance is that already showing high variability in the ancestral controls (P3.p division frequency, variant #14), followed by P4.p and P8.p division frequency (variant #12 and #13; class C). Variants causing likely defects in vulval function (class A) were overall less frequent than variants in classes B or C. In addition, several of these variant patterns have not been found by mutagenesis in the laboratory, presumably because they were too subtle for efficient phenotypic scoring. On the other hand, we did not uncover all possible variant vulval patterns, which suggests that certain of these variants are either fully lethal and could not be propagated in MA lines, or their appearance through mutational effects is too improbable. Such variants include lateral inhibition defects with vulval cells showing adjacent 1° fates as seen in lin-12/Notch mutants [54]. Although a fully penetrant loss of lateral inhibition may be lethal, it is interesting that we did not find this variant at low penetrance like other fate pattern variants. This suggests that the mutational target size for this variant (relying on Notch pathway regulation) is small. Taken together, these observations provide clear examples of developmental bias [13]–[15],[18],[19], with certain phenotypic variants being more easily induced by mutation than others.
Several results show that biases in the production of vulval variants are genotype-dependent. First, overall rates of mutational decay differ among ancestral controls, most likely due to higher molecular mutation rates in the C. briggsae isolates compared to the C. elegans isolates [25],[30]. The approximately two-fold greater change in the trait mean in C. briggsae was roughly consistent with previous results concerning other traits [28],[29]. Second, we observed differences in the relative mutability of the same canonical pattern to different types of variant pattern. These differences in the mutationally inducible phenotypic spectra may be explained by one of two possible mechanisms. First, the mutation rate at specific loci may vary among wild isolates. For example, a microsatellite repeat present at these loci in some isolates and absent in others may dramatically change mutation rates at the locus [55]. Second, a distinct bias in the developmental system may occur if the internal system variables are slightly offset in some isolates towards the production of a given variant pattern. For example, C. elegans PB306 may mutate more frequently to genotypes producing hyperinduction defects if the Ras pathway involved in vulval induction is in average slightly more active in individuals of this isolate (compared to other wild isolates). More mutations of small effect on the system may then tip the balance towards hyperinduction when acting on the C. elegans PB306 isolate, and remain silent in other isolates. In this case, the different relative mutability to the hyperinduced phenotype of different starting genotypes may thus depend on cryptic genetic variation causing variation in system parameters, also termed intermediate phenotypes [32].
Apparent cryptic variation in such a quantitative developmental parameter may be confirmed by introgression of mutations or by measurements of signalling pathway activity. A higher Ras pathway activity in the C. elegans PB306 isolate is indeed supported by the higher induction index of let-60(n1046gf/ras, lin-3(n378rf)/egf mutants and of the ark-1(sy247lf); gap-1(n1691lf) double mutant [35]. Our present results using a reporter gene further confirm that the Ras pathway is significantly more active in C. elegans PB306 compared to C. elegans N2 (Figure 4). This result demonstrates the presence of intraspecific variation in the activity of vulval signalling pathways and agrees with the proposed second mechanism of evolution of the mutational variance through a bias in mutational effects. In the future, the determination of the molecular lesions and their introgression in different genetic backgrounds may definitively answer whether this difference accounts for the increased frequency of hyperinduced variants in PB306.
Mutational and environmental perturbations can both cause de-canalization of the phenotype [56]. Yet, there is limited experimental evidence whether these two sources of variation also affect the same elements of developmental systems. When comparing the phenotypic effects of mutational vs. environmental perturbation, analyses are often restricted to a single or few environmental conditions using a single or few genetic variants. MA lines provide a more extensive and unbiased sampling of genotypic space. Yet, unlike mutation, environments cannot be systematically sampled. We therefore limit our comparison to six environments examined in an earlier study [34], showing that certain vulval variants are specifically generated in certain environments and genotypes. Several of these previously observed variant patterns were also frequently found after MA. Specifically, vulval centering shift variants on P7.p were never found in C. elegans N2 MA lines, but occurred often in MA lines derived from the other three ancestral genotypes. Similarly, N2 never generated P7.p centering shifts under starvation stress, while C. briggsae showed increased and increased frequency of this variant pattern. Mutational perturbations therefore may mirror environmental perturbations, so that both sources of variation reveal the genotype-dependence of developmental bias.
Examination of different Caenorhabditis MA lines allows us to detect axes of high mutational variability in the vulval developmental system. Whether or not such high mutational variance translates into actual evolution then depends on selection. Some of these phenotypic axes of least resistance upon mutation may correspond to traits under purifying selection. In this case, the available mutational variance does not result in phenotypic evolution. For other variant types, however, the high mutational variance may correspond to phenotypic evolution observed in the species or among closely related species. In the Caenorhabditis genus, intra- and interspecific variation in vulval patterning traits is limited to the frequency of P3.p adopting a 3° versus 4° fate, and to a lesser extent that of P4.p [37],[39],[57]. For these two vulval phenotypes we also found the greatest mutational variance. The mutational bias and the evolutionary trend in the vulva system thus mainly affect the same trait. At a larger evolutionary scale, a similar match between mutational pattern and evolution is found in the Oscheius genus, but for vulva variants that concern the second round of 3° cell divisions (variants #10–11). In this case, the mutational variance in the occurrence of the second round of 3° cell divisions appears high in Oscheius tipulae CEW1 (from EMS-induced mutant lines) [58] and the same trait varies greatly within the Oscheius genus [24],[37],[39]. By contrast, we found very little mutational variation in the occurrence of a second division round for the 3° cells (variants #10–11), and these traits are invariant within the Caenorhabditis genus, presumably because of developmental constraints. Such studies of relative trait mutability are thus crucial to understand variation in evolutionary trends between taxa and thereby bridge the gap between micro- and macro-evolutionary variation.
In conclusion, our results provide an empirical view on the developmental variation induced by spontaneous random mutation. In the case of the highly canalized vulval developmental system, this variation is generally very subtle and difficult to quantify. In addition, the induced phenotypic variation is very complex despite the seeming molecular and developmental simplicity of this process. Nonetheless, we could uncover a number of developmental and genetic biases in the introduction of phenotypic variation, supporting the notion that such asymmetries bias the range of phenotypes available for selection to act upon [11]–[15],[18],[19]. Many more studies characterizing biases in the production spontaneous phenotypic variation (and its correspondence to evolutionary variation of the studied phenotypes) are required to evaluate whether such asymmetries play important roles as direction-giving forces in the evolutionary process.
The main set of mutation accumulation (MA) lines in this study is that of Baer et al. [25] (called CFB lines). The lines were originated from a single highly inbred individual from each of two isogenic wild isolates of C. elegans (N2 and PB306 isolates) and C. briggsae (HK104 and PB800 isolates). Criteria for choice of these isolates are given in [25]. The mutation accumulation experiments began with 100 replicate MA lines per isolate. Details of the mutation accumulation protocols are given in the original paper. Briefly, highly inbred stocks of each isolate were replicated 100 times and perpetuated by single-hermaphrodite transfer for 250 generations. This protocol results in a genetic effective population size of Ne≈1 (the approximation is the result of occasionally having to use backup stocks of worms when the original worm did not survive), thereby minimizing the efficiency of natural selection and ensuring that all but the most deleterious mutations behave according to neutral dynamics. Worm stocks, including G0 ancestral controls and ultimate generation MA lines, were cryopreserved using standard methods [59].
Wild isolates of C. elegans (N = 25) and C. briggsae (N = 10) used in this study are listed in Table S6. Both species display a high selfing rate in natural populations [52],[60]. The (isogenic) wild isolates were originally established by selfing populations derived from a single individual isolated from the wild.
Worms were kept on Petri dishes (55 mm diameter) filled with NGM (Nematode Growth Medium) agar, seeded with approximately 200 µl bacterial suspension of the E. coli strain OP50. All experiments were carried out at 20°C. For each of three experimental blocks, a random set of MA lines and the four ancestral controls were thawed (for samples size, see below). To eliminate potential genetic variation in the stock culture, a single individual from each line was selected to initiate the experimental populations. After population expansion, 20–30 adult hermaphrodites per line were hypochlorite treated to clear individuals form potential microbial contaminations. (At this time, for each of the four ancestral controls, multiple replicates were established except for the first block). The resulting eggs were allowed to develop into adults at which stage 20 hermaphrodites (from the same NGM plate) were transferred to a new NGM plate. When the majority of the offspring had reached the L4 stage (after approximately 2–5 days depending on the line), 50 offspring/line were randomly selected to score their vulval phenotype. The vulval cell phenotype was determined during the early to mid L4 stage using Nomarski microscopy on individuals anaesthetized with sodium azide [59]. We counted induced cells and determined the fates of the cells P3.p to P8.p as described previously [44]. MA and control lines underwent approximately 4–6 generations on NGM plates (at low densities) between thawing and scoring.
We defined different types of vulval developmental variants (shown in Figure 2) by taking into account developmental features of the system. Note that due to replacement regulation between vulval precursor cells [31], the fate of each individual cell is not independent from that of the other cells. For example, when the anchor cell is positioned on P5.p, the entire pattern is displaced anteriorly and four Pn.p cell fates are affected simultaneously; if P5.p is missing, P4.p adopts a 2° fate; if the anchor cell is missing, the fates of P(5–7).p switch to a 3° fate, etc. Defining 14 distinct variant types allowed us to greatly lower the number of variant types compared to the combination of each fate for each cell (1°/2°/3°/4°/missing x 6 = 30 classes). Some of these variants correspond to changes due to independent developmental events as defined by mutational analysis [24],[53],[61]. For example, hypoinduction phenotypes through cell fate change from a vulval fate to a non-vulval fate (trait #2) likely occur through low activities of Ras or possibly Wnt pathways (Induction Vulvaless in [61]). In contrast, hypoinduction phenotypes arising by lack of Pn.p cells (trait #3) occur because of cell death or earlier switch in cell fate (Generation Vulvaless in [61]).
The following number of MA and control lines were analyzed for each isolate: HK104 (44 MA lines, 17 control lines), PB800 (53 MA lines, 17 control lines), PB306 (51 MA lines, 17 control lines) and N2 (52 MA lines, 17 control lines). For each MA and control line, 50 individuals were scored for their vulval phenotype.
There are two fundamental observable quantities of interest in a MA experiment—the change in the trait mean and the change in the variance. In this study, vulval character state is a binary random variable X with state 0 = wild-type and state 1 = non-canonical” (for traits 1–13). The data are binomially-distributed with parameter p = Pr(X = 1). Within a genotype/treatment group (“treatment” = MA or G0 ancestral control), each line provides a single independent estimate of p.
If wild isolates are homozygous at all loci (a plausible approximation for a highly-selfing species; see above), the standing genetic variance (VG) can be estimated from the among-line component of variance [65]. However, for 22/25 wild isolates of C. elegans, we only have a single estimate of the binomial parameter p and therefore cannot meaningfully partition the variance in p into within and among-isolate components. Instead, we use the variance in isolate means V L̂ as an upper bound on VG. Using ΔV and V L̂ to approximate the mutational variance VM and VG, respectively, the relationship VG≈VM/S provides an estimate of the strength of selection against new mutations (S), provided the system is at mutation-(purifying) selection balance (MSB) [47]. For the isolates for which we have multiple independent estimates of p, we partitioned the variance into within- and among-isolate components using REML as implemented in the MIXED procedure of SAS v. 9.2. We can then compare the variance components of these isolates to V L̂ to gain a rough idea of the relative fraction of the variance that is among isolates.
To establish confidence intervals on ΔV and V L̂ we used a delete-one jackknife method [66] to estimate the standard error of the statistic, which was then used in the standard Student's-t calculation of the 95% confidence limits [67],
To estimate Ras pathway activity level in the C. elegans N2 and PB306 isolates, we used a previously generated transgenic strain containing an integrated transcriptional reporter construct for the LET-60/Ras pathway, egl-17::cfp-lacZ (strain GS3582) [41]. This construct contains a nuclear localization sequences upstream of the CFP coding sequence and was generated using the isolate N2 [41]. We then generated the egl-17::cfp-lacZ strain JU480 from the strain GS3582 by genetically removing the transformation marker unc-4(e120). Each integrated transgenic array generated in the N2 background was outcrossed ten times to PB306, by crossing at each generation the male progeny to wild hermaphrodites. After ten backcrosses, the introgressed line was made isogenic by selfing for several generations, yielding strain JU488.
The CFP fluorescence quantification experiment was performed as described in [34] in standard conditions at 20°C, for JU480 and JU488 in parallel. For each individual/image, we quantified signal (pixel) intensity of P5.p, P6.p and P7.p. For each examined developmental stage, we carried out an ANOVA (JMP 7.0 for Mac) testing for the fixed effects of isolate, individual (nested in isolate), cell, and the interaction between isolate and cell type using mean signal intensity as a response variable. The inclusion of the effect individual(isolate) allowed us to control for the non-independence between measures of P5.p, P6.p, and P7.p taken from a single individual. Post-hoc tests (Tukey's HSD) were then performed to determine differences in signal expression between isolates and cells (P5.p. P6.p, P7.p).
|
10.1371/journal.pgen.1003761 | Convergent Transcription Induces Dynamic DNA Methylation at disiRNA Loci | Cytosine methylation of DNA is an important epigenetic gene silencing mechanism in plants, fungi, and animals. In the filamentous fungus Neurospora crassa, nearly all known DNA methylations occur in transposon relics and repetitive sequences, and DNA methylation does not depend on the canonical RNAi pathway. disiRNAs are Dicer-independent small non-coding RNAs that arise from gene-rich part of the Neurospora genome. Here we describe a new type of DNA methylation that is associated with the disiRNA loci. Unlike the known DNA methylation in Neurospora, disiRNA loci DNA methylation (DLDM) is highly dynamic and is regulated by an on/off mechanism. Some disiRNA production appears to rely on pol II directed transcription. Importantly, DLDM is triggered by convergent transcription and enriched in promoter regions. Together, our results establish a new mechanism that triggers DNA methylation.
| DNA methylation in eukayrotes refers to the modification of cytidines at 5th position with methyl group (5mC). Though absent in some species, DNA methylation is conserved across fungi, plants and animals and plays a critical role in X chromosome inactivation, genomic imprinting, transposon silencing etc. In addition, DNA methylation also occurs at the promoter sequence to regulate gene expression. Filamentous fungus Neurospora crassa has a well-known mechanism of DNA methylation for genomic defense. During sexual stage repetitive sequences (e.g. transposons) are recognized and point mutations are introduced. During vegetative stage these mutations serve as signals for establishing static DNA methylation to silence all copies of the sequences. In this study, we report a new type of DNA methylation in Neurospora. It is tightly linked to a type of non-coding small RNA termed dicer-independent siRNA (disiRNA) and therefore was termed disiRNA loci DNA methylation (DLDM). DLDM is dynamic regulated and shows an on/off pattern, i.e. most alleles contain no 5mC but some are densely methylated. Interestingly, DLDM can be triggered by convergent transcription and is accumulated at promoter regions. In summary, our findings demonstrate a new type of dynamic DNA methylation.
| DNA methylation at the 5th position of cytosine to form 5-methylcytosine (5mC) is an important epigenetic gene silencing mechanism conserved from plants, fungi to animals [1], [2]. Even though most of the DNA methylation is relatively stable, dynamic DNA methylation has been observed during specific stages of animal development [3]. DNA methylation occurs in three different nucleotide sequence contexts: CG, CHG, and CHH (where H is C, A, or T). Small non-coding RNAs have been shown to be involved in the establishment and maintenance of heterochromatin formation in different organisms. In the fission yeast Schizosaccharomyces pombe, small RNAs and the RNAi pathway mediate histone H3 lysine-9 methylation at the centromeric regions [4]–[6]. In plants, the asymmetrical CHH methylation is maintained by de novo DNA methylation mediated by 24-nt small interfering RNAs (siRNAs) [review in 7], [8]. In mammalian germ cells, the Dicer-independent Piwi-interacting RNAs (piRNAs) are thought to be involved in DNA methylation [9], [10].
In the filamentous fungus Neurospora crassa, about 2% of cytosines in the genome are methylated [11]. Nearly all of the known methylation sites are within transposon relics of repeat-induced point mutation (RIP) and the repetitive ribosomal DNA locus [11]–[13]. RIP is a genome defense mechanism that results in mutation and methylation of duplicated DNA sequences during the sexual cycle [11], [14]. All previously known Neurospora DNA methylation is dependent on the histone methyltransferase, DIM-5, which meditates trimethylation of histone H3 at the lysine9 (H3K9me3) [15], [16]. Heterochromatin protein 1 (HP1) recognizes H3K9me3 and recruits DIM-2, the only confirmed DNA methyltransferase in Neurospora, to methylate DNA [17]–[19]. For the natural RIP'd sequences, DNA methylation is more or less stable and is generally not required for the maintenance of H3K9 methylation [12]. In addition, the known DNA methylation events are not dependent on the canonical RNAi pathway [20].
Neurospora produces many types of small RNAs, including microRNAs, siRNAs, QDE-2 interacting RNA (qiRNAs), and dicer-independent siRNAs (disiRNAs), through diverse small RNA biogenesis pathways [reviewed in 21]. disiRNAs are a distinct class of small RNAs as they symmetrically are mapped to both strands of DNA and their production is independent of the known canonical RNAi components, including Dicer [22]. The disiRNA loci range from a few hundred base pairs to more than 10 kilobases in size and are located in gene-rich regions of the genome. The function of disiRNA is unknown.
In this study we identified a new mechanism of DNA methylation that is associated with disiRNA loci. Our results showed that this type of DNA methylation, which we call disiRNA loci DNA methylation (DLDM), is very different from the previously known DNA methylation in Neurospora. DLDM is highly dynamic, depends on transcription at disiRNA loci, and is triggered by convergent transcription in some loci.
DNA methylation is an important regulatory mechanism of transcription silencing. Therefore, we examined whether in Neurospora the disiRNA loci are associated with DNA methylation using the methylation-sensitive restriction enzyme-based PCR (MSP) assays [23]. The Neurospora genomic DNA from a wild-type strain was digested with the isoschizomers DpnII or BfuCI. Both enzymes digest unmethylated GATC sites, but only DpnII can cut at sites when C is methylated. Primer sets were then used for semi-quantitative and quantitative PCR (qPCR). As shown in Figure 1A, a non-disiRNA locus (NCU06312) was not methylated, as indicated by the lack of PCR amplification product after digestion by BfuCI, whereas a PCR product was readily detected for the ζ-η region, a relic of RIP that forms constitutive heterochromatin and was previously shown to carry DNA methylation [14], [24]. For the 9 selected disiRNA loci with high level of disiRNA and with size larger than 5 kb, PCR products were detected in all samples after BfuCI digestion, indicating that these disiRNA loci are methylated. From quantitative PCR (qPCR) analyses, we estimated that percentages of methylation at the DpnII sites in these loci ranged from ∼3.5% to 28.8%, levels much lower than that of the ζ-η region (58.4%).
We then applied qPCR to MSP to further determine the DNA methylation profile in disi-47, disi-6, and disi-29 loci, which are three large loci with high levels of disiRNAs (Figure S1). The DNA methylation profiles at these loci are correlated with disiRNA profiles: DNA methylation peaked at regions of peaked disiRNA expression, and there was little or no DNA methylation outside of the disiRNA loci (Figure 1B and Figure S2).
Since the MSP results only reflect the methylation status of cytosines within the cleaved GATC sites, the detected methylation levels might be biased. To exclude this possibility, we performed methylated DNA immunoprecipitation (MeDIP) to detect DNA methylation based on 5mC density in disiRNA loci. For 13 loci (disi-6, 8, 22, 23, 29, 34, 35, 39, 42, 47–50) with relatively high disiRNA level and with size larger than 5 kb, all of them harbor DNA methylation (data not shown). We further analyzed disi-6, disi-29 and disi-47 loci with more primer sets for better resolution. As expected, the levels of DNA methylation were high in regions of high disiRNA expression and were low or absent outside of the disiRNA loci (Figure 1C and Figure S3). In constrast, no DNA methylation was detected at two negative control loci al-1 and NCU06312.
Since the previously known DNA methylation in Neurospora is located in regions derived from RIP, we calculated the RIP indices of the methylated disiRNA loci [25]. We observed no significant difference in RIP indices between that of the whole genome (1.11±0.33) and all of the 50 identified disiRNA loci (1.13±0.32). Moreover, for each disi-6, disi-29 and disi-47 loci, the lowest RIP indices are well above the threshold (0.7) for RIP-induced DNA methylation [25] (Figure S4). These data suggest that the DNA methylation in these disiRNA loci is not a result of RIP. We therefore named this type of DNA methylation DLDM (disiRNA loci DNA methylation) to distinguish it from the RIP-induced DNA methylation.
To determine the nature of DLDM, we performed bisulfite sequencing of selected genomic regions. As controls, we first sequenced the ζ-η region, a relic of RIP process and the am locus, which was previously shown to be unmethylated [26]. As expected, every clone of the DNA at the ζ-η region was methylated to various degrees at both symmetric and non-symmetric cytosine sites with an average methylation frequency of 36% (Figure 2A), similar to the estimated frequency (42%) determined in the MSP assay. In contrast, the genomic DNA was unmethylated at the am locus. In our experiment, 99.5% of all cytosines were converted to uracils. This percentage was similar to the conversion rate (99.8%) of an unmethylated PCR fragment after the bisulfite treatment, indicating that our bisulfite conversion was complete (data not shown). To our surprise, bisulfite sequencing of the disiRNA regions with peak DNA methylation revealed that vast majority of the clones were not methylated, as indicated by the nearly 100% conversion of all cytidines into uracils. For a few of the disiRNA loci clones, however, almost all cytosines were maintained after the bisulfite treatment (data not shown).
To rule out the possibility that these highly methylated clones were due to incomplete bisulfite conversion, we first treated the genomic DNA with bisulfite, then performed PCR and subcloned the fragments of either the disi-6 or the am locus into plasmids. PCR was then performed by using the plasmids as templates, and the resulting PCR products were subjected to DpnII digestion, which can only digest the PCR products if GATC site remains intact after bisulfite conversion due to the protection of 5mC (Figure 2B). As expected, all 111 clones of DNA from the am locus examined were resistant to the DpnII digestion, indicating that the bisulfite treatment of the genomic DNA was complete. For the disi-6 locus, however, 18 out of 123 DNA clones were cleaved by DpnII. Sequencing of the DpnII digestible (“cut”) and resistant (“noncut”) clones of disi-6 locus showed that in all “noncut” clones nearly 100% of cytosines were converted to uracils, indicating that the initial genomic DNA carried no 5mC (Figure 2C). In contrast, most of the cytosines in the DpnII digestible clones were methylated. Similar results were obtained from the disi-47 and disi-29 loci (Figure S5). Since Neurospora is haploid, these results indicate that the DNA methylation at disiRNA loci is actively regulated and is either on or off within each nucleus: disiRNA loci are either not methylated or are heavily methylated once DNA methylation process is switched on.
To confirm this finding, we used Southern blot analysis to visualize DNA methylation at several disiRNA loci. As shown in Figure 2D, both DpnII and BfuCI resulted in the same digestion pattern of the am locus, consistent with the lack of DNA methylation. For the ζ-η region, the BfuCI digestion resulted in the disappearance of one DNA fragment and the appearance of several additional higher molecular weight DNA fragments, consistent with all DNA molecules from this locus being methylated to some degree. The largest DNA fragment from BfuCI digestion was less than 2 kilobases (kb), indicating that methylation in the ζ-η region is limited to a small region. For the disi-47 and disi-29 loci, however, BfuCI digestion did not significantly change the levels or the relative ratios of the DpnII-digested bands, suggesting that most of the DNA lack methylation. We did observe, however, a ladder of high molecular weight (up to 8–10 kb) DNA fragments in the BfuCI-digested DNA, indicating that when methylated, the DNA at the disiRNA loci are heavily methylated across a large DNA region. Taken together, these results suggest that DLDM is highly dynamic and is regulated differently from the known DNA methylation events within relics of RIP in Neurospora.
The histone modification of H3K9me3 is mediated by the H3K9 methyltransferase DIM-5. This enzyme is essential for DNA methylation at relics of RIP, but H3K9 methylation at these loci is generally maintained in the absence of DNA methylation [12], [27]. To determine whether DLDM is mediated by H3K9me3 modification, we performed H3K9me3 chromatin immunoprecipitaton (ChIP) assays. Our results showed that the disiRNA loci are enriched by histones containing the H3K9me3 mark (Figure 3 and S6). However, in contrast to the ζ-η locus, in which the levels of H3K9me3 were not affected by the deletion of dim-2KO, the levels of H3K9me3 at the disi-47, -6 and -29 loci decreased dramatically in the dim-2KO mutant. These results indicate that DLDM requires DIM-2 for the maintenance of H3K9me3 at the disiRNA loci.
As disiRNA loci are gene rich, we examined whether transcription was important for disiRNA and DLDM. Our previous EST analyses showed that at least some disiRNA loci harbor fully or partially overlapped antisense transcripts, suggesting that convergent transcription might be a trigger of disiRNA production and/or DLDM [22]. We chose disi-47 locus to test this hypothesis since it harbors the well-studied circadian clock gene frequency (frq, NCU02265), which is known to produce the sense frq transcript and an overlapping antisense transcript qrf [28], [29]. The frq gene encodes a core component of the Neurospora circadian clock. MeDIP and MSP assays showed that the promoter region of the frq gene was methylated; however, only low levels of DNA methylation were detected in the frq coding region (Figure S3).
The transcription of both frq and qrf is activated by light in a process that mediated by the WHITE COLLAR (WC) complex, which consists of two PAS domain-containing transcription factors WC-1 and WC-2 [29]. Previous studies showed that the expression of both frq and qrf are high in constant light (LL) and are low in constant darkness (DD) [28]. Our qRT-PCR assays also confirmed that the level of frq mRNA was significant higher in LL than in DD (Figures 4A). Importantly, our mRNA deep sequencing [30] and RT-qPCR assays also demonstrated the existence of light-dependent expression of transcripts from the promoter region of frq (Figure 4B and S7), where disiRNA level is high. In a wc-2KO mutant, frq mRNA and the transcripts originating from the promoter region were both abolished, indicating that, like frq mRNA, the promoter-specific transcription requires WC-2. MeDIP assays showed that the level of DNA methylation in the promoter of frq was significantly higher in LL than that in DD in the wild-type strain and was completely abolished in the wc-2KO mutant (Figure 4C). Together, these results suggest that DNA methylation at the frq promoter is dependent on WC-2-mediated transcription and that transcription at promoter region may be necessary for the DNA methylation process and disiRNA production. Similarly, divergent promoter transcripts are also clearly seen in other disiRNA loci such as disi-29 (Figure S7), suggesting that this architecture of transcription might be the reason of DLDM.
To determine whether disiRNA production is dependent on the WC-dependent transcription, we performed small RNA deep sequencing analyses in wc-2KO strain. As shown in Figure 4D, the disiRNA abundance in the mutant is completely abolished at the disi-47 locus. This result is consistent with the loss of DLDM in the wc-2KO strain (Figure 4C) and suggests that disiRNA and DLDM are tightly linked and both triggered by pol II-dependent transcription.
Neurospora is a well-established model system for DNA methylation. All previously known DNA methylation in Neurospora occurs in relics of RIP [12], [32]. RIP is a process that silences repetitive DNA sequences during sexual stage (prior to meiosis) by converting cytosine to thymine in target sequences and occurs mostly at CpA dinucleotide context [33]. The resulted A/T rich region then serves as a signal that induces methylation of the nearby region to silence gene expression. In this study, we showed that DLDM is established and maintained very differently from the RIP-induced DNA methylation.
First, DLDM occurs in the gene-rich disiRNA loci that contain no relics of RIP or other repetitive elements. Second, in contrast to RIP'd regions in which DNA methylation is more or less constitutive and occurs in all alleles, most of the alleles are not methylated at disiRNA loci. In disiRNA loci, only a small percentage of alleles are extensively methylated with most of cytosines modified over a region that extends several kilobases. The dense cytosine methylation is similar to recently demonstrated dense methylation/hydroxylmethylation of cytosines in mouse embryonic stem cells [34]. The on-off pattern of DLDM indicates that DLDM is highly dynamic and that there is an inducible mechanism that mediates the establishment of DLDM. On the other hand, a de-methylation process may also exist to convert methylated alleles back into unmethylated alleles. Third, unlike the DNA methylation in the RIP'd regions, which is generally not required for maintenance of H3K9 methylation, DLDM is required for the maintenance of H3K9 methylation at the disiRNA loci. It suggests that an unknown mechanism should exist to recognize DNA methylation and in turn trigger H3K9me3.
Finally, DLDM is dependent on transcription. We demonstrated that DLDM is induced by convergent transcription from artificial constructs expressed in Neurospora. This conclusion is in agreement with the fact that most of disiRNA loci are known to produce sense and antisense RNAs [22]. Interestingly, the peaks of DNA methylation occurred at the promoter regions, where disiRNA expression also peaked. In addition, promoter-specific RNA transcripts were detected, and levels of these transcripts correlated with the levels of DNA methylation, suggesting that these non-coding RNA transcripts are involved in DLDM and are the precursors for disiRNAs. The induction of DLDM by transcription may explain its on/off pattern and suggests that a certain threshold level of transcription may be required for the establishment of DNA methylation. These results indicate that DLDM differs substantially from the typical DNA methylation in RIP'd DNA regions. It should be noted that transient DNA methylation was recently reported at the frq locus and was shown to be involved in setting the proper phase of circadian clock during the preparation of this paper [35]. In addition, the distribution of DNA methylation induced by convergent transcription is mainly accumulated upstream and peaks at about 3 kb from the TSS, suggesting that DLDM might also be involved in suppressing the promiscuous transcription at promoter region during transcriptional initiation. Indeed, recent studies suggest that pol II-directed gene transcription may adopt a gene loop structure by tethering promoter and terminator sequence, which enhances the transcriptional directionality toward the gene body [36]–[38]. Therefore, it is possible that the DNA methylation is a result of complex interaction between both ends of the gene for transcription initiation and termination, which strengthens the directionality of both sense and antisense transcription.
In this study, FGSC 4200 (a) was used as wild-type (WT) strain. Mutant strains wc-2KO, qde-1KO, qde-2RIP, qde-2RIP;sms-2RIP double mutant, and dcl-1RIP;dcl-2KO double mutant (dclDKO) were generated in previous studies [31], [39], [40]. The dim-2KO and dim-5KO strains were generously provided by Dr. Qun He [23]. Liquid cultures were grown in minimal medium (1× Vogel's, 2% glucose) at 30°C overnight and then at room temperature with shaking at 130 r.p.m. for 24 h [41]. For liquid cultures containing QA, 0.01 M QA, pH 5.8, was added to the liquid culture medium containing 1× Vogel's, 0.1% glucose and 0.17% arginine.
To make the his-3 targeting Pqa-2:cul:1-gccP constructs, a PCR fragment containing the promoter of ccg-1 was inserted into the plasmid pDE3dBH-Pqa-2 [42] to generate Pqa-2::1-ccgP. Then a PCR fragment of luciferase gene (luc) was inserted between the two promoters, with the luc sense transcripts and antisense transcripts driven by ccg-1 and qa-2 promoter, respectively. The control construct, Pqa-2:cul, was created by inserting the luc gene into pDE3dBH-Pqa-2, with qa-2 promoter driven antisense transcription of luc. The resulting constructs were introduced into the his-3 locus of dclDKO, his-3 strain [39], a his-3 strain and an eri-1lKO, his-3 recipient strain.
Approximately 10 µg genomic DNA was digested with BfuCI or DpnII, fractioned in 1.0% agarose gels, and transferred to nylon membrane. Hybridization probes were prepared from PCR products of interest (primer sequences in Table S1) with Rediprime II DNA Labeling System (GE Healthcare).
Approximately 10 µg genomic DNA was sonicated into small fragments (size ∼300–1000 bp). In each reaction, 1 µg of the 5-methylcytosine monoclonal antibody (Epigentek) was used to perform the MeDIP assay as previously described [12], [43]. MeDIP samples were analyzed with qPCR with corresponding primers listed in Table S1. In order to compare methylation in different regions, relative enrichment of DNA was calculated as the ratio of MeDIP sample over its input (set as 1), and the qPCR result of a primer pair of the am locus was used for normalization to correct for possible primer efficiency bias [44]. To compare MeDIP results of different samples or treatments, we performed the MeDIP at the same time with same batch of anti-5mC antibodies, due to the variation of MeDIP efficiency for different batches of antibodies.
The ChIP assay was performed as previously described [45]. The immunoprecipitation was performed with an H3K9me3 antibody (Abcam ab8898). The relative enrichment was calculated as the MeDIP assay and the qPCR result of a primer pair of the am locus was used for normalization.
Methylation specific PCR (MSP) was performed as previously described [23]. The methylation rate, determined by quantitative PCR, was calculated as the ratio of BfuCI-digested DNA signal to its input. A primer pair for (113–114), whose PCR product carries no BfuCI/DpnII recognition site (GATC), was used to normalize for loading and primer efficiency.
The bisulfite PCR methylation analysis was carried out in three steps: 1) The bisulfite treatment of genomic DNA was carried out as described in the manual of EpiTech Bisulfite Kit (Qiagen) except that we used a modified thermal cycler condition: 99°C for 5 min followed by 60°C for 25 min; 99°C for 5 min followed by 60°C for 85 min repeated 3 times; and 99°C for 5 min followed by 60°C for 90 min. 2) Two rounds of nested PCR were performed; the PCR product of first round was diluted 10–100 fold and 1 µL was used for second round of PCR. The second round PCR products of the expected size were cloned into TOPO clone kit (Invitrogen) and individual clones were sequenced.
Two strategies were used to examine DNA methylation in disiRNA loci. Strategy 1, shown in Figure 2, used plasmids as templates to amplify the cloned fragment. The genomic counterpart of the cloned fragment carries a GATC site. If the site was methylated, the fragment would be resistant to bisulfite treatment, whereas the unmethylated sites were converted into uridine and no longer recognized by DpnII or BfuCI. By identifying whether the PCR product was resistant to DpnII or not, we could distinguish whether the cloned fragment was methylated in the GATC sequence. Strategy 2, used in experiments in Figure S5, is similar to the first strategy except that one aliquot of genomic DNA was treated and one was not treated with BfuCI before bisulfite treatment, PCR, cloning, and sequencing. The average methylation rate was calculated by dividing total number of 5-methylcytosines by the total number of cytosines in the amplified sequence. The primers used for bisulfite sequencing are shown in Table S2 and S3.
Total RNA was extracted with TRIzol (Invitrogen), digested with Turbo DNase (Ambion) and reverse transcribed into cDNA with SuperScript II (Invitrogen). β-tubulin transcripts (primer pair tub) were used as loading control for quantitative PCR.
Total RNA of wc-2KO strain, wild-type strain and dicerDKO Pqa-2:cul:1-gccP strain were extracted with the TRIzol reagent (Invitrogen) and small RNAs were enriched with 5% polyethylene glycol (MW8000) and 500 mM NaCl as previously described [31]. Library construction and small RNA sequencing was performed by the Beijing Genomic Institute (Shenzhen, China) with Illumina standard protocol. All small RNA analyses were performed as described previously [22] except that an alignment tool Bowtie (ver 0.12.7) was used to map the small RNAs onto the N. crassa genome. In order to compare the density of small RNAs between samples, a standard normalization method was applied by scaling total reads of different samples to those of the same library size [46]–[48]. To correct bias induced by ribosomal RNA degradation products, we filtered out the reads matching rDNA regions from the total reads and used the remaining reads for scaling. The density of small RNA is presented as the relative number of small RNAs in a 100 nt non-overlapping sliding window along the Watson or Crick strand of each chromosome. The sRNA sequencing data was visualized with Generic genome browser (version 1.70) [49].
The NCBI accession number of the sRNA deep sequencing data reported in this study is GSE47666.
|
10.1371/journal.ppat.1000036 | A Novel Secretion Pathway of Salmonella enterica Acts as an Antivirulence Modulator during Salmonellosis | Salmonella spp. are Gram-negative enteropathogenic bacteria that infect a variety of vertebrate hosts. Like any other living organism, protein secretion is a fundamental process essential for various aspects of Salmonella biology. Herein we report the identification and characterization of a horizontally acquired, autonomous and previously unreported secretion pathway. In Salmonella enterica serovar Typhimurium, this novel secretion pathway is encoded by STM1669 and STM1668, designated zirT and zirS, respectively. We show that ZirT is localized to the bacterial outer membrane, expected to adopt a compact β-barrel conformation, and functions as a translocator for ZirS. ZirS is an exoprotein, which is secreted into the extracellular environment in a ZirT-dependent manner. The ZirTS secretion pathway was found to share several important features with two-partner secretion (TPS) systems and members of the intimin/invasin family of adhesions. We show that zirTS expression is affected by zinc; and that in vivo, induction of zirT occurs distinctively in Salmonella colonizing the small intestine, but not in systemic sites. Additionally, strong expression of zirT takes place in Salmonella shed in fecal pellets during acute and persistent infections of mice. Inactivation of ZirTS results in a hypervirulence phenotype of Salmonella during oral infection of mice. Cumulatively, these results indicate that the ZirTS pathway plays a unique role as an antivirulence modulator during systemic disease and is involved in fine-tuning a host–pathogen balance during salmonellosis.
| Bacteria of the Salmonella genus are important human pathogens and a leading cause of food-borne illness. Like for all other living organisms, protein secretion is a fundamental process, which is required for many different aspects of Salmonella biology including biogenesis of organelles, nutrient acquisition, and virulence. In this work we describe a new secretion pathway in Salmonella termed ZirTS. This pathway consists of an exported protein (ZirS) and a designated membrane translocator (ZirT), which mediates the secretion of ZirS to the extracellular milieu. Using a mouse model of Salmonella infection, we found that the ZirTS system is induced in Salmonella colonizing the small intestine and in Salmonella shed in fecal pellets during acute and persistent infections. Interestingly, inactivation of ZirTS results in a hypervirulence phenotype of Salmonella during oral infection of mice. These observations indicate that the ZirTS pathway plays a unique role as an antivirulence modulator and is involved in fine-tuning host–pathogen interactions during disease. Our study elucidates an emerging theme in pathogenesis emphasizing the importance of pathogens to limit their effects upon the cells they infect in order to achieve a balance with their host.
| Salmonella spp. are Gram-negative enteropathogenic bacteria that infect a variety of mammalian, avian and reptile hosts. Infection by this highly versatile pathogen can lead to different outcomes including asymptomatic carriage, gastroenteritis, or severe, life-threatening systemic disease, known as typhoid fever. The nature and the severity of the disease depend upon the serovar of the infecting Salmonella as well as the species and immunological status of the infected host [1].
The two hallmarks of Salmonella enterica serovar Typhimurium (S. Typhimurium) pathogenesis are the invasion of non-phagocytic cells such as epithelial cells of the intestinal mucosa, and the survival and replication inside infected phagocytic cells. Both mechanisms, as well as many of the virulence determinants used by S. Typhimurium, are directly linked to genes encoded within large horizontally acquired regions of the chromosome termed Salmonella pathogenicity islands.
Protein secretion is a ubiquitous cellular function found in organisms of all kingdoms. Gram-negative bacteria secrete a wide range of proteins whose functions include biogenesis of organelles, nutrient acquisition, virulence, efflux of toxins, and injection of virulence factors (effectors) into host cells. Protein export from the bacterial cytoplasm to the surface or the extracellular milieu requires transport across the inner membrane (IM), periplasm, and outer membrane (OM) of the cell envelope. In Gram-negative bacteria, several secretion pathways have evolved to fulfill this task [2],[3]. The auto-transporters (ATs) and the two-partner secretion (TPS) systems (often classified as the Type V Secretion System) have been the focus of much interest in recent years due to their prime role in virulence traits of Gram-negative pathogens [4],[5].
ATs are single functional units consisting of modular domains including: an N-terminal signal sequence that targets the protein to the general secretion (Sec) machinery at the IM; the passenger domain, which harbors the specific effector function; and the C-terminal translocation unit that forms, once inserted into the OM, a β-barrel secondary structure that mediates the secretion of the passenger domain. ATs are synthesized as pre-pro-proteins, and after cleavage of the signal peptide, the pro-protein is released into the periplasm. The passenger domain is then exported through the OM via the translocation unit, often cleaved off and released into the extracellular milieu [6],[7].
In contrast to the ATs, which are synthesized as a single polypeptide, in TPS systems the passenger domain and the transporter domain are translated as two separate proteins, referred to by the generic terms TpsA and TpsB, respectively [8]. TpsA proteins are synthesized with an N-terminal cleavable signal peptide and transported across the IM by the Sec machinery. Subsequently, TpsA substrates transit through the periplasmic space to their cognate secretion partner (TpsB) which then facilitates their secretion [9]. The TpsB cluster members show characteristic features of integral OM proteins and like TpsA, are thought to be exported across the IM by the Sec apparatus [9]. Conserved amphipathic motifs throughout their sequence indicate that TpsB proteins are likely to contain high numbers of transmembrane β-strands [10]. This secondary structure is believed to adopt a β-barrel conformation forming a pore in the OM that enables the translocation of TpsA across the OM into the extracellular environment.
Another group of proteins, which are conceptually analogous to ATs is the intimin/invasin (Int/Inv) family of adhesins. These family members are specialized OM proteins found in strains of Yersinia spp. (Inv), pathogenic E. coli (Int), and Citrobacter spp. (Int) that mediate adhesion of these pathogens to their hosts. Both invasins and intimins are translocated from the cytoplasm across the IM via the Sec-translocase and are related to each other both in terms of sequence and structure. The structure of Int/Inv includes a C-terminal C-type lectin receptor-binding domain, which is separated from a membrane-embedded N-terminal domain by several tandem Ig-like repeats, four in invasin and three in intimin. The conserved N-terminus domain is believed to form a β-barrel in the OM, which is used for the export of the C-terminal region. The extracellular C-terminus of Int/Inv is responsible for the receptor binding (Tir and β1 integrin, respectively) [11],[12]. Not much is known about the secretion mechanisms of Int/Inv, but based on existing similarities with ATs [13], it has been proposed that Int/Inv are secreted by an ATs-like mechanism [14].
In this report we describe the identification and characterization of a novel secretion pathway in Salmonella, named ZirTS. We show that ZirTS share important characteristics with the TPS systems and the Int/Inv family, and demonstrate that ZirTS play a unique role as an antivirulence modulator during systemic disease in mice.
Many virulence factors are pathogen-specific, however, a growing group of identified virulence determinants has been shown to harbor homology to various eukaryotic proteins or domains [15],[16], presumably as a result of continuous co-evolution with the eukaryotic host. Based on this idea, we developed a bioinformatic screen aimed at identifying Salmonella open reading frames (ORFs) that: (1) are absent from related non-pathogenic bacterial genomes and; (2) possess homology to known eukaryotic domains.
Screening the Salmonella Typhimurium LT2 genome while applying these bioinformatic filters led to the identification of an unknown ORF designated STM1668, located 26-bp downstream to an ‘invasin-like’ annotated gene (STM1669). Herein we rename STM1668 and STM1669 zirS and zirT, respectively, (see below).
No homologs of ZirS were found in the currently available genome databases in any bacterial genome outside of the Salmonella genus; however, weak homology was found to several eukaryotic proteins including a human zinc finger protein (NP_065798, 24% identity and 39% similarity over 199 amino acids). Additionally, zirS was found to be A+T rich (59.7%) in comparison to the rest of the S. Typhimurium genome (47%) and was located within a previously identified genomic island, GEI 1664/1678 [17]. These observations indicate that the zirS region was most likely acquired by a lateral gene transfer event during the evolution of Salmonella.
Interestingly, a highly conserved organization of the zirS region was found in all of the available Salmonella serovar genome sequences including S. bongori (Figure 1), implying that the lateral transfer event occurred before the divergence of S. enterica from the species S. bongori (∼35 to 40 million years ago), but after the split of Salmonella from the genus Escherichia (∼120 to 160 million years ago) [18],[19].
The neighboring ORFs adjacent to zirT and zirS include STM1670 and STM1667. STM1670 is annotated as a putative serine/threonine protein kinase and located 74-bp upstream to zirT. STM1670 homologs are currently found only in Salmonella databases, suggesting that it might be a unique Salmonella protein. STM1667, which is encoded 98-bp downstream to zirS contains a conserved peroxiredoxin domain and is annotated as a putative thiol peroxidase.
ZirS is predicted to be a 276 amino acid protein with an estimated molecular mass of 30.8 kDa. Sequence analysis of ZirS using the SignalP 3.0 program [20] (http://www.cbs.dtu.dk/services/SignalP/) predicted a typical prokaryotic Sec-dependent signal sequence at its N-proximal region, with a potential cleavage site between amino acids 24 and 25 (VLA▾DS). In Gram-negative bacteria, the presence of a signal sequence suggests that the protein is processed and exported across the IM in a Sec-dependent fashion. This process involves the cleavage of the signal peptide by a LepB leader peptidase (type I signal peptidase) and requires ATP hydrolysis by a designated ATPase, which provides the driving force for translocation (reviewed in [21]). To examine this hypothesis experimentally, a tagged version of ZirS was constructed using a C-terminal two-hemagglutinin (2HA) tag and cloned into a low copy number vector in the presence of the upstream gene, zirT (pOG-zirTS-HA). As illustrated in Figure 2A, expression of zirS-HA (in the presence of zirT) in S. Typhimurium led to prominent secretion of ZirS-HA into the medium as detected by Western-blot analysis.
To assess the contribution of the Sec-translocon to the extracellular export of ZirS, the secretion of ZirS-HA in the absence and presence of azide was studied. Low concentrations of azide (2 mM) specifically inhibit SecA, the ATPase component of the Sec-complex, and therefore interfere with Sec-dependent protein secretion, resulting in accumulation of pre-proteins in the cytoplasm [22],[23]. In performing this experiment, Salmonella strain expressing ZirS-HA (pOG-zirTS-HA) that was grown in LB to late logarithmic phase was washed, resuspended in fresh medium, and incubated for 90 or 120 min in the presence or absence of 2 mM sodium azide. Subsequently, the intracellular and the secreted ZirS-HA were analyzed by Western-blot. As demonstrated in Figure 2B, the presence of low concentrations of azide strongly reduced the secretion of ZirS-HA into the medium and led to accumulation of a higher molecular-weight (pre-ZirS-HA) isoform in the cytoplasm. We concluded from these experiments that the secretion of ZirS into the extracellular environment is dependent on the function of the Sec-translocon and involved signal peptide cleavage at the N-terminus of ZirS.
ZirT, encoded by the gene immediately upstream to zirS, is predicted to be a 660 amino acid protein with an estimated molecular mass of 72.7 kDa. In contrast to ZirS that showed no prokaryotic homologs outside of the genus Salmonella, a bioinformatic search against non-redundant protein databases revealed several bacterial protein groups that share significant homology with ZirT. All are known OM proteins including various invasins and intimins from different Gram-negative pathogens (Figure 3). More precise comparison of ZirT to these proteins revealed that the sequence similarity is concentrated within the mid-N-terminal region of ZirT, spanning from amino acid 88 to 368. Importantly, these homologous regions are thought to form porin-like β-barrels in the bacterial OM [13].
Further sequence analysis of ZirT also predicted a Sec-dependent signal sequence in the N-terminus with a potential cleavage site between amino acids 27 and 28 (VIA▾DS), which supported the possible export of ZirT from the cell cytoplasm. In agreement with the sequence homology found, other localization prediction tools (PSORTb v.2.0 http://www.psort.org/) [24] suggested a subcellular localization of ZirT in the OM, with 31 predicted trans-membrane β-segments (TMBETA-NET, http://psfs.cbrc.jp/tmbeta-net/) [25]. In order to investigate the subcellular localization of ZirT, a C-terminus HA tagged version was constructed and cellular fractionation analysis of Salmonella cells expressing ZirT-HA (pOG-zirT-HA) was performed. This experimental approach showed the localization of ZirT-HA in the cellular membranes fraction (Figure 4A).
To further characterize the precise localization of ZirT in the cell envelope, we utilized sucrose density gradient ultracentrifugation fractionation. S. Typhimurium total membranes fraction was isolated and subjected to ultracentrifugation through a sucrose density gradient (30–60% sucrose, w/v). The specific localization of ZirT was determined based on the presence of OmpA and β-ATPase used as controls for the OM and the IM fractions, respectively [26],[27]. As shown in Figure 4B, ZirT-HA was found to be distinctively localized into the OM fractions.
In general, many OM proteins with β-barrel structures exhibit heat-modifiable electrophoretic mobility behavior, in which strong resistance to denaturation in the presence of 1% SDS is observed, unless heated to 100°C. Consequently, the folded and the compact β-barrel conformations migrate more quickly in SDS-PAGE than their denaturated forms [28]. As ZirT was found to have sequence similarity to known β-barrel OM proteins and predicted to be amphipathic β-strand rich, we investigated whether ZirT also exhibited heat-modifiable electrophoretic mobility. When protein extracts from Salmonella cells expressing ZirT-HA (pOG-zirT-HA) were incubated in a sample buffer without boiling and analyzed by SDS-PAGE, instead of running at its expected denatured position of ∼75 kDa, non-boiled ZirT-HA migrated mainly as a faster protein band at ∼62 kDa (and a secondary band at ∼55 kDa). Heat denaturation of the samples before analysis reproduced the unfolded form in a time and temperature-dependent manner (Figure 4C). These data suggest that the folded, mature ZirT is arranged into β-barrel architecture in the OM.
It is noteworthy that Western-blot analyses against the denatured tagged version of ZirT allowed the detection of 2–3 distinct molecular-weight bands (Figure 4). This observation might indicate possible processing of ZirT, resulting in different protein isoforms.
We next focused our interest on examining possible interactions between ZirT and ZirS. The secretion of ZirS-HA was, therefore, analyzed while expressed from a low-copy number construct harboring ZirS-HA alone (pOG-zirS-HA), or from a plasmid containing both ZirT and ZirS-HA (pOG-zirTS-HA). This assessment was done in three different S. Typhimurium genetic backgrounds (wild-type, ΔzirS, and ΔzirT strains) using immunoblots against the cellular and the secreted protein fractions. Interestingly, expression of ZirS-HA was observed in both the presence and absence of ZirT, as evidenced in the cellular fractions. However, in contrast to protein expression, secretion of ZirS-HA was only detected when ZirT was co-expressed (Figure 5A). Thus, secretion, but not expression, of ZirS was found to be dependent on the presence of ZirT.
Unexpectedly, the same results were observed in all three genetic backgrounds, including the wild-type and a ΔzirS mutant strain, both carrying a chromosomal copy of zirT that did not seem to support the secretion of ZirS-HA expressed from an episomal construct. Two possible interpretations of this result were: (1) the nature of the interaction between ZirS and ZirT requires specific stoichiometry that was not achieved from chromosomal expression of ZirT, or that (2) the secretion of ZirS demands the presence of ZirT in cis. To examine these possibilities we analyzed the expression and the secretion of ZirS-HA in the presence of ZirT that was provided either in cis from the same episomal construct (pOG-zirTS-HA) or in trans from a different vector with a similar (low) copy-number (pOG-zirS-HA and pOG-zirT-4). As can be seen in Figure 5B, providing ZirT either in cis or in trans from a similar copy number vector was able to complement the secretion of ZirS, implying that particular stoichiometry of ZirS and ZirT might be required for efficient secretion of ZirS.
To gain further insight into the nature of ZirS secretion, we expressed the HA-tagged version of ZirS alone (pOG-zirS-HA) or together with ZirT (pOG-zirTS-HA), in a heterologous E. coli K-12 host that does not possess any homologs of ZirTS (or any neighboring genes). Introducing ZirS-HA alone into an E. coli host resulted in detectable expression, but not secretion of ZirS-HA. In contrast, introducing ZirS-HA together with ZirT led to prominent secretion of ZirS-HA into the medium by E. coli (Figure 5C). We concluded from this experiment that zirTS encodes an autonomous and self-sufficient secretion system, in which ZirS is secreted in a strict ZirT-dependent manner.
Cumulatively, the data presented describing the nature of ZirT and ZirS are consistent with several key characteristics of TPS systems and/or the Int/Inv family. These similarities includes: (1) primary sequence homology to various intimins and invasins; (2) ZirT being an outer membrane β-barrel protein, similar to TpsB or the N-terminus module of the Int/Inv members; (3) like TpsA, ZirS seems to be translocated from the cytoplasm across the IM via the Sec translocase; (4) ZirT containing a prototypical N-terminal signal sequence, as the TpsB and the Int/Inv members; and (5) ZirS being secreted into the extracellular milieu in an explicit ZirT-dependent manner, analogous to the relation between TpsA and TpsB.
Nevertheless, despite the shared similarities, some fundamental differences exist between the ZirTS and the compared systems (see Discussion). Based on this, we suggest that the ZirTS secretion system is functionally similar to the TPS pathway, but represents a distinctive secretion pathway in Salmonella.
In order to understand better the regulation the ZirTS pathway, we were interested in identifying different regulatory factors that govern the expression of zirTS. Using the Virtual-Footprint program (http://www.prodoric.de/vfp/) [29] we were able to identify a potential binding site for OxyR located 97-bp upstream from the start codon of zirT. The OxyR transcription factor is a LysR-type regulator that activates the expression of numerous genes in response to oxidative stress [30]. To test whether OxyR affects zirTS expression, S. Typhimurium SL1344 oxyR mutant strain and a reporter-gene construct harboring a fusion between zirTS and a promoterless β-galactosidase gene (zirTS::lacZ) were constructed. Since zirTS::lacZ was found to be most strongly induced in M9 minimal medium (pH 7.4) in comparison to LB (742±73 and 258±21 M.U., respectively), the expression of this reporter-gene fusion was compared in both strains under these conditions. As demonstrated in Figure 6A, zirTS::lacZ expression was found to be about 2.3 fold higher in the ΔoxyR background than the wild-type strain (P<0.0001), suggesting that OxyR is involved in the regulation of these genes. To further confirm these results we applied a qualitative real-time PCR approach and compared the abundance of zirT and zirS transcripts in the ΔoxyR background vs. the wild-type strain. In agreement with the lacZ reporter-gene results, the expression levels of zirT and zirS were about 1.5 (P = 0.0033) and 2.5 (P<0.0001) fold, respectively, higher in the ΔoxyR background, compared to the wild-type strain (Figure 6B). Together, these results suggest that OxyR plays a role as a negative regulator of the ZirTS pathway.
The induced expression of zirTS::lacZ in M9 minimal medium in comparison to LB and the initial identification of ZirS as a Salmonella protein, which presented some sequence homology to eukaryotic zinc-binding proteins, prompted us to examine possible effect of different metals ions on the regulation of zirTS. To investigate this, we complemented defined M9 medium with different metal salts and examined the expression levels of zirTS::lacZ under these conditions. Addition of Mg2+, Fe2+ or Fe3+ ions to the medium did not alter the expression of zirTS::lacZ; however, addition of subinhibitory concentrations of Zn2+ ions resulted in moderate but statistically significant (P<0.0001) reduction of zirTS::lacZ expression, in a dose-dependent manner (Figure 6C). These results implied that Zn2+ may repress the expression of zirTS. If this assumption were true, we expected that addition of metal chelators to LB broth would lead to an increased expression of zirTS::lacZ. Indeed, addition of different divalent metal chelators (DTPA, EDDA, and TPEN) resulted in a significant (P<0.0001) increase of zirTS::lacZ expression. Furthermore, when the presence of these metal chelators was counteracted by the addition of excessive Zn2+, induction was prevented. Addition of excessive Fe2+ did not prevent zirTS::lacZ induction, indicating that the observed repression is zinc-specific (Figure 6D).
In order to further support these results, we implemented a quantitative RT-PCR methodology and compared the abundance of the zirT and zirS transcripts in minimal medium supplemented with or lacking zinc, as well as in LB in the presence or absence of metal chelators. As illustrated in Figure 6E, the addition of zinc salt to an M9 defined medium, decreased the expression of zirT and zirS by about 2 and 2.5 fold, respectively (P<0.0001). As oppose to that, when the metal chelators DTPA was added to LB broth, a moderate induction of zirS expression, by more than 2.6 fold (P<0.0001) was observed. Strikingly, when LB was supplemented with the intracellular zinc-specific chelator, TPEN [31], stronger induction by more than 2 and 5 fold, was observed (P<0.0001) in the expression of zirT and zirS, respectively.
Collectively, we concluded from these experiments that zinc significantly contributes to negative regulation of the ZirTS pathway and therefore we named STM1668 and STM1669, zinc regulated secreted protein (zirS) and zinc regulated transporter (zirT), respectively.
To investigate the role of ZirTS in vivo, we used the murine model for systemic salmonellosis and evaluated the survival-time of BALB/c mice infected orally with ∼1×106 cfu of wild-type, ΔzirS, and ΔzirT S. Typhimurium strains. Surprisingly, the median survival-times of ΔzirS and ΔzirT strains were 7.5 and 6 days, respectively, while the median survival-time of the wild-type strain was longer (8.5 days), implying the possibility that the ΔzirS and ΔzirT strains might possess virulent capability higher than the wild-type. However, although a trend was apparent, with a sample size of 8 mice in each group, these differences were not statistically significant (P>0.05).
In many cases, comparing survival time is not sensitive enough to reveal virulence differences, especially when two virulent strains are compared. In contrast, the competitive index (CI) approach [32] is considered to be more sensitive to subtle differences. In these experiments, mice were challenged orally with a mixed inoculum containing equal numbers of wild-type bacteria and a mutant strain carrying an in-frame deletion of zirS. Six days post infection (p.i.) mice were sacrificed and the recovered cfu ratio between the mutant and the wild-type strain (i.e. CI value) was evaluated. As the main sites of Salmonella replication during systemic infection are the spleen and the liver [33]–[35], the CI geometrical mean was calculated for these sites. In 129X1/SvJ mice, the mean CI was found to be 4.4 and 4.2 for the spleen and liver, respectively (Figure 7A), indicating that the ΔzirS mutant strain outcompeted the wild-type strain by more than 4 fold during the infection. Comparable results were also obtained when a ΔzirT mutant was competed against the wild-type strain (data not shown). This CI analysis correlates with the single infection results, which suggested a shorter survival-time of mice infected with ΔzirS and ΔzirT mutant strains in comparison to the wild-type, and together demonstrated a hypervirulent phenotype for ΔzirS and ΔzirT mutant strains in mice.
Nramp1 (natural resistance-associated macrophage protein-1; also known as Slc11a1) is a host resistance gene that provides protection against several intracellular pathogens, including S. Typhimurium [36]. In order to examine a possible effects of Nramp1 on the observed hypervirulent phenotype, and to further validate these results, we repeated the CI analysis in 129Sv/J (Nramp1+/+) and isogenic Nramp1-deficient (Nramp1−/−) mouse genetic backgrounds. As can be seen in Figure 7B and C, a similar trend was observed in these mouse strains, with an even more pronounced difference in the Nramp1−/− background. In the latter, the mean CI values were 10.3 and 12.7 in the spleen and the liver respectively, demonstrating a significant overgrowth of the ΔzirS mutant in comparison to the wild-type strain.
Next, we examined whether the apparent hypervirulence behavior of the ΔzirS mutant is dependent on the route of infection. To test this, 10 129X1/SvJ mice were infected intraperitoneally (i.p.) with approximately 2×104 cfu and sacrificed 3 days p.i.. Intriguingly, as indicated in Figure 7D, when the bacteria were administrated i.p., both strains reached equal numbers and no growth advantage of the ΔzirS mutant was observed, suggesting that following i.p. infection, wild-type and the ΔzirS mutant are equally virulent. Similar results were obtained during CI infection of a ΔzirT mutant versus a wild-type strain (data not shown).
We infer from these experiments that the absence of the zirS (or zirT) gene leads to hypervirulence of Salmonella in vivo, resulting in a significant (3–12 fold, P<0.05) overgrowth of the ΔzirS strain in systemic sites. Interestingly, the differential growth of the ΔzirS mutant was evident only following oral infection but not when the mice were infected i.p.. Since the lack of ZirTS leads to an increased virulence of Salmonella, we propose that the ZirTS secretion pathway functions as an antivirulence modulator during systemic disease in mice.
The results attributing an antivirulence function to zirTS during Salmonella infection in mice, prompted us to study the expression pattern of the ZirTS pathway during the course of both acute and persistent infections in the murine model. A reporter-gene fusion between zirT and the luxCDABE operon was constructed and introduced into wild-type S. Typhimurium. As a positive control for in vivo expression, we used a similar construct containing a Salmonella sigma 70 (RpoD) dependent promoter [37]. Both reporter strains, designated zirT::lux and rpoD::lux, were used to infect C57BL/6 mice orally. Three days post-infection mice were sacrificed and immediately examined for luciferase activity in the gastrointestinal (GI) tract and systemic sites. Although zirT::lux had about two logs lower expression levels in vitro in comparison to rpoD::lux (data not shown), induced and distinct expression of zirT::lux was observed in vivo. In contrast to rpoD::lux, which showed strong expression in systemic sites (liver, spleen, and mesenteric lymph nodes) and the small intestine, zirT::lux expression was localized primarily within the small intestine, while at systemic sites, zirT::lux expression was low, despite being heavily colonized with Salmonella (Figure 8). These results suggest that the in vivo expression of zirT is induced mainly throughout the gastrointestinal tract of infected animals rather than at systemic sites. This unique expression pattern is in agreement with the CI experiments that showed a virulence difference between the wild-type and a zirS mutant strain following oral, but not i.p., infection. In light of these data, suggesting that zirT is not abundantly induced at systemic sites; we propose that by administrating the bacteria i.p., the induction of zirT in the GI tract was bypassed and differences between the wild type and the zirTS mutants were not noticeable.
Besides the induction in the small intestine, strong expression of zirT::lux was apparent in the fecal pellets from both C57BL/6 (Nramp1 negative) mice that developed an acute systemic disease and 129X1/SvJ (Nramp1 positive) mice carrying a persistent Salmonella infection (Figure 9). Remarkably, the induced expression of zirT::lux in Salmonella shed within fecal pellets of both mouse strains was evident starting from day one p.i. and lasting at least 8 weeks p.i. in 129X1/SvJ mice during persistent Salmonella infection. At 6 weeks p.i., due to low levels of Salmonella shedding (as determined by cfu counts), expression of zirT::lux was not detected in the feces, but reappeared at 8 weeks p.i., when high numbers of Salmonella were shed. These results indicate strong and continuous expression of zirT in the feces of shedding mice.
In this study we described the identification and characterization of a Salmonella-conserved autonomous and previously unidentified secretion pathway, termed ZirTS. We showed that this novel secretion system consists of an OM protein, ZirT that is expected to adopt a β-barrel structure and the exoprotein ZirS, which is secreted into the extracellular environment in a strictly ZirT-dependent manner.
The ZirTS secretion system was found to share several important features with members of the TPS system and the Int/Inv family. However, despite the mutual similarities, some fundamental differences exist between the ZirTS and the compared systems. Obviously, as opposed to Int/Inv members, ZirTS function as a secretion system comprised of two separate components. Likewise, comparison of ZirTS to TPS systems reveals several hallmarks of TPS, which are absent from ZirTS. The most characteristic feature of the TPS exoproteins is a conserved N-terminal domain, known as the TPS domain. This region has been identified in all of the TpsA proteins characterized thus far and includes highly conserved NPNGI and NPNL motifs. The TPS domain is necessary for secretion and has been proposed to mediate recognition of the exoprotein by the transporter [8], [9], [38]–[40]. An additional feature of TPS is the sequence conservation of TpsB family members, particularly within the C-terminus [41]. Amino acid sequence analysis and comparison of ZirS and ZirT to other known TPS members indicated the lack of these conserved modules in ZirS and ZirT and therefore differentiate ZirTS from the TPS pathway. Based on this we contend that zirTS encode a novel secretion pathway in Salmonella.
Adapting the currently accepted mechanism of TPS systems, we propose the following model for the modus operandi of ZirTS: pre-ZirS is translocated across the IM, via the Sec translocase. Upon translocation, the signal peptide of pre-ZirS is cleaved off by a specific periplasmic signal peptidase and the mature ZirS is released in the periplasmic space. As was shown for other TPS systems, there appears to be no accumulation of periplasmic intermediates [8] and ZirS is expected to transit through the periplasmic space, only briefly. Following Sec-dependent export across the IM, ZirS interacts with ZirT and then is translocated across the OM to the extracellular environment through a hydrophilic β-barrel pore formed by ZirT.
One of many intriguing questions related to ZirTS is its evolutionary origin. A conceivable scenario for the evolution of such a system might be a molecular separation of an ancestral Int/Inv related protein (or an AT) into two distinct functional polypeptides, early in the evolution of Salmonella. Other circumstances that might explain the development of ZirTS include a later acquisition of zirS into the genome and the “exploitation” of an already existing OM β-barrel porin for its secretion.
Thus far, Int/Inv and TPS systems have been shown to be primarily involved in different virulence traits of Gram-negative pathogens. In contrast, the ZirTS secretion pathway demonstrated an interesting and somewhat surprising antivirulence activity, rather than a virulence determinant, in the murine salmonellosis model. Presently, it is still unknown if the ZirTS pathway is unique to Salmonella or related systems exist in other bacteria. A bioinformatic search for non-invasin/intimin ZirT homologs has led to the identification of a predicted 934 amino acid OM protein (SG0602) in Sodalis glossinidius, which share 28% identity and 45% similarity with ZirT (E-value 5e-47). The gene localized immediately downstream to SG0602 is expected to encode a 371 amino acid protein (SG0603), with a predicted signal sequence. Interestingly, S. glossinidius is an endosymbiont residing intracellularly in tissues of the tsetse flies and utilize for cell invasion a type III secretion system, which is phylogenetically related to the SPI-1 type III secretion system of Salmonella [42]. Although at this point, experimental evidence is still needed, de facto secretion of SG0603 in a SG0602-dependant manner will indicate a broader distribution of the ZirTS-like pathways among the Enterobacteriaceae.
Despite the fact that ZirS is being secreted into the extracellular environment, in mixed infection experiments, secretion of ZirS from a wild-type strain did not seem to complement the phenotype of a co-infecting zirS mutant. This observation may imply that ZirS exerts its effect only topically or has a short in vivo half-life.
Another interesting feature of ZirTS is its location on a horizontally acquired genomic island known as GEI 1664/1678 [17]. The acquisition of genomic islands by horizontal gene transfer enables bacteria to rapidly gain complex functions from other species and are crucial for the interaction of S. Typhimurium with eukaryotic host cells. From an evolutionary and ecological standpoint, infections caused by microbial pathogens that have sustained a long-standing association with their hosts are most often self-limiting or go unnoticed. Salmonella enterica, as an example of such a pathogen, has maintained a coexistence with vertebrate hosts for millions of years and evolved extremely sophisticated mechanisms to engage vertebrate hosts. When examined at the cellular and molecular levels, this functional interface reveals an impressive array of bacterial determinants designed to manipulate the host immune system, to sense the host environment or to modulate a variety of cellular processes. The overall picture emerging from close examination is perhaps one of balance, self restraint and sophistication rather than one of uncontrolled hostility towards its host [43]. Most studies in bacterial pathogenesis are directed towards finding genes that promote virulence in the host. Nonetheless, recent studies have demonstrated that indeed, a fine balance during host infection is kept due to the function of a subset of Salmonella genes known as ‘antivirulence genes' [44]. Deletion mutations of these genes have led to an overgrowth phenotype in the mouse model. Mouslim and colleagues have shown that a PhoPQ regulated gene, pcgL, is involved in stimulating the host immune system to prevent bacterial overgrowth in mouse organs [45]. Another example of an antivirulence gene is sciS. A sciS mutant strain has been shown to display increased intracellular numbers in J774 macrophages and hypervirulence in mice, when administered intragastrically [46]. A null mutation in a Gifsy-2 phage harbored gene, grvA, increased virulence as measured by competitive index experiments in mouse spleens and small intestine [47]. Cumulatively, these studies contribute to a recurring theme in pathogenesis emphasizing the importance of pathogens to limit their effects upon the cells they infect in order to achieve a balance with their host. These examples prove that it is possible for an inactivated gene to lead to an increase in bacterial numbers in host tissues. Increased bacterial loads in the murine host would likely lead to more rapid sepsis and toxic shock, thus increasing lethality and decreasing transmission of the pathogen. The hypervirulence of ΔzirTS observed in the mouse model during oral infection and its induction in the fecal pellets at an early stage of the infection are consistent with this concept. As with all Salmonella species, S. Typhimurium is primarily transmitted through the fecal-oral route. Infected animals excrete Salmonella in the feces, which will then gain access to an uninfected host, starting a new infection cycle. Supported by induction of zirT in the small intestine and its constant induction in feces, we hypothesize that ZirTS play a role as a ‘virulence modulator’ in the early stages of infection. We propose that ZirTS contribute to a host-pathogen balance after the transmission from an infected to naïve host. The role ZirTS play during an early stage of the infection may be in prevention of premature host death and, perhaps, demonstrate at the molecular level, what was understood by MacFarlane Burnet almost 70 years ago that “there is little point in a microorganism destroying its host in a spectacular fashion if this leaves it with no prospect of being ferried to other vulnerable hosts” [48].
Bacterial strains and plasmids used in this study are listed in Table 1. S. Typhimurium SL1344 was used as the wild-type strain, and all mutants used in this study were isogenic derivatives of SL1344. Bacterial liquid cultures were maintained in Luria-Bertani (LB) broth or M9-glucose minimal medium supplemented with 0.0021% (w/v) histidine (since SL1344 is a histidine auxotroph). The appropriate antibiotics were used at the following concentrations: chloramphenicol, 25 µg ml−1; kanamycin, 50 µg ml−1; ampicillin, 100 µg ml−1; and streptomycin, 100 µg ml−1. The metal chelators diethylenetriaminepentaacetic acid (DTPA), N,N′-ethylenediaminediacetic acid (EDDA), and NN′N′-tetrakis(2-pyridylmethyl)ethylene diamine (TPEN) were added to LB in the indicated concentrations.
S. Typhimurium OG2006 carrying an in-frame deletion (amino acid 9–270) of zirS, and OG2007 harboring a Kan cassette in zirT, were generated by allelic exchange using the counter-selectable suicide vector pRE112 [49] containing the levansucrase-encoding sacB gene [50]. pRE112 based plasmids were transformed into E. coli DH5αλpir and then electroporated into E. coli SM10λpir [51] that was used as the conjugative donor strain to S. Typhimurium SL1344. Streptomycin/chloramphenicol-resistance merodiploid colonies were grown for 4 h in LB broth without antibiotic selection, diluted and then plated onto agar containing 1% (w/v) tryptone, 0.5% (w/v) yeast extract, 5% (w/v) sucrose and incubated at 30°C. Sucrose-resistant colonies were selected, and the presence of the constructed mutation was confirmed by PCR. An SL1344 oxyR mutant strain was constructed by P22 transduction from TA4101 (S. Typhimurium LT2). The growth of these constructed strains was indistinguishable from the parental strain while growing in liquid culture in vitro.
Primers used in the study are listed in Table 2. Two-hemagglutinin (2HA) tagged version of ZirS was constructed using the primers OG-61 and OG-63. The resulting PCR product containing the intact sequence of zirT following by zirS (without the stop codon) was cloned into the vector pOG-WSHA after digest with SacI and XbaI. The resulting plasmid harbors a C-terminal fusion of the 2HA tag with ZirS (pOG-zirTS-HA). The primers OG-62 and OG-63 were used to amplify a PCR fragment (containing zirS and ∼1-kb upstream to zirS) that was cloned using SacI and XbaI into pOG-WSHA to generate pOG-zirS-HA. To construct a 2HA tagged version of ZirT we used the primers OG-61 and OG-158 to amplify a PCR fragment that was cloned using SacI and XbaI into pOG-WSHA resulting in pOG-zirT-HA.
A reporter gene construct containing a translational fusion between zirTS and a promoterless lacZ gene was generated using the primers OG-89 and OG-92. The resulted PCR product was cloned into pCR-Blunt (Invitrogen), digested with EcoRI and SmaI and subcloned into pMC1403 to generate pOG-zirTS::lacZ.
Reporter-gene fusion of zirT with the luxCDABE operon from Photorhabdus luminescens was generated by PCR amplification using the primers OG-186 and OG-187. The resulted product was cloned into pCR-Blunt (Invitrogen), digested with XhoI and BamHI and cloned into pSC26 [37] to generate pOG-zirT::lux.
To examine the expression and secretion of ZirS and ZirT, culture supernatant was filtered through a 0.2 µm pore-size filter membrane, concentrated by precipitation with 10% (vol/vol) trichloroacetic acid (TCA), and washed with acetone. The secreted protein fraction and the corresponding bacterial cell pellets were resuspended in 1× sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) sample buffer and subjected to Western blot analysis using the appropriate primary antibodies: rat monoclonal anti-HA (α-HA; 1∶2,000; Roche Applied Science) or mouse α-DnaK (1∶2,000; Stressgen Biotechnologies). Rabbit polyclonal antibodies against subunit beta of E. coli ATP synthase and OmpA were generous gifts from Gabriele Deckers-Hebestreit and Francisco Garcia-del Portillo, respectively, and were both used in a 1∶10,000 dilution. Goat α-rat, mouse, or rabbit immunoglobulin G conjugated to horseradish peroxidase were used as a secondary antibodies (1∶7,500) followed by detection with ECL reagents (Amersham Pharmacia).
β-galactosidase assays were performed according to [52]. The assays were performed with 100 µl of culture, and the substrate for β-galactosidase hydrolysis was o-nitrophenyl-β-D-galactopyranoside (ONPG, Sigma). The background expression of the vector (pMC1403) was 2.76±0.5 and 1.21±0.14 Miller Units (M.U.) when cultures were grown in LB and M9 minimal medium, respectively.
Salmonella RNA was extracted from mid-exponential phase cultures using the Qiagen RNAprotect Bacteria Reagent and the RNeasy mini kit according to the manufacture instructions, including an on-column DNase digest using the RNase-free DNase set (Qiagen). The quantity and quality of the extracted RNA were determined by a ND-1000 spectrophotometer (NanoDrop Technologies). To diminish any genomic DNA contamination, RNA was secondly treated with an RNase-free DNase I (Invitrogen). 0.5 µg of DNase I-treated RNA was subjected to a first strand cDNA synthesis using the QuantiTect Reverse Transcription Kit (Qiagen). Real-time PCR reactions were performed in an Applied Biosystems 7500 Fast Real-time PCR System. Each reaction was carried out in a total volume of 10 µl on a 96-well optical reaction plate (Applied Biosystems) containing 5 µl FastStart Universal SYBR Green Master (ROX) mix (Roche Applied Science); 1 µl cDNA; and two gene-specific primers in a final concentration of 0.3 µM each. Real-time cycling conditions were as follows: 50°C for 2 min; 95°C for 10 min; and 40 cycles of 95°C for 15 s, 60°C for 1 min. No-template and no reverse-transcriptase controls were included for each primers set and template. Melting curve analysis verified that each reaction contained a single PCR product. Reported gene expression levels were normalized to transcripts of rpoD, a housekeeping gene that serves as an internal control. Gene-specific primers were designed using PRIMER3 software (http://primer3.sourceforge.net/), are listed in Table 2, and correspond to the following genes: rpoD, OG-220 and OG-221; zirT, OG-216 and OG-217, OG-216 and OG-229; zirS, OG-212 and OG-215, OG-228 and OG230.
S. Typhimurium SL1344 strains expressing ZirT-HA or ZirTS-HA were grown for 3 h in LB to late log phase (O.D.600∼1.0). Cells were harvested at 5,000 g for 10 minutes at 4°C, and washed with ice-cold phosphate buffered saline (PBS). All the following steps were performed at 4°C. Cells were resuspended in 1 ml of cold lysis buffer [50mM Tris (pH 8.0), 20% (w/v) sucrose, protease inhibitor cocktail (Roche Applied Science), and lysozyme (100 µg/ml)] and incubated on ice for 1 h to generate spheroplasts. MgSO4 was added to final a concentration of 20 mM and spheroplasts were collected by centrifugation for 10 min at 5,000 g. The supernatant containing the periplasmic fraction was recovered and the pellet containing the cytoplasm and the membranes fractions was resuspended in 1 ml cold sonication buffer [50 mM Tris (pH 8.0), 20 mM MgSO4, RNase A (10 µg/ml), DNase I (5 µg/ml), and protease inhibitor cocktail] and lysed by sonication. Unlysed bacteria were removed by low-speed centrifugation at 5,000 g for 10 min. The supernatant was recovered and subjected to ultracentrifugation for 1 h at 100,000 g (in a TLA 100.3 fixed angle rotor in Beckman TL100 ultracentrifuge) to pellet the membrane fractions. The supernatant represented the cytoplasmic fraction was recovered and the membrane pellet was washed in cold sonication buffer, repelleted for 30 min at 100,000 g and resuspended. This fraction represented the total membranes fraction.
S. Typhimurium membranes were prepared from 100 ml cultures expressing ZirT-HA that were grown to late log phase (O.D.600∼1.0) in LB. Cells were harvested at 5,000 g for 10 minutes at 4°C, and washed with 20 ml of ice-cold PBS. All steps were performed at 4°C afterwards. Cells were resuspended in 5 ml of cold PBS containing protease inhibitor cocktail and passed twice through a French Pressure cell at 10000 psi. Debris was removed by centrifugation at 5,000 g for 10 min and the clarified cell extract was centrifuged for 1 h at 100,000 g (30,000 rpm in a SW41Ti rotor; Beckman). Membrane pellets were resuspended in 250 µl of PBS by repeated passage through a syringe equipped with a 25 gauge needle. To separate inner and outer membranes, 200 µl of membranes were layered on top of a discontinuous sucrose gradient composed of 0.5 ml of 60%, 1 ml of 55%, 2.4 ml of 50%, 2.5 ml of 45%, 2.4 ml of 40%, 1.4 ml of 35%, and 0.8 ml of 30% sucrose in PBS (w/v) and centrifuged for 16 hours at 100,000 g in a Beckman SW41Ti rotor. Membrane fractions (800 µl) were recovered from the gradient by using a 20 gauge needle from the bottom of the gradient. Fractions aliquots were separated on a SDS-10% Polyacrylamide gel. The presence of OmpA and the β-subunit of ATPase in the different fractions were used as markers for the analysis of the inner and outer membranes using Western-blot with the respective antibodies.
129X1/SvJ female mice were purchased from the Jackson Laboratories. A breeding colony of inbred strain of 129Sv/J (Nramp1+/+) mice and isogenic Nramp1-deficient (Nramp1−/−) strain have been previously described [53] and are maintained at the University of British Columbia Animal Facility. All the mice were kept in sterilized filter-topped cages and given food and water ad libitum. Experiments were carried out under specific-pathogen-free conditions according to the standard animal care guidelines and protocols of the UBC Animal Care Committee and the Canadian Council on Use of Laboratory Animals. For competitive index (CI) infection experiments, 6–7 weeks old female mice were infected orally or i.p. with mixed inoculum containing a marked wild-type strain resistant to chloramphenicol (ushA::cat, NB24) and a ΔzirS or ΔzirT mutant strain (OG2006 and OG2007, respectively). For oral infection, mice were administered with 1×106 cfu in 0.1 ml of infection buffer (0.1 M HEPES pH 8.0, 0.9% NaCl) and sacrificed after 6 days. For i.p. infection, 2×104 cfu were injected in 0.2 ml PBS and mice were sacrificed after 3 days. Spleen and liver were homogenized in cold PBS, diluted and plated on LB plates containing streptomycin for determination of total Salmonella cfu. Colonies were replica-plated under chloramphenicol selection for enumeration of SL1344 ushA::cat. The competitive index was calculated as [mutant/wild-type]output/[mutant/wild-type]input. CI experiments, in which mice were co-infected with the chromosomally marked strain (SL1344 ushA::cat) and an unmarked SL1344 strain, or strains harboring mutation in non-virulent genes, demonstrated a CI value of 1, indicating equal virulence capability in mice [54],[55].
Wild-type Salmonella Typhimurium harboring pOG-zirT::lux or pSIG70-16, carrying the lux operon under an RpoD dependent promoter [56] as an expression control were grown for 16 h in LB+Kan at 37 °C. Female C57BL/6 mice were orally infected with 2×107 cfu of the reporter strains in 0.1 ml of PBS. At 3 days p.i. mice were anaesthetized with 2% isofluorane carried in 2% O2 and imaged using IVIS 100 (Xenogen Imaging Technologies). Greyscale reference images taken under low illumination were collected and overlaid with images capturing the emission of photons from the lux-expressing bioluminescent S. Typhimurium using Living Image Software version 2.50 (Xenogen). To determine the total numbers of colonizing Salmonella (cfu), the spleen, liver, MLN, ileum, caecum and colon were homogenized in PBS using a high-speed mixer mill (MM301; Retsch), diluted and spread plated on LB agar supplemented with streptomycin.
Data of the β-galactosidase assays are expressed as mean±standard deviation. The statistical significance between two mean values was calculated by the unpaired t-test with two-tailed P value. The statistical significance of the measured mean fold-change by the qRT-PCR was evaluated by t-test against hypothetical value of 1 with two-tailed P value. Data of the CI experiments in mice are expressed as geometrical mean. The statistical significance was calculated by the Wilcoxon Signed Rank Test, against theoretical median of 1 with two-tailed P value. P<0.05 was considered to be statistically significant. |
10.1371/journal.pcbi.1005882 | Feature reliability determines specificity and transfer of perceptual learning in orientation search | Training can modify the visual system to produce a substantial improvement on perceptual tasks and therefore has applications for treating visual deficits. Visual perceptual learning (VPL) is often specific to the trained feature, which gives insight into processes underlying brain plasticity, but limits VPL’s effectiveness in rehabilitation. Under what circumstances VPL transfers to untrained stimuli is poorly understood. Here we report a qualitatively new phenomenon: intrinsic variation in the representation of features determines the transfer of VPL. Orientations around cardinal are represented more reliably than orientations around oblique in V1, which has been linked to behavioral consequences such as visual search asymmetries. We studied VPL for visual search of near-cardinal or oblique targets among distractors of the other orientation while controlling for other display and task attributes, including task precision, task difficulty, and stimulus exposure. Learning was the same in all training conditions; however, transfer depended on the orientation of the target, with full transfer of learning from near-cardinal to oblique targets but not the reverse. To evaluate the idea that representational reliability was the key difference between the orientations in determining VPL transfer, we created a model that combined orientation-dependent reliability, improvement of reliability with learning, and an optimal search strategy. Modeling suggested that not only search asymmetries but also the asymmetric transfer of VPL depended on preexisting differences between the reliability of near-cardinal and oblique representations. Transfer asymmetries in model behavior also depended on having different learning rates for targets and distractors, such that greater learning for low-reliability distractors facilitated transfer. These findings suggest that training on sensory features with intrinsically low reliability may maximize the generalizability of learning in complex visual environments.
| Training can modify the visual system to produce improvements on perceptual tasks (visual perceptual learning), which is associated with adult brain plasticity. Visual perceptual learning has important clinical applications: it improves the vision of adults with visual deficits, e.g. amblyopia and cortical blindness, and even presbyopia (aging eye). A critical issue in visual perceptual learning is its specificity to the trained stimulus. Specificity gives insight into the processes underling experience-dependent plasticity but can be an obstacle in the development of efficient rehabilitation protocols. Under what circumstances visual perceptual learning transfers to untrained stimuli is poorly understood. Here we report a qualitatively new phenomenon: specificity in visual search depends on intrinsic variations in the reliability of feature representations; e.g., vertically oriented lines are represented in V1 with greater reliability than tilted lines. Our data and computational model suggest that training on sensory features with intrinsically low reliability can maximize the generalizability of learning, particularly in complex natural environments in which task performance is limited by low-reliability features. Our study has possible implications for the development of efficient clinical applications of perceptual learning.
| Training in fundamental visual perceptual tasks can lead to substantial improvement, a phenomenon known as Visual Perceptual Learning (VPL), which is associated with adult brain plasticity. VPL has powerful real-word applications [1–3] including improving the vision of adults with cortical blindness [4], amblyopia [5–7] and presbyopia [8].
VPL is often specific to the trained feature and location (reviewed by [9]). From a theoretical point of view, specificity can provide important insight into the neuronal mechanisms that underlie VPL. For example, specificity has been taken to imply plasticity in early-stage visual processing (e.g., [10,11]). However, from a practical or clinical viewpoint, specificity can be a major obstacle in the development of effective training protocols, and it is therefore critical to understand the factors that determine VPL specificity and the conditions that lead to transfer.
For complete transfer to occur, the visual system needs to apply learning for one stimulus to another stimulus. The ability to generalize improvements across stimuli may be most likely when the representation of the stimuli is intrinsically similar. However, the visual system has intrinsic variations in its representation of different feature values. In particular, the reliability with which different feature values are represented can vary considerably within a feature dimension. For example, the reliability of orientation representation in V1 strongly depends on the orientation value. Cells responding to orientations around cardinal are larger in number and have smaller response variability compared to cells responding to orientations around oblique [12,13]. In human V1-V3, sensory uncertainty estimated from the fMRI BOLD signal is higher near oblique orientations than near cardinal orientations, which correlates with orientation estimation behavior [14]. These studies show a gradual variation in representational reliability as a function of orientation, with higher reliability for orientations closer to cardinal (especially horizontal) and lower reliability for more oblique orientations.
These intrinsic differences have been linked to substantial behavioral effects unrelated to learning. They explain the advantage that observers have in discriminating orientations around cardinal compared to around oblique [13–16]: the oblique effect [17]; and in detecting oblique targets among cardinal [18–20] or near-cardinal [21] distractors over the reverse: orientation search asymmetry [22]. Explanations of search asymmetry propose that oblique distractors have less reliable representations than cardinal distractors and thus hinder target detection more [19,20]. Intrinsic variations in representational reliability are not limited to orientation; for example, stimulus processing also varies across spatial frequency [23]. Thus far, however, no study has directly investigated the effect of these preexisting variations in representational reliability on VPL transfer and specificity.
Investigation of VPL has focused instead on the manipulation of task properties. By varying task difficulty [10,24] and task precision (e.g., orientation difference in a discrimination task; [25]), researchers varied the representational precision required to successfully perform the task, and studied its effect on VPL specificity. However, variability in task demands is distinct from initial variability in the underlying representation and may invoke different learning mechanisms. For example, increased specificity in difficult or high-precision tasks has been attributed to changes in the modulated level of representation in the visual processing hierarchy [10,24], whereas intrinsic differences in representational reliability are present within the same hierarchical level.
Here, we asked whether variations in representational reliability alone can explain VPL and its specificity and transfer, when task properties such as difficulty and precision are the same. Our results show that near-cardinal and oblique orientations not only yield an orientation search asymmetry [18–22] but also show asymmetric transfer of VPL in visual search. Conversely, task difficulty, which was independently manipulated by varying the stimulus onset asynchrony (SOA) between a mask and the search display, did not affect the pattern of transfer. To test the sufficiency of a reliability-based account, we fit a computational model that combines learning-related increases in the reliability of stimulus representations with a Bayesian search strategy based on Ma et al. [26]. This Bayesian search model was well-suited to test our hypothesis, because it explicitly represents orientation reliability. Using an unchanging optimal decision rule, the model accounts for both search and transfer asymmetry via initial differences in near-cardinal and oblique orientation reliability.
We trained observers in visual search for an odd orientation (Fig 1A & 1B). One group of observers trained with a near-cardinal target and oblique distractors (near-cardinal group) and the other group trained with an oblique target and near-cardinal distractors (oblique group). Near-cardinal was 80° counter-clockwise from vertical and oblique was 50° from vertical. In both groups, the search stimulus color (task irrelevant) was either green or red. Following training, observers completed an orientation test and a control, color test. In the orientation test, only the orientation swapped with respect to training, i.e. the near-cardinal group was tested with the oblique target and the oblique group was tested with the near-cardinal target (Fig 1C). In the color test, only the color of the stimulus was different with respect to training, i.e. observers that trained with red stimuli were tested with green stimuli and vice versa. Comparing these tests controlled for the involvement of high-level cognitive factors during test sessions (e.g. observers are more engaged due to any new aspect of the stimuli). Perceptual sensitivity (d′) and bias (c) were calculated for each SOA in each session. Incorrect trials and trials with reaction time (RT) ≥4 SDs above the observer’s mean (≤0.5% of the trials) were removed from the RT analysis.
To test for learning, we compared the first and the last training days. For all three dependent measurements (sensitivity, bias and RT) we conducted a (2X2X2) three-way mixed design analysis of variance (ANOVA) with training effect (training day 1 vs. 6) and SOA (35, 59, 94 and 129 ms) as within-observers factors and group (near-cardinal vs. oblique training) as a between-observers factor.
To test for the transfer of learning for each of the three dependent measurements, we conducted a (2X2X2) three-way mixed design ANOVA with tests (color test vs. orientation test) and SOA (35, 59, 94 and 129 ms) as within-observers factors and group (near-cardinal vs. oblique group) as a between-observers factor. As Fig 1E reveals, whereas color test performance was very similar to the last day of learning in both groups, orientation test performance was dependent on the group.
Because baseline performance (training day 1) for the near-cardinal condition was lower than for the oblique condition, it may be that during the orientation test (when target and distractor orientations swapped) specificity was inflated by the baseline difference. In order to control for this possibility, we additionally assessed transfer and specificity by comparing performance (d′) in the orientation transfer test from one group with the baseline performance (training day 1) and trained performance (training day 6) of the other group, such that the orientation condition was the same within each comparison (Fig 2B). First we tested whether transfer performance is higher than baseline, which would indicate that at least some learning partially transferred to the untrained orientation. Two independent sample one-tailed t-tests revealed significant transfer both to near-cardinal and to oblique orientations, t(8) = 3.08, p = 0.007, Cohen’s d = 1.94, t(8) = 2.01, p = 0.039, Cohen’s d = 1.28, respectively. Next we tested whether transfer performance is different than trained performance; a difference would indicate specificity, while no difference would indicate full transfer of learning to the untrained orientation. Two independent sample t-tests revealed significant partial specificity following oblique orientation training, t(8) = 2.96, p = 0.018, Cohen’s d = 2.05, but not following near-cardinal orientation training, t<1. The same results were obtained when a nonparametric test was used (S1 Table). Thus, learning only partly transferred to the near-cardinal orientation but fully transferred to the oblique orientation.
Because we found that VPL specificity and transfer depended on the trained orientation–despite equated task difficulty and task precision–we hypothesized that differences in the representational reliability of near-cardinal and oblique orientations may lead to both search and VPL transfer asymmetries. To investigate this possibility, we used computational modeling. We developed a model that consists of two parts: optimal orientation search [26,28] and reliability improvement over the course of learning. The goal was to determine whether orientation reliability and its improvement with learning could explain the behavioral data.
We compared four models to test different hypotheses about the role of orientation reliability in learning and transfer in the orientation search task. We tested whether initial reliability differences between near-cardinal and oblique orientations alone (Reliability model), different learning rates for targets and distractors alone (Learning model), both of these factors together (Reliability-and-Learning model), or these factors with independent learning rates for the two groups (Reliability-Learning-Group model) best accounted for the data. Detailed descriptions of the models can be found in the Methods, and all model fits are shown in S1 Fig.
Model comparison using the AICc metric indicated that initial reliability differences between near-cardinal and oblique orientations were critical to explain the data. The Reliability model (three parameters, AICc = 10.61) and the Reliability-and-Learning model (four parameters, AICc = 13.00) outperformed the Learning model (three parameters, AICc = 21.05) and the Reliability-Learning-Group model (six parameters, AICc = 24.56). When we compared cross-validated r2, the Reliability-and-Learning model fit the data better than the Reliability model. For the Reliability-and-Learning model, cross-validated r2 was 0.81 (SD 0.09), falling within the noise ceiling (lower and upper bounds, [0.75 0.84]), Model performance was therefore as good as possible given the noise in the data. For the Reliability model, cross-validated r2 was 0.70 (SD 0.24), falling below the noise ceiling.
To determine whether transfer and specificity in the two best models could be predicted based only on the learning phase, we fit the models to the training days only and predicted the transfer test performance for each group. For the Reliability-and-Learning model, the predicted orientation test performance was similar to the observed performance, namely, transfer in the near-cardinal group and specificity in the oblique group (Fig 3A, stars). The Reliability model predicted more transfer in the oblique group than was observed in the data (Fig 3A, plus signs), similar to its fit to all data points (S1 Fig). The pattern of transfer and specificity therefore did not depend on including the test session data when fitting the model, and the Reliability-and-Learning model better explained transfer behavior.
The Reliability-and-Learning model, then, captured the three key features of the data: 1) the search asymmetry at baseline, 2) the performance improvement with learning, and 3) the orientation dependence of VPL specificity and transfer (Fig 3A). Learning in the oblique group maintained the difference in reliability between the near-cardinal and oblique orientations, thereby maintaining the search asymmetry present at baseline and preventing full transfer. Conversely, learning in the near-cardinal group decreased the reliability difference between orientations, effectively overcoming the search asymmetry and allowing similar near-cardinal and oblique performance by the end of training.
Fig 3B shows Reliability-and-Learning model estimates of near-cardinal and oblique reliability as a function of training session for each group. The model estimated greater sensory uncertainty (lower reliability) for the oblique than for the near-cardinal orientation, consistent with physiological and behavioral findings [12,13]. For the best-fitting parameter estimates, the distractor learning rate was 0.65 and the target learning rate was 0.24.
Existing models of VPL predict the same level of specificity across the same levels of task-difficulty [24], task precision [25,29] and feature exposure during training [30]. The demonstration that a mere difference in the trained feature value, near-cardinal vs. oblique orientation, determined VPL specificity challenges these views. Supported by computational modeling, we suggest that intrinsic differences in the representational reliabilities of near-cardinal and oblique orientations governed VPL specificity and transfer in orientation search.
Our design enabled us to control for the involvement of task-related factors and to assess the effect of representational reliability per se. In both groups the equal orientation difference between targets and distractors (30°), equated performance controlled by SOA, and identical exposure to the transfer feature insured independence from task precision [25], difficulty [24], and feature exposure [30], respectively. Our analyses confirmed that both learning rate and magnitude were equal for the two groups. In addition, our results cannot be explained in terms of differences in number of difficult trials during training. A larger number of difficult trials during training has been found to increase specificity [31]. This relationship would predict a result opposite to ours: specificity in the near-cardinal group, which was more difficult on average (across SOAs). Thus, stimulus-related properties, rather than task, determined specificity here.
The dependence of transfer on the specific orientation value has implications for the investigation and interpretation of VPL transfer and specificity using oriented stimuli. Indeed such stimuli have been commonly used to investigate VPL, including in orientation discrimination tasks (e.g., [25,30,32–34]), visual search (e.g., [10,30,35,36]) and texture discrimination tasks (e.g., [37–41]). Some VPL studies have varied orientation values to manipulate task properties, such as task difficulty, and then linked those task properties to the resulting feature specificity (e.g., [10,30,37]). Our study suggests that orientation differences alone can affect the pattern of feature specificity and transfer and therefore should be controlled, particularly in displays with more than one orientation.
Researchers have inferred the site of the underlying plasticity in VPL based on specificity and transfer results (reviewed by [42]). Specificity and transfer have been taken to indicate learning in early and late visual areas, respectively [10,11,24,30]. Here we show that preexisting variation in representational reliability, which can occur within the same level of processing, can determine VPL transfer. Our findings, therefore, suggest that specificity and transfer are not always appropriate diagnostic tools for the level of VPL plasticity.
Our model combined orientation-dependent reliability, improvement of reliability with learning, and an optimal search strategy. We based the search strategy on the optimal visual search model by Ma et al. [26], because that model provides a parsimonious explanation of orientation search with minimal parameters. We found that a single change to the model–letting reliability depend on orientation–captured orientation search asymmetry prior to learning. According to the model, the lower reliability of oblique compared to near-cardinal stimuli leads to more uncertainty during the local decision regarding the identity of an item. The disrupting effect of this uncertainty on visual search performance is larger with oblique distractors (near-cardinal target) than with near-cardinal distractors (oblique target), simply because there are many distractors but only one possible target in any given display.
The improvement of reliability across training days captured the behavioral pattern of both learning and transfer. Importantly, the model uses the same optimal decision rule throughout training and during the transfer tests. Search asymmetry and learning, therefore, could be attributed to variation in sensory reliability only, rather than changes in decision strategy and rule based learning [30].
Comparing alternative versions of the model allowed us to determine which factors were critical to explain the behavioral data. Preexisting differences in reliability were essential–a model without this component failed to fit the data–but independent learning for targets and distractors also improved model performance, particularly in capturing transfer behavior. This result is consistent with a previous study that found independent target and distractor learning in an orientation search task [43,44]. Our learning rate estimates correspond well to that study’s finding of about twice as much learning for distractors as for targets [43]. It is therefore the combination of preexisting reliability differences and greater learning for distractors than targets that best explained behavior, in this family of models. Specifically, greater learning for the initially low-reliability oblique distractors eliminated the search asymmetry and enabled full transfer for the near-cardinal group.
Our model follows the account that differences between the reliabilities of the cardinal (or near-cardinal) and oblique representations cause orientation search asymmetry [18–20]. A key component of these accounts is the ratio of target signal to background noise, which depends on the target and distractor identities [18,19,45]. Alternative accounts have also been proposed. One influential theory explains visual search asymmetries by considering a map of feature dimensions and their interactions [46]. This theory suggests that targets with larger feature values (e.g. more oriented, i.e. oblique) are inherently more detectable than targets with smaller values (e.g. less oriented, i.e. cardinal) (e.g., [46,47]). Based on this theory, a neural computational model was developed that explains search asymmetry in terms of a salience map in V1 [48]. However, it is unclear how the elimination of search asymmetry following near-cardinal training could be predicted if search asymmetry arises from inherent feature properties like “more tilted” [46,48,49]. Moreover, no previous model addresses VPL in orientation search.
Analogous to the reliability differences between orientation values represented by our model, neurons responding to oblique orientations have larger tuning curve widths than those responding to near-horizontal orientations in macaque V1 [13], and there is more cortical area tuned to near-cardinal orientations than to oblique orientations in ferret cortex [12]. Higher sensory uncertainty has also been estimated for oblique compared to near-cardinal orientations in human V1-V3 [14]. The Ma et al. [26] model on which the orientation search component of our model is based has been implemented as a biologically plausible neural network model, strengthening the connection between the physiological literature and our current computational results.
Learning was modeled as an increase in the representational reliability of the stimulus orientations. This increase could be implemented either as a reduction of the tuning curve width of V1 or V4 neurons with training [50–53] or as an improvement in readout from the early sensory response [29]. Both mechanisms have been proposed previously for an orientation discrimination task. Our model, therefore, applies VPL principles derived from orientation discrimination tasks to explain VPL for more complex visual search tasks.
Our findings are limited to VPL in orientation search, and more study is required to determine whether they generalize to other stimuli and tasks. Our study also does not rule out alternative models for orientation search asymmetry and VPL in visual search, but it shows that a parsimonious optimal decision rule, preexisting differences in orientation reliability, and reliability learning suffice to explain both search and transfer asymmetry.
For simplicity our model assumes the same representational reliability for all stimulus locations. However, stimulus reliability can vary as function of eccentricity (e.g., [18,23]) and polar angle [54,55]. It will be interesting to test the relation between location-dependent feature reliability and VPL transfer and specificity.
Researchers have sought to understand the perceptual and neuronal processes that underlie VPL by studying how task demands affect VPL specificity. In the present study we control for task while testing the effect of the intrinsic reliability of feature representations on VPL specificity in visual search. We found a striking difference in VPL transfer depending on the orientation of the trained target, which we interpret as an effect of representational reliability. This interpretation is supported by both previous neurophysiological findings and computational modeling of the present data. We conclude that preexisting variation in the reliability of feature representations within a single level of processing may have a critical effect on VPL transfer and specificity, calling into question the logic that the degree of feature specificity can be used to infer the neural level at which VPL occurs, especially for complex visual displays.
A growing body of research demonstrates the potential benefits of VPL in clinical (e.g., [4–8,56,57]) and professional (e.g., [58]) applications. Our study suggests a testable hypothesis: to increase the generalizability of perceptual learning in real-world applications, efficient training protocols should focus training on low-reliability features–oblique orientations and motion directions [59], peripheral spatial locations [60], and so forth–which may limit performance in a variety of natural tasks.
We developed a model that consisted of two parts: optimal search (based on Ma et al. [26]) and reliability improvement over the course of learning. We compared alternative versions of the model to determine which parameters were required to explain, in a single fit, the data from both observer groups, including the initial search asymmetry, performance improvement over the course of training, and the transfer asymmetry.
|
10.1371/journal.pcbi.1001115 | Reactive Oxygen Species Production by Forward and Reverse Electron Fluxes in the Mitochondrial Respiratory Chain | Reactive oxygen species (ROS) produced in the mitochondrial respiratory chain (RC) are primary signals that modulate cellular adaptation to environment, and are also destructive factors that damage cells under the conditions of hypoxia/reoxygenation relevant for various systemic diseases or transplantation. The important role of ROS in cell survival requires detailed investigation of mechanism and determinants of ROS production. To perform such an investigation we extended our rule-based model of complex III in order to account for electron transport in the whole RC coupled to proton translocation, transmembrane electrochemical potential generation, TCA cycle reactions, and substrate transport to mitochondria. It fits respiratory electron fluxes measured in rat brain mitochondria fueled by succinate or pyruvate and malate, and the dynamics of NAD+ reduction by reverse electron transport from succinate through complex I. The fitting of measured characteristics gave an insight into the mechanism of underlying processes governing the formation of free radicals that can transfer an unpaired electron to oxygen-producing superoxide and thus can initiate the generation of ROS. Our analysis revealed an association of ROS production with levels of specific radicals of individual electron transporters and their combinations in species of complexes I and III. It was found that the phenomenon of bistability, revealed previously as a property of complex III, remains valid for the whole RC. The conditions for switching to a state with a high content of free radicals in complex III were predicted based on theoretical analysis and were confirmed experimentally. These findings provide a new insight into the mechanisms of ROS production in RC.
| Respiration at the level of mitochondria is considered as delivery of electrons and protons from NADH or succinate to oxygen through a set of transporters constituting the respiratory chain (RC). Mitochondrial respiration, dealing with transfer of unpaired electrons, may produce reactive oxygen species (ROS) such as O2− and subsequently H2O2 as side products. ROS are chemically very active and can cause oxidative damage to cellular components. The production of ROS, normally low, can increase under stress to the levels incompatible with cell survival; thus, understanding the ways of ROS production in the RC represents a vital task in research. We used mathematical modeling to analyze experiments with isolated brain mitochondria aimed to study relations between electron transport and ROS production. Elsewhere we reported that mitochondrial complex III can operate in two distinct steady states at the same microenvironmental conditions, producing either low or high levels of ROS. Here, this property of bistability was confirmed for the whole RC. The associations between measured ROS production and computed individual free radical levels in complexes I and III were established. The discovered phenomenon of bistability is important as a basis for new strategies in organ transplantation and therapy.
| Reactive oxygen species (ROS) are side products of electron transport in the mitochondrial respiratory chain, the principal component of energy transformation in mitochondria. ROS generation starts with the formation of a superoxide radical (O2−) as a result of interaction between molecular oxygen and free radicals, e.g. semiquinone (Q−): O2+Q−→O2−+Q [1]. This extremely active compound can be deactivated in cells, mainly through superoxide dismutase [2]. However, H2O2 formed in this process can interact with various intracellular compounds to produce ROS. ROS production serves as a metabolic signal [3]–[5]. However, when released in excess under certain stress conditions such as hypoxia/reoxygenation, ROS can also directly damage cells [6]. This destructive function of the electron transport chain represents the main problem in organ transplantation [7] and in many systemic diseases, as diverse as Parkinson disease [8] and diabetes [9]. The problem can be so great that in some organisms disruption of the electron transport chain can be a positive factor in increasing lifetime [10].
Although electron transport and coupled ROS production have been the focus of intensive research, important details are still not understood. There is currently debate regarding the relative contribution of various sites of the respiratory chain to overall ROS production [11], [12] and the factors that may alter this contribution [13]. The use of specific inhibitors can localize the sites of ROS production, but their contribution under normal and stress conditions without inhibitors in vivo is not clear. It is generally accepted that electron transport from succinate through complex I to NAD+, the phenomenon known as reverse electron transport [14], is important for respiration and ROS production [15], [16]. However, the mechanism of ROS production as a result of electron transfer upstream in the respiratory chain is not understood. Some details of the general mechanism of electron transport, such as the interaction of complex I with quinones that results in translocation of four protons through the membrane and reduction of one ubiquinone molecule per two electrons transported, remain the subject of discussions [17], [18]. Answering these questions will help in understanding the mechanisms of electron transport and coupled ROS production, and will be useful for advances in transplantology and therapy.
The solution to such problems requires not only improvements in experimental techniques and new experiments, but also modification of methods for theoretical analysis. Specifically, kinetic modeling, which is an efficient method for investigating complex systems, still needs to be adopted for the mitochondrial respiratory chain. In fact, kinetic modeling in its classical form has been used for analysis of mitochondrial respiration. However, even the most detailed models [19] could consider only simplified scenarios. Huge number of differential equations is necessary to describe the behavior of respiratory complexes, so an automated procedure is required for their construction. Previously we developed a rule-based methodology for the automated construction of large systems of differential equations for analysis of 13C isotope tracing experiments in metabolic flux analysis [20]–[22]. We extended this methodology to the mathematical description of multienzyme complexes, specifically mitochondrial respiratory complex III based on a Q-cycle mechanism [23]. A detailed description of complex III operation revealed that in a certain range of parameters complex III has the property of bistability, where two different steady states exist for the same parameters and the system can reach one or the other, depending on its initial state. Perturbations, such as fluctuations in succinate concentrations or temporal hypoxia, can switch the system from low to high ROS producing steady state. Such a switch explains the damaging increase in ROS production on reoxygenation after hypoxia.
The prediction of bistability for the mitochondrial respiratory chain was based on analysis of the Q-cycle mechanism for complex III. The contribution of other parts of the respiratory chain and linked processes that provide substrates must affect the properties of the respiratory chain. To study mitochondrial respiration as a whole, we extended the model of complex III [23]. The extended model includes the following elements: a detailed mathematical description of complex I; the stoichiometry of electron transport and proton translocation by the respiratory chain; the transmembrane potential; proton leak; oxidative phosphorylation; the TCA cycle that produces NADH and succinate as substrates for complexes I and III; and the transport of TCA cycle metabolites. The objective of this extension to the whole respiratory chain and linked processes was to create a tool for analysis of the basic behavior of the respiratory chain, in particular under conditions defining different fluxes in the forward and reverse directions. The ultimate aim was to reveal characteristics that have not been measured, such as the content of various free radicals, and thus to provide an insight into the relationship between states of the respiratory chain operation and ROS production.
Figure 1 shows the components of the respiratory chain connected in the extended model. Respiratory complexes I to III are components of the electron transport chain connected through ubiquinone. Complex III is linked to complex IV through reduction/oxidation of cytochrome c. NADH, which is a substrate for complex I, is produced in the TCA cycle. Since the total concentration of NAD+ and NADH is conserved, NADH consumption, which fuels electron transport in the respiratory chain, defines the levels of NAD+, which is a substrate for several reactions in the TCA cycle. In this way, the extended model links electron transport with central energy metabolism, in particular with the reactions of the TCA cycle.
As described in the Methods section, the model of the respiratory chain and linked substrate transport and TCA cycle reactions contains 51 parameters. Out of 22 parameters of complex III, six ratios for forward and reverse rate constants were expressed through midpoint potentials. The order of magnitude for the rate constants for forward electron transport reactions in complex III can be estimated based on previous studies [19]. A qualitative reproduction of measured triphasic dynamics of cytochrome bH reduction by succinate in isolated cytochrome bc1 complex [24], as described in Text S1 and Figures S1, S2, S3 and S4, provides some restrictions for rate constants for binding/dissociation of complex III with ubiquinone species.
The rates of respiration in the presence of ADP (state 3) or an uncoupler characterize the maximal capacity of the respiratory chain. In the absence of ADP (state 4), the respiration rate is characterized by proton leaks, which must be compensated by respiration. According to our measurements, the respiration rate is 480±40 and 170±30 ng atom O/min/mg protein in the uncoupled and in state 4 in succinate-fueled mitochondria, and 410±30 and 80±20 ng atom O/min/mg protein in mitochondria fueled by pyruvate and malate, respectively.
If mitochondria fueled by succinate do not expend the energy of the transmembrane electrochemical potential on ATP synthesis (state 4), succinate oxidation results in fast reduction of intramitochondrial NAD+. In the presence of rotenone, an inhibitor of electron transport in complex I, NAD+ reduction is characterized by NAD+-dependent reactions of the TCA cycle and in particular the forward respiratory flux resulting from succinate oxidation. In the absence of rotenone, reverse electron transport [14] also participates in NAD+ reduction, which makes the process much faster (Figure 2A). These data define the rate constants for reverse electron transport.
While succinate fuels complex III through succinate dehydrogenase, the oxidation of malate and pyruvate in the TCA cycle fuels complex I by reducing NAD+ to NADH. Respiration under such conditions defines the characteristics of complex I.
To evaluate the model parameters, we used a procedure that simulates all the different types of data listed above for the same set of parameters. The ratio of forward and reverse constants defined by a known midpoint potential or dissociation constant was kept fixed, and the conditions of substrate supply or membrane permeability for protons were changed in accordance with experimental conditions. The procedure fitted all the data by changing the free parameters within the order of magnitude indicated in [19], summarizing and minimizing the deviations in several calculations that simulated measurements. Minimization was performed using a standard stochastic procedure in the global space of parameters as described in Methods.
The best fit reproduces well the dynamics of NAD+ reduction measured in brain mitochondria in the presence and absence of rotenone using the same set of parameters (Figure 2A). The insets in Figure 2A show respiration rates and ΔΨ in the presence and absence of rotenone. These characteristics remain practically the same in both conditions. Without rotenone inhibition reversible electron flow through complex I, which fits the experimental data shown in Figure 2A, is directed to NAD+ reduction (is negative) only during a short period of time (Figure 2B), although ROS are constantly produced for a much longer time under such conditions [15], [25]. Reverse electron flow is believed to induce excessive ROS production, but evidently these two processes are not correlated.
Rotenone essentially changes the dynamics of NADH measured before succinate addition. It is slightly oxidized by the RC in the absence of rotenone, but slowly reduced in its presence. This reduction is a result of oxidation of internal substrates while electron flow through the RC is blocked. We found that the metabolites of TCA cycle cannot be substrates that provide NADH reduction, because oxidation of TCA cycle metabolites results in much faster initial reduction of NADH. If the parameters of TCA reactions are changed to slow down and reproduce the initial dynamics of NADH, maximal respiration rate with pyruvate becomes inconsistent with experimental data (not shown). Rather, slow oxidation of other metabolites, probably aminoacids or lipids, contributes to NADH reduction. The simulation of such slow oxidation did not prevent NADH oxidation in the absence of rotenone, and reproduced NADH reduction in its presence.
While, in the absence of rotenone, succinate induced much faster NADH reduction due to reverse electron transport, the steady state levels are lower than in the presence of rotenone. The steady state levels are defined by NADH production and consumption in respiration. Rotenone blocks the consumption, therefore NADH levels further increase when rotenone is added after succinate. The model parameters were adjusted without considering subsequent NADH increase and the reproduction of this phenomenon validates the model.
The model reproduces measured maximal and state 4 respiratory electron flows for succinate-fueled mitochondria, as well as for mitochondria fueled by pyruvate/malate (Figure 2C). The change in ΔΨ in the same simulations qualitatively corresponds to known changes measured under such conditions (Figure 2D). The parameters for simulations shown in Figure 2 are listed in Table 1 (column indicated as best fit).
These simulations of measured data provide an insight into important hidden characteristics, such as the capacity of ROS production. ROS are produced by the respiratory chain as a consequence of one-electron transfer directly to oxygen from free radicals of electron transporters such as the semiquinone radical (SQ) at the Qo site in complex III [26]–[28] or FMNH [29], SQ bound to complex I [30], or N2 centers [30] in complex I. Simulating the experimental data as presented in Figure 2 the model at the same time simulates the dynamics of these free radicals.
The model describes various states of respiratory complexes formed in the process of electron transport, including those containing free radicals. Such radicals could be responsible for passing unpaired electrons to oxygen thus forming superoxide radicals and other forms of ROS. The contributions of various radicals to ROS production remain unknown; to clarify it we compared measured ROS production and the levels of various free radicals predicted by the model for the same conditions. A similar change in radical content and measured ROS production indicates qualitative accordance between the model and the described process and thus validates the model.
It is generally known that inhibition of reverse electron transport by rotenone decreases ROS production in succinate-fueled brain mitochondria [15], . In our measurements, ROS accumulation was inhibited immediately after rotenone addition (Figure 3A). The model predicts that the SQ content at site Qo in complex III in succinate-fueled mitochondria is practically unchanged by the presence of rotenone (Figure 3B) and this remains valid for simulations with any set of parameters describing the data well. This is the reason for the coincidence of intervals for SQ at Qo for the first two types of simulation shown in Table 1. Thus, the ROS-stimulating role of reverse electron transport and the ROS-inhibitory effect of rotenone cannot be explained at the level of complex III. Apparently, reverse electron flow mainly affects complex I by increasing the concentrations of free radicals able to pass electrons to oxygen.
The model predicts that rotenone essentially decreases initial levels of SQ bound on site Qn (Figure 3C), FMNH (Figure 3D), and the content of reduced N2 centers (inset). After an initial decrease, levels of SQ and FMNH increase, in agreement with the acceleration of ROS production measured after initial inhibition induced by rotenone (Figure 3A). The reason for accumulation of free radicals and acceleration of ROS production is the production of malate from succinate, which then reduces NAD+ in malate dehydrogenase reactions. This supply of substrate for complex I increases ROS production in rotenone inhibited mitochondria. The increase in the rate constant for malate–succinate exchange eliminates a slow increase in free radical content when rotenone is present (dashed curve in Figure 3C and D). It should be noted that the acceleration of ROS accumulation is not always observed experimentally and this agrees with the predicted disappearance of this slow component after acceleration of malate-succinate exchange. Such similarity of experimental and simulated behavior supports the mechanism accepted for its simulation and in this way validates the model.
The fact that the species of complex I can contain more than one radical makes it more difficult to understand the contribution of each site. In particular, the species 1101001 (the positions of digits correspond to Qp-Qp-Qn-Qn-N2-FMN-FMN), which contain SQ and FMNH radicals, slowly accumulate after inhibition by rotenone. This accumulation defines the dynamics of SQ and FMNH, whereas 1101100 defines the fast component in levels of Qn-bound SQ and reduced N2 (inset in Figure 3D). It is possible that only one of coupled radicals makes the major contribution to ROS production, but in this case the levels of other radicals would also correlate with ROS production. On the other hand, radicals situated inside the same species could interact, so that the specie as a whole produce superoxide. In the considered example the behavior of the whole ensemble of radicals in complex I agrees with the observed effect of rotenone, and this validates the model.
Overall, according to the model predictions, rotenone hardly affects SQ levels in complex III, but initially it significantly decreases the levels of free radicals produced in complex I; this is the reason for the decrease in ROS production induced by rotenone in succinate-fueled mitochondria. The model also explains the subsequent increase in ROS production as a result of the formation of malate in rotenone-inhibited mitochondria.
Rotenone induces a large increase in ROS production in pyruvate/malate-fueled mitochondria (Figure 4A). The corresponding simulations show that rotenone greatly increases the levels of FMNH and SQ at Qn site, but decreases the levels of reduced N2 (Figure 4B). Since the changes in N2 disagree with measured ROS production, probably N2 center does not make essential contribution in ROS production under the considered conditions. The same species (1101001) that defined the slow component in the increase in free radicals now change faster and defines the main part of the response to rotenone. Species 1101100 also constitute an essential part of the total radical content, but their levels decrease in response to rotenone in accordance with the decrease of total N2 radical levels.
Stimulation of electron transport by addition of ADP or an uncoupler such as FCCP to succinate-fueled mitochondria results in a decrease in ROS production (Figure 5A). This generally known phenomenon [15] validates the model prediction that the levels of all free radicals decrease when electron transport is stimulated by addition of ADP or an uncoupler.
Mitochondria fueled by pyruvate/malate also produce less ROS when electron transport is stimulated by an uncoupler (Figure 6A). Such measurements also validate the model, which predicts a decrease in the levels of free radicals (Figure 6B).
At high succinate concentrations, brain mitochondria produce much more ROS than those fueled by pyruvate (Figure 7A). The model also predicts higher levels of free radicals in complex III, as well as in complex I, for mitochondria fueled by succinate (Figure 7B).
Thus, the study of associations between measured ROS production and predicted radical levels in RC revealed qualitative consistency of measurements with all types of radicals and therefore validated the model, or showed a way of discrimination between possible sites of ROS production, and even between possible ROS producing species. However, in the latter case, a special, quantitative study is needed, which currently is beyond of the scope of presented study.
It has been predicted that the Q-cycle mechanism of complex III can in principle induce bistable behavior [23]. The whole respiratory chain considered here, with the parameters that fit the experimental data, also has two different steady states for the same parameters. Figure 8A shows that the SQ content at site Qo in complex III could persist at different values, depending on whether the respiratory chain is initially reduced or oxidized. Figure 8B shows how the steady states for free radicals of complexes I and III change with the external succinate concentration for a set of parameters that reproduces the experimental data described above.
With increase in succinate concentration at some point the system switches to the state with the highest levels of semiquinone radicals at Qo site of complex III. The difference here from the similar curve in Figure 7 is that pyruvate is present in addition to succinate. Once the system is switched to the state of highest SQ content at Qo, it remains in this state even if the succinate concentration decreases back to low values. Thus, if the system is initially in an oxidized state, the steady state SQ levels at Qo depend on the succinate concentration, in accordance with the blue curve in Figure 8B. If the system is initially in a reduced state, it remains in this state until succinate concentrations decrease to the micromolar range. Since complex III is directly connected to complex I through a common substrate (ubiquinone), the bistable behavior of complex III induces bistability in complex I. However, when complex III enters the state with high SQ levels at Qo, SQ levels at Qn decrease (Figure 8B), as well as the levels of other free radicals in complex I (not shown). In some range of succinate concentrations total amount of radicals in the two presented steady states can be similar, but this does not necessary means similar ROS production in the two states since the probability of ROS production can be different for various radicals.
Thus, bistable behavior remains valid for the extended model of the RC with proton translocation and transmembrane potential (ΔΨ) generation, and with parameters defined by fitting the experimental data and validated by qualitatively similar predicted and measured ROS production. The model predicts also that a pulse of succinate is associated with decrease of ΔΨ. Such counterintuitive decrease of ΔΨ induced by increase of substrate for respiration is shown in Figure 9A. The value of ΔΨ decrease, induced by the same pulse of succinate, can be different, depending, for instance, on membrane leak, as illustrated by two curves in Figure 9A. Measurements of ΔΨ using safranine O fluorescence revealed that the mean ΔΨ at low succinate (0.2 mM) is greater than at high succinate (2 mM) (Figure 9B), thus validating the paradoxical prediction of the model.
The succinate threshold for a switch to the reduced state depends on the parameters of pyruvate transport and TCA reactions; here we do not investigate the quantitative details with respect to bistability, but emphasize only the qualitative similarity of predicted and measured behavior. With regards to the considered in the previous sections normal “working” steady state, the predicted levels of free radicals are robust with respect to the model parameters, as the next section shows.
The sensitivity of simulations to variations in model parameters is shown in Table S1 for each type of experimental data presented in Figure 2 (dynamics of NAD+ reduction, maximal and state 4 respiratory fluxes). The sensitivity is also listed for simulated levels of free radicals shown in Figure 3. The results indicate that significant changes in some parameters hardly affect the simulations (e.g. kqp_FS). Evidently, the data do not restrict the parameter values and they could not be defined unambiguously. However, changes in these parameters within the range, for which fitting remains good, do not affect the predictions in terms of free radical levels. The parameters shown in red highly affect the simulations. However, it is possible that different combinations of such parameters could fit the measured data equally well because of mutual compensatory changes. In this case, despite the high sensitivity, the parameters can have a wide range of values for which a good fit is obtained. Confidence intervals rather than sensitivity are used to characterize the robustness of parameter determination.
Different sets in the global space of parameters that fit the experimental data could be identified using our stochastic algorithm for minimization of the objective function χ2 (sum of squares of deviations from measured data normalized by standard deviations). The algorithm identified confidence intervals for parameters based on fixed thresholds of χ2 [31].
Table 1 shows the 99% confidence intervals for the free parameters. The ranges for which the values give a good fit to the data are large. Thus, even though the measurements cover various modes of respiratory chain operation, the data do not restrict the parameters sufficiently to define them unambiguously. Various sets over a wide range of parameters can describe the data equally well. However, the situation is different for free radical levels predicted for the simulated experimental conditions. Table 1 lists intervals for predicted free radical levels simulated using the parameters sets that fit the data with χ2 that is below the threshold. The confidence intervals for free radical levels are generally much narrower, so the predicted values are more robust. Although the intervals for SQ at Qo sites in succinate-fueled mitochondria are relatively large, they are clearly almost the same for both conditions (with or without rotenone). This result agrees with data indicating that the SQ content at Qo practically shows no dependence on the presence of rotenone (Figure 3B). The levels of all free radicals in complex I under the conditions for the first two simulations are very robust, despite the high parameter variability. If the parameters give a good fit, the model predicts similar levels of complex I radicals. Although the intervals are relatively large under the third condition (pyruvate/malate supply), it is evident that they are much lower than the intervals for the condition of succinate supply, as well as the levels of radicals in complex III.
To construct a detailed mathematical model that accounts for all redox states formed during electron and proton transport in complexes III and I, we used our rule-based methodology for automated construction of large systems of ODE [23]. This model further extends our methodology previously used to model the distribution of 13C isotopes in central metabolism [20]–[22], development of which occasionally coincided in time with that of similar rule-based methodology for signal transduction pathways [32], [33]. For the study of mitochondrial processes our methodology gives a deep insight into the mechanics of respiration and ROS production. Here, rule-based algorithms for mathematical description of mitochondrial respiration coupled to proton translocation and ΔΨ formation was linked to a classical kinetic model that accounts reactions of the TCA cycle, which provides succinate and NADH as substrates for respiration and substrate transport in mitochondria.
After fixing the ratios of forward and reverse rate constants for electron transport reactions, free parameters were defined by fitting of forward and reverse electron flows measured under various conditions. High variability of parameters with a good fit to experimental data precluded definition of their values. However, the levels of free radicals calculated in the model showed much less variability. Different sets of parameters with a good fit to experimental data define very similar patterns for free radicals formed in complexes I and III. Thus, the analysis gives a valid insight into the mechanism of respiration and ROS production, even without precise evaluation of the model parameters.
A substantial body of experimental data on mitochondrial ROS production cannot be satisfactorily explained within the current experimentally based paradigm. Some of these results were obscure, such as acceleration of succinate-driven ROS production after initial inhibition by rotenone (Figure 3). Others, such as a lower membrane potential in mitochondria fuelled by higher succinate concentration (Figure 9), were even counterintuitive. Calculation for mitochondrial constituents not measurable by current techniques represents a powerful tool for mechanistic explanation of accumulated data and for directing experimental research to test model predictions.
A body of evidence indicate that either FMNH [29], [30], or SQ bound to Qn sites of complex I [30], or reduced N2 centers [30], [34], [35] may be a major contributor to ROS production, depending on the tissue, substrate, energy demand and oxygen tension [36], [37]. The simulations revealed correlations between measured ROS production rates and levels calculated for each type of free radical. In this first step of the study we did not assume any explicit link between any specific radical and ROS, but qualitatively compared all of them, taken separately, with measured ROS production. However, the method, which we use, opens a direction for future studies of quantitative contribution of various radicals of electron transporters, and even specific species of complex I and III, into total ROS production.
The similarity between changes in the ROS production rate and in the levels of specific free radicals validates the model and also provides an insight into the mechanism of ROS production. Rotenone inhibition of ROS production in succinate-fueled mitochondria correlated with the free radicals formed in complex I, but not in complex III. Evidently, under the given conditions, reverse electron transport must contribute to free radical formation in complex I, although the net flux reducing NAD+ through complex I exists for only a very limited period of time.
In accordance with our previous study that revealed bistability for complex III [23], the extended model confirms the existence of two steady states for the same set of parameters. In one of these states (oxidized), mitochondria can develop a maximal rate of respiration, ΔΨ, and a capacity for ATP synthesis. This is the usual working state. In the presence of pyruvate high succinate concentrations can induce a switch of respiration to the reduced steady state, where lack of electron acceptors strongly restricts electron flow. The levels of free radicals in complex III greatly increase in this state, but decrease in complex I, in contrast. The switch to a more reduced state results in ΔΨ decrease. Indeed, we observed a ΔΨ decrease in isolated mitochondria in conjunction with an increase of succinate concentrations in the presence of pyruvate.
Q-cycle mechanism of complex III operation assumes bifurcation of electron flow at Qo site: one electron goes to Rieske center and further to complex IV, and another one reduces cytochrome b. This bifurcation of electron flow underlies the bifurcation between the two steady states. If in some moment the rate of first electron transition to Rieske center is higher than that for the second electron (because cytochrome b is reduced), semiquinones at Qo accumulate, thus preventing Qo liberation, binding and oxidation new molecules of ubiquinol, and thus limiting electron flow. In the case, shown in Figure 9A, greater proton leak resulted in greater transient discrepancy between the two electron flows at the point of bifurcation, which ultimately leaded to more significant inhibition of respiration and deeper descent of ΔΨ.
The decrease of ΔΨ, in the case shown in Figure 9B, is relatively small, however, in living cardiomyocytes a much higher decrease of ΔΨ can be observed, accompanied by high ROS production, and associated with mitochondrial permeability transition (MPT) [38]. Although the presented study does not touch a possible link between bistability and MPT, it puts forward some hypotheses, which verification in future can essentially clarify the mechanism of MPT. There are at least two phenomena, which do not find appropriate explanation in terms of current state of knowledge. First, the increase of permeability in this process does not increase electron flux and proton recirculation, as in case of uncouplers. Second, ROS production is high despite the decrease of ΔΨ. If we assume that the switch into the reduced state precedes MPT, both phenomena would find a natural explanation. This hypothesis, although not proved yet, opens avenues for deeper investigation of the MPT mechanisms.
The presented new insight into mitochondrial respiration was possible due to the application of novel methodology of modeling that allowed a detailed mathematical description of mitochondrial respiration. The phenomenon of bistability, predicted based on this methodology, has a potential to be a basis of new paradigm for the mechanism of ROS production, which will initiate new research with outcome on academic and practical levels.
The file executable in Linux, which runs the simulations, and the C++ code of the program could be downloaded free from http://www.bq.ub.es/bioqint/ros_model/plcb2010.cpp.tar.gz.
The model of complex III described elsewhere [23] was used as a part of the extended model presented here. For each reaction two values, forward (Kf) and reverse (Kr) rate constants, were used as parameters. The order of magnitude of Kf was set based on [19] and then Kr was determined as described in [23] using midpoint electrochemical potentials, which determinations was variable and allowed refinement by fitting the data presented in “Results”. Table 2 summarizes the reactions and values of parameters for complex III that simulate the data.
The overall process catalyzed by complex I is oxidation of NADH coupled with ubiquinone reduction and pumping 4 protons from negative to positive side of the membrane:(I.0)This is a complex process that involves electron transport through a chain of intermediates coupled with proton translocations through inner mitochondrial membrane. The structure and mechanism of catalysis of complex I is reviewed in [39] and the data from this review are used for the construction of model of complex I.
It is assumed that proton translocation is a result of Q reduction (with proton binding) at the negative side and its oxidation (and proton release) at the positive side. If several protons are translocated per one electron, then this electron must pass several cycles of Q reduction and oxidation. Such mechanism, similar to that accepted for complex III, called Q-cycle, was suggested for complex I (see e.g. [17]). We constructed a model based on electron cycling that is in accordance with the measured stoichiometry of proton translocations per one electron passed through the chain.
The initial step of such transport is the oxidation of NADH coupled with the reduction of FMN; further, electrons from FMN pass through a relay of eight different iron-sulfur (Fe-S) containing centers [40], which possibly form a relay for electron transport from FMN to the last Fe-S center N2 (see e.g. the review [40], [41]). The Fe-S centers have similar midpoint potential close to that for FMN (E∼−350 mV) with an exception of N2, which is much more positive (−150 mV, [40]). In this model the relay of Fe-S centers is simplified, so that electrons pass from FMN directly to the N2 center, which can interact with quinones. In this way, two-electron transporter FMN and one-electron transporter N2 form the core of complex I, N2-FMN- FMN (referred as core).
The mechanism of interaction of N2 center with ubiquinone that results in the translocation of four protons from matrix to cytosol and one ubiquinol synthesized is not fully understood. Here we implemented in the model a proposed mechanism, which we consider as a working hypothesis that could be checked by the analysis of model behavior. In this way the model could serve as a tool for checking different possible mechanisms.
According to the EPR data [42], [43] there are two ubiquinone-binding sites; bound ubiquinones possess different EPR characteristics, one of them is fast- and another is slow-relaxing. The former one bound in oxidized form in the proximity of N2, in Qn site, could be reduced by N2 and bind protons taking them from negative (matrix side of the membrane) (indicated as Qn below). The other one, bound in the reduced form in Qp site, situated in the proximity of Qn, can interact with Qn-bound semiquinone releasing protons to the positive (cytosolic) side of the membrane. This interaction of two quinones in fact is in agreement with the idea outlined in [39] that complex I contains a single, but very large, binding domain for its hydrophobic substrate. Binding Qn and Qp gives additional three species of the complex I, Qn-Qn-N2-FMN- FMN, Qp-Qp-N2-FMN- FMN, and Qp-Qp-Qn-Qn-N2-FMN- FMN.
The proposed mechanism of N2-ubiquinone interactions, which we implemented in the model, is shown in Figure 10, and the individual reaction steps are described in the legend. It satisfies the known stoichiometry of proton translocation and ubiquinone reduction (four protons translocated and one ubiquinol synthesized per two electrons taken from NADH).
0. Reduction of oxidized FMN by NADH.
In traditional form this equation is expressed as(I.0.0)Any of the forms of complex I with reduced FMN can receive two electrons from NADH, however, subsequent transitions require the interaction of three centers, N2, Qn and Qp. Therefore effective outcome produces only the reduction of FMN in the specie qnpc with ubiquinone bound at Qn and ubiquinol bound at Qp, which is reflected by binary number 1100000 corresponding to decimal 96. The reduction of FMN results in the production of redox state 1100011 (decimal 99):The forward and reverse reaction rates for this transformation are expressed in accordance with mass action law:(I.0.1)Here, as described for the complex III, “0” designates oxidized and “1” reduced states.
The ratio of rate constants from I.0.1 could be found from the known redox potentials. Equilibrium constant for this reaction as a function of midpoint electrochemical potentials could be found from the condition of equality of electrochemical potentials at equilibrium:(I.0.2)(I.0.3)since , expression (I.0.3) could be rewritten as(I.0.4)(I.0.5)taking into account that the difference between midpoint potentials for NADH (−320 mV) and FMN (−340 mV) [40] is ΔEm = −20 mV.
1. Reduction of the N2 center by FMN (step 1 in Figure 7):(I.1)First electron of FMNH2, which by convention occupied second position from the right in binary representation, passes to N2 converting 0 into 1 in the third position from the right:(I.1.1)The relationship between forward and reverse rate constants could be defined similar to (I.7). For the first transition at equilibrium(I.1.2)since , eq (I.1.2) can be written as(I.1.3)since at pH = 7 Em(FMN−/FMNH2) = −350 mV [43], and Em for N2 iron-sulfur center Em = −150 mV [40].
2. Reduction of Qn by the reduced N2 center (first electron) and by QH2 bound at Qp center (second electron):(I.2a)In binary form:The semiquinone Q−n is very active [17], so it reacts with QH2 bound at p-site:(I.2b)In binary form:This reaction is symmetrical: p-side quinol and n-side semiquinone give p-side semiquinone and n-side quinol. The distance between the two quinone binding sites can be estimated as follows. Fast-relaxing semiquinone (bound to n-side oriented proton well) situated at the distance of ∼12 Å from N2, slow-relaxing semiquinone (bound to p-side oriented proton well) situated at the distance of ∼30 Å from N2 [42]. The distance between the bound quinones could be around 18 Å, which makes possible the interaction between them, taking into account the high energy of electron coming from FMN to Qn-bound quinone through N2 center. The assumption of such interaction fulfills the known stoichiometry of translocation of four protons and overall reduction of one ubiquinone coupled with the transport of two electrons through complex I.
We grouped together these two reactions:(I.2)Overall in this reaction the oxidation of N2 center is coupled with the reduction of Qn.Here is considered that Em for N2 center is −150 mV [40] and Em for ubiqunone one-electron reduction is −45 mV [42].
3. Dissociation of QH2n at n-site, transition of p-site SQp to the n-site and binding of dissociated QH2n at p-site.
In this step the three reactions are combined: dissociation of QH2 formed at n-site, change of position of p-site bound semiquinone, and binding QH2 at p-site. Overall in binary form:(I.3)and the forward and reverse rates are:
4. Second electron (from radical FMN−, which by convention occupied the right position) passes to N2 converting 0 into 1 in the third position from the right:(I.4)
The transition of second electron characterized by the same ΔEm as accepted in (I.1), but it is not related with proton binding or release, therefore the right hand side value of (I.1.3) equals to the ratio kf/kr.
5. Reduction of N2 by FMN in step 4 induces the interaction of n-site semiquinone with p-site quinol resulted in the production of n-site quinol and p-site semiquinone coupled with the translocation of two protons:(I.5)The dissociation of n-site quinol produced and the change of position of p-site semiquinone:(I.6)
6. The reduction of n-site semiquinone by N2 coupled with the binding of two protons:(I.7)In equilibriumSince ,and the ratio of forward and reverse rate constants could be expressed astaking into account that Em for ubiquinol reduction in −63 mV [42] and Em for reduced N2 center oxidation is −150 mV [40].
7. QH2 dissociates, Q binds at n-site and QH2 binds at p-site, overall:(I.8)Above, the ratios of forward and reverse rate constants for the redox reactions of complex I are defined and summarized in Table 3. The particular values were defined from fitting the experimental data presented in “Results” using these ratios as restrictions. In some cases the fitting required different value of midpoint potential. This may indicate the differences in the operation of complex I in situ and under the specific conditions of midpoint potentials determination. Recognizing the importance of this subject we leave its studying for future because it deserves a separate consideration.
Although the mathematical description of complex I and complex III are similar, they differ in the strictness of rules for electron transport and proton translocation. For complex III the transition between two transporters allowed for any states of other transporters. This assumes participation of all 400 redox forms in electron transport. For complex I the rules accepted in the model allow participation in electron transport only several selected form. This illustrates the flexibility of methodology applied.
Proton binding to ubiquinone at the matrix side of the membrane and their dissociation from ubiquinol to the intermembrane space results in the translocation of protons and arising the transmembrane gradient of H+ concentration and electric potential. As described above for complex I, the reactions (I.2), (I.5) and (I.7) reduce ubiquinone each time taking two protons from the matrix. In complex III the reduction of ubiquinone at Qi site by reduced cytochrome bH is coupled with binding two protons taken from the matrix. The rate of this process (v35) is calculated as described in [23]. The electron flow (vO) through complex IV results in the reduction of oxygen with the uptake of two protons from the matrix and additional two protons are transferred from the matrix to cytosol. Proton leak (vlk) and ATP synthesis (vsyn) return the protons transferred back to the matrix: vlk is leak of protons through the membrane:(H.1)vsyn is the reaction rate of ATP synthase, which uses the energy of three protons translocating them back to matrix to synthesize one ATP:(H.2)The total rate of reversible uptake of the matrix protons is expressed as follows:(H.3)The reactions (I.2) and (I.5) also release protons into the intermembrane space. For complex III the rate v30 release protons as described in [23]. The flux vO transfers two electrons outside and leak and ATP synthesis consume the gradient:(H.4)The rates of proton translocations (H.1) and (H.2) change cytosolic (outside) and matrix (inside) proton concentrations (Ho and Hi) as described by the following differential equations:(H.5)(H.6)Here bo, bi, Vo, Vi are the buffer capacity and volume of outer and inner intracellular space with regards to mitochondria respectively.
The differential equation for electric potential difference (ψ) used the same terms as that for proton concentration, but multiplied by a coefficient, which transforms the flux of ions into the change of electric potential:(H.7)where F is Faraday number (96000 cu/mol or 0.96·10−4 cu/nmol), C is electric capacity of the membrane (2·10−4 F/mg of protein, as computed based on [44].
Substrates for respiration, i.e. NADH and succinate are produced in TCA cycle inside mitochondria and in the model the connection of this part of intracellular metabolism with respiration through these common metabolites is taken into account by the simulations of following reactions.
Since the emphasis of work described here is the operation of respiratory chain, the reactions of TCA cycle were simulated in simplified form, as linear function of each substrate. Such expressions assume that the substrate concentrations are far from saturation, which should be true for the most cases. In this case the usual hyperbolic dependence of enzymatic reactions on substrate concentrations is close to the linear dependence. On the other hand, this simplification allows to avoid such unfavorable situation, when choosing inappropriate Km makes reactions artificially insensitive to substrate changes. Therefore we used such assumption as a first approximation, which could be easily corrected with obtaining more information about the properties of system.
Pyruvate transport and transformation to acetil coenzyme A:(T.1)Citrate Synthase :(T.2)here the conversion of pyruvate into acetyl coenzyme A, linked with NAD+→NADH transformation, is included in the same reaction.
The reactions converting citrate into succinate were joined together, taking into account that NAD+ is used in these reactions:(T.3)Then succinate is transformed into fumarate in succinate dehydrogenase reaction, which reduces Q taking two protons from matrix:(T.4)Here the total content of reduced and oxidized ubiquinone is conserved at the levels defined by [45] (6 nmol/mg prot).
Succinate not only could be produced in TCA cycle but also transported from outside of mitochondria in exchange to fumarate or malate (which are lumped in one pool in the present version of the model):(T.5)Succinate could also be exchanged to phosphate:(T.6)Malate dehydrogenase reaction transforms lumped fumarate/malate pool into oxaloacetate producing NADH:(T.7)Malic enzyme transforms malate into pyruvate producing NADH:(T.8)The concentrations of metabolites were calculated by numerical solving the differential equations constructed using the above expressions for reaction rates:(T.9)(T.10)(T.11)(T.12)(T.13)The differential equation, which describes the evolution of NADH takes into account the stoichiometry of its production in TCA cycle and consumption by complex I (reaction (I.0) described above.(T.14)The total concentration of NAD and NADH is assumed to be constant, so that NAD+, which defines the rates of TCA cycle reactions is computed as CNAD = CNADt−CNADH.
The reactions linked with electron transport and respective parameters are summarized in Table 4. In total, without the reactions of pyruvate transport and ATP synthase, which were switched out in accordance with experiments analyzed, this module contains 11 parameters.
As the presented equations show, although the expressions for reaction rates are simplified, the stoichiometry of succinate and NADH production and succinate transport is reflected precisely in the model and this was the most important for the presented step of study of the link between central metabolism and ROS production by electron transport chain and the role of reverse electron transport in this process.
The whole model contains (22−6)+(18−7)+11 = 51 parameter (22 for complex III, 18 for complex 1, and 11 for the rest of reactions simulated). The six parameters of complex III and seven parameters of complex I are defined by the known values of midpoint potential. The other parameters were validated by fitting experimental data. To fit the experimental data our modification of Simulating Annealing algorithm was implemented in the way similar to that in [22]. The specificity of this algorithm was defined by the specificity of experimental data. The dynamics of NAD+ reduction was measured under the two different conditions, in the presence and absence of rotenone, an inhibitor of reduction/oxidation of quinones in complex I. The presence of rotenone was simulated by decreasing to zero the rate constants of step 5 in the reactions performed by complex 1 (kf15 = kr15 = 0), and the two conditions were fitted simultaneously for the same values of all other parameters. The procedure consisted of minimization of χ2, normalized sum of squares of deviations from experimental data. χ2 was calculated based on two simulations, first, normal conditions and, second, the presence of rotenone (kf15 = kr15 = 0, and all other parameters as in the first simulation).
The fitting algorithm made the following actions:
The cycles of perturbations and coordinate descent repeated thousands times and saved sets of parameters were analyzed: program read the saved sets with corresponding values of χ2, defined the best fit (absolute minimum of χ2), the set of parameters, corresponding to the best fit, and defined confidence intervals for the parameters using the criterion of Δχ2 [31].
All procedures involving animals were approved by Children's Hospital of Pittsburgh and were in compliance with “Principles of Laboratory Animal Care” and the current laws of the United States.
Brain mitochondria were isolated from the cortex of adult Wistar rats. After removal, tissue was minced and homogenized in ice-cold isolation buffer I (IB I) which contained: 225 mM mannitol, 75 mM sucrose, 5 mM HEPES buffer (pH adjusted to 7.3 with KOH), 0.1 mg/ml fatty acid free BSA, 1 mM tetrapotassium EDTA and 12% Percoll. The homogenate thus obtained was carefully layered on the top of a discontinuous gradient of Percoll (24% and 42%) prepared using the same buffer. The preparation was then centrifuged at 16,000×g for 10 min. The fraction containing the mitochondria located between 42% and 24% Percoll was carefully withdrawn by a syringe and washed from Percoll twice by pelleting in IB I. The resulting mitochondrial suspension was diluted in isolation buffer II (IB II), which was similar to IB I, except for the concentration of EDTA (0.1 mM) and lack of albumin, and spun down at 12,000×g for 10 min. The deposit of mitochondria was homogenized in IB II at a final protein concentration of ∼20 mg/ml and stored on ice until use. The protein concentration in the mitochondrial samples was determined using a Protein Assay kit (Pierce Chemical Company, Rockford IL) according to the manufacture's instructions. Mitochondria prepared in this way were active for at least 5–6 hours, as determined by their ability to maintain a stable transmembrane potential in the presence of oxidizable substrates.
Fluorescence measurements were performed in a stirred cuvette mounted in a Shimatzu RF-5301 spectrofluorimeter maintained at 37°C. Mitochondria (0.2 mg/ml of protein) were added to 1.5 ml of the basic incubation medium that contained: 125 mM KCl; 2 mM KH2 PO 4; 2 mM MgCl2; 10 mM Tris;10 mM HEPES (pH 7.0); 100 µM EGTA; and oxidizable substrates as indicated in a particular experiment. Concentration of rotenone, when indicated, was 1 mkM.
Fluorescence of NAD(P)H was measured at excitation/emission wavelengths 365 nm (slit 5 nm)/463 nm (slit 10 nm), respectively. To quantify the measurements a calibration curve was constructed using standard concentrations of commercial NADH.
Hydrogen peroxide was measured by fluorescence of Amplex red (2 µM), which increased upon oxidation to resorufin in the presence of 1 U/ml of horseradish peroxidase (HRP) as previously described [15]. Measurements were carried out at excitation/emission wavelengths of 560 nm (slit 1.5 nm)/590 nm (slit 3 nm), respectively. Amounts of H2O2 released by mitochondria were estimated by constructing calibration curves using known H2O2 concentrations in the standard incubation buffer together with Amplex red and HRP, but without mitochondria.
Mitochondrial transmembrane potential, ΔΨm, was estimated using fluorescence quenching of the cationic dye safranine O. Since polarized mitochondria have a negative charge inside, positively charged molecules of safranine O are accumulated inside the matrix; increase in dye concentration inside the matrix leads to fluorescence quenching, thus, a decrease in fluorescence corresponds to an increase of membrane potential. The excitation wavelength was 495 nm (slit 3 nm) and emission 586 nm (slit 5 nm), and the dye concentration used was 2.5 µM [15].
Mitochondrial respiration rates were measured by an Oroboros High Resolution Respirometer (Innsbruck, Austria) in a stirred 2 mL chamber at 37°C in the same incubation media as indicated above. Oxygen sensor was calibrated at each experiment according to the manufacture's instructions. Calculations of respiratory rates were performed by software built into the instrument.
|
10.1371/journal.pcbi.1003261 | Lipid Receptor S1P1 Activation Scheme Concluded from Microsecond All-Atom Molecular Dynamics Simulations | Sphingosine 1-phosphate (S1P) is a lysophospholipid mediator which activates G protein–coupled sphingosine 1-phosphate receptors and thus evokes a variety of cell and tissue responses including lymphocyte trafficking, endothelial development, integrity, and maturation. We performed five all-atom 700 ns molecular dynamics simulations of the sphingosine 1-phosphate receptor 1 (S1P1) based on recently released crystal structure of that receptor with an antagonist. We found that the initial movements of amino acid residues occurred in the area of highly conserved W2696.48 in TM6 which is close to the ligand binding location. Those residues located in the central part of the receptor and adjacent to kinks of TM helices comprise of a transmission switch. Side chains movements of those residues were coupled to the movements of water molecules inside the receptor which helped in the gradual opening of intracellular part of the receptor. The most stable parts of the protein were helices TM1 and TM2, while the largest movement was observed for TM7, possibly due to the short intracellular part starting with a helix kink at P7.50, which might be the first helix to move at the intracellular side. We show for the first time the detailed view of the concerted action of the transmission switch and Trp (W6.48) rotamer toggle switch leading to redirection of water molecules flow in the central part of the receptor. That event is a prerequisite for subsequent changes in intracellular part of the receptor involving water influx and opening of the receptor structure.
| The activation of G-protein-coupled receptors (GPCRs) depends on small differences in agonist and antagonist structures resulting in specific forces they impose on the helical bundle of the receptor. Having the crystal structures of GPCRs in different stages of activation it is possible to investigate the successive conformational changes leading to full activation. The long molecular dynamics simulations can fill the gap spanning between those structures and provide an overview of the activation processes. The water molecules are recognized to be crucial in the activation process which link shifting of ligand in the binding site, the actions of molecular switches and finally the movements of fragments of TM helices. Here, we present five 700 ns MD simulations of lipid S1P1 receptor, either in Apo form, or bound to antagonist ML056 or natural agonist S1P. The Apo and antagonist-bound receptor structures exhibited similar behavior, with their TM bundles nearly intact, while in the case of the agonist-bound receptor we observed movements of intracellular ends of some of TM helices.
| Sphingolipids together with glycerol-based phospholipids are major structural components of cell membranes. In response to various extracellular stimuli, including growth factors, inflammatory cytokines, antigens, and agonists of some GPCRs, the sphingolipids can be metabolized into potent mediators, such as sphingosine-1-phosphate (S1P) [1]. This sphingolipid has emerged as an important signaling mediator participating in the regulation of multiple physiological and pathological processes taking place in cancer, cardiovascular diseases, wound healing, atherosclerosis and asthma but also is important in pathological conditions such as inflammation and stress. It can also trigger a range of biological effects such as cell migration, differentiation, apoptosis, immunity, proliferation and angiogenesis [2]–[5]. The functioning of S1P receptors in the maintenance and modulation of the activity of the biological barrier is of the profound biological importance and has many therapeutic implications including treatment of multiple sclerosis, prevention of the transplant rejection and probably the adult respiratory distress syndrome as well [6]–[11]. Within the five known high-affinity S1P receptors the S1P1 receptor subtype is the most commonly expressed in various cell types including cardiac cells, endothelial cells and neurons [11]–[15]. Studies on deletions in the S1P1 gene have revealed its essential endothelial function in the arterial smooth muscle cell migration [16]. The S1P1 knockout mice exhibit embryonic lethality or abnormalities in the development of the immune system [11], [17], [18].
The recently published crystal structure of S1P1 with antagonist ML056 by Stevens group [19] (PDB code: 3V2Y) showed a detailed ligand binding mode including the precise position of a long hydrophobic tail of a ligand regardless of lack of directional bonds establishing its location in the binding site. The authors also predicted the binding mode of an agonist S1P by docking it to the same binding site as the antagonist. Based on the docking results they concluded that the long hydrophobic tail of the agonist is responsible for the receptor activation as it was not possible to fit it to the antagonist-bound crystal structure with preserved interactions of a zwitterionic head. Only after allowing the receptor structure to adapt to the agonist it was possible to fit the hydrophobic tail and simultaneously preserve the polar interactions of the ligand head. However, the exact mechanism of the S1P1 activation is still not known and it is particularly interesting to learn how these changes are evoking passing of a signal to the cytoplasmic side of the receptor. To address that issue, we conducted five all-atom 700 ns MD simulations for the Apo form of S1P1, antagonist ML056-bound S1P1 and agonist S1P-bound S1P1. We studied movements of amino acid residues in centrally located area where the transmission switch operates. We also proposed the pathway of the activation mechanism involving the movement of water molecules as it was recently detected during simulations of the model of the formyl peptide receptor FPR1 [20].
The S1P agonist coordinates were obtained from the PUBCHEM online database [21]. The ligand preparation utility in MacroModel [22] was used to optimize the geometry of the initial structure. The systematic conformational search was also performed in MacroModel and top five conformers of the lowest potential energy were kept for docking. The docking procedure was performed using Glide [23], [24] (Schrödinger 2012 suite). The protonated state of primary amine of S1P and ML056 at physiological pH was predicted by Epik [25], [26] and resulted in zwitterionic head group of both ligands. The S1P molecule was initially placed in the binding pocket with a pose similar to the antagonist molecule in the S1P1 crystal structure (PDB: 3V2Y). Cubic box defining the docking area was centered on the ligand mass center with a box size of 10 Å. Next, the flexible ligand docking was performed. Ten poses out of 10,000 were included in the post-docking energy minimization and the best scored pose was chosen for MD simulation. For an antagonist ML056 present in the crystal structure no ligand optimization was performed but only addition of hydrogen atoms according to the calculated protonated state.
To obtain the atomic partial charges for S1P and ML056 ligands, the structures obtained from docking were energy-minimized and the electrostatic potentials were obtained. The quantum mechanical calculations were done in GAUSSIAN 09 program [27] with 6-31G* basis set. The obtained potentials were used as input for the RESP (Restrained-Electrostatic Potential) fit method [28] performed by the R.E.D. tools [29]. All ligand topology parameters were generated using SwissParam web server [30].
The crystal structure of the S1P1 receptor lacks of two intracellular loops ICL2 (amino acids 149–155) and ICL3 (amino acids 232–244). The latter one, between helices TM5 and TM6, was substituted by T4-lysozyme to stabilize the structure. The original missing loops were modeled in Modeller 9v10 [31] and Rosetta loop modeling tools [32]. Initial 5000 loop conformations were generated in Modeller, and conformations with the lowest DOPE score were submitted to the Rosetta loop modeling for an all-atom refinement (the kinematic closure method). The unstructured part of C-terminus, the residues 327–330 after helix H8, was removed in our model.
Pre-equilibration of the lipid bilayer composed of POPE phospholipids (1-palmitoyl-2-oleoyl-sn-glycero-3-phosphoethanolamine) and embedding of the receptor into lipid bilayer was done in Maestro 9.2 program [33] and in Desmond [34] program. We used 23 Na+ and 44 Cl− ions to make the system neutral and to set the ionic strength to 0.15 M. The total number of atoms in the investigated system was approximately 50,000 including about 8,300 water molecules and 132 POPE phospholipids. The periodic box dimensions were set to 7.0 nm×7.0 nm×10.4 nm. Equilibration of the system was performed at the constant pressure and temperature (NPT ensemble; 310 K, 1 bar) employing Berendsen temperature and pressure coupling scheme [35] under CHARMM36 force field [36]. All bond lengths to hydrogen atoms were constrained using M-SHAKE algorithm [37]. Van der Waals and short-range electrostatic interactions were cut off at 10 Å. Long-range electrostatic interactions were computed by the particle mesh Ewald (PME) summation scheme [38]. A RESPA (time-reversible reference system propagator algorithm) integrator [39] was used with a time step of 2 fs. Long-range electrostatic interactions were computed every 6 fs. Harmonic positional restraints on the protein backbone were tapered off linearly from 10 to 0 kcal/mol−1A−2 over 20 ns. Additional 20 ns NPT equilibration without restraints was executed afterwards. Finally, 700 ns simulations were performed for Apo receptors, and with agonist and antagonist bound structures. All simulations were performed in Desmond [34]. To facilitate comparison of our structure to other GPCRs the Ballesteros-Weinstein numbering scheme [40] was used (numbers in superscript) apart from the sequence numbers of S1P1 residues. The Desmond force field parameters for both ligands, S1P and ML056, are provided as a supplementary information (Protocol S1).
After the non-restrained final step of equilibration procedure the backbone of the TM core and loops were matching the crystal structure. Only the loose, unstructured N-terminus (amino acids 16–21) was freely moving during equilibration. The amino acids in the binding site of Apo receptor structure were nearly in the same positions as in the crystal structure with exception of S1052.64 at extracellular end of TM2 (movement of whole residue 1.5 Å outside of the receptor) and a rotamer of M1243.32 side chain which was oriented in such a way that it took a position occupied in the crystal structure by the ligand's hydrophobic tail. Contrary, those two residues, S1052.64 and M1243.32, in the MD simulation of the antagonist-bound receptor were matching the crystal structure. After the equilibration the antagonist molecule took a slightly shifted position compared to that of the crystal structure as its phosphate group lost a direct contact with R1203.28 though preserving the interaction with K34, located in a short linker between two helices in N-terminus. What is more, the charged amino group of antagonist gained another favorable interaction, apart from E1213.29. This happened due to the N1012.60 residue, which flipped and started to interact with the nearby E1213.29 and amine group of antagonist. We also observed a solvent-mediated hydrogen bond between the antagonist and R1203.28. The carbonyl group of antagonist formed a hydrogen bond with Y982.57 which was not present in the crystal structure (too large distance 4.7 Å). The binding modes of investigated ligands are shown in Figure 1 while detailed interactions with adjacent amino acids are shown in Figure S1. The interactions of ML056 in the binding site of S1P1 receptor were preserved until the end of 700 ns MD simulation, apart from residue Y382.57 which rotated away and formed a hydrogen bond with S3047.46 located few residues to the highly conserved NPxxY motif on helix TM7. The same bond was formed during MD simulation of Apo receptor but not during a simulation with agonist (Figure 2).
Water molecules, which were not visible in the crystal structure due to its low resolution were found to fill the empty binding site of Apo receptor after equilibration and during MD simulation. In case of ligand-bound receptor structures a number of water molecules in the binding site was only slightly smaller than that in Apo receptor because both ligands took positions mostly inaccessible to water molecules. Only the polar and charged groups of zwitterionic head had a contact with water (Figure S1). In case of the structure of agonist bound receptor, at the beginning of MD simulation, the zwitterionic head interacted indirectly with amino acids via water molecules but this changed during the simulation (Figure S1B and S1D). After equilibration of the S1P/S1P1 complex the phosphate group of S1P interacted directly with K34 (similarly to antagonist) but also with Y29 (as in the crystal structure of antagonist-bound complex). Those interactions were stable through the whole MD simulation. However, in contrast to the antagonist case, both residues E1213.29 and N1012.60 did not interact directly with agonist, but only via water molecules. However, during simulation, the phosphate group started to interact with R1203.28 and the OH group of S1P formed a hydrogen bond with N1012.60, while S1052.64 interacted with both the hydroxyl and the amine group of agonist. The superimposition of both studied ligands, ML056 and S1P, in the receptor binding site is shown in Figure 1B. The hydrophobic tail of both ligands is located mostly in the same area surrounded by helices TM3 and TM5-7 as well as hydrophobic residues from extracellular loop ECL2. The ends of both ligands are pointing toward the same region of TM5, however, a tail of S1P is longer and reaches a hydrophobic cluster composed of three phenylalanine residues, F1253.33, F2105.47 and F2736.52, centered at TM5.
During the MD simulations we observed several movements of aromatic residues (Figure 2) which can be interpreted as possible rotamer switches. In Apo and antagonist bound receptor complex structure a residue Y982.57 changed its conformation, which led to the formation of a hydrogen bond with S3047.46 (Figure S2). Although one cannot exclude that such movement is a result of slightly different binding of antagonist compared to the crystal structure, the analogous rotameric change in Apo receptor structure is striking. Additionally, in case of the antagonist complex the residue W2696.48 is fluctuating and its χ2 angle is changing between 0 and 90 degrees, until the rotation of Y982.57 occurs (Figure 2B). The changes of W2696.48 are much smaller in the second simulation with antagonist (Figure 2B′). Contrary, in the case of agonist S1P-bound complex, a stable rotamer of Y982.57 is a result of a hydrogen bond between Y982.57 and a backbone carbonyl group of L2977.39. Such a bond was created during equilibration period and was stable until the end of simulation. In the crystal structure the residue Y982.57 is bound neither to the ligand nor to any other receptor residue. Instability of residues W2696.48 and F2656.44 together with Y982.57 rotamer “up” in case of agonist-bound receptor (Figure 2C and 2C′) may be a prerequisite to rearrangement of residues located close to the highly conserved W2696.48. Such rearrangement is called a transmission switch [41], [42] (previously called a tryptophan rotamer toggle switch) and can lead to the movement of cytoplasmic parts of helices TM6 and TM7 outward of the receptor center.
Similar scheme of activation was recently described for adenosine A2A receptor based on its 1.8 Å high resolution antagonist-bound structure [43]. The structure contains 177 structured water molecules, 57 of which occupy the interior of the 7TM bundle. In the antagonist-bound A2AR (PDB id: 4EIY) there is so called water channel (Figure S3A). The channel has two bottlenecks close to residues W2466.48 and Y2887.53, respectively, reducing its diameter to slightly less than one water molecule (2.4 and 2.0 Å, respectively) dividing the channel into three parts. Rearrangement of the receptor backbone and side chains due to agonist binding (PDB id: 3QAK) makes the structure more open in bottleneck areas suggesting possibility of formation of continuous hydrogen bond network involving water (Figure S3B). The importance of water molecules for GPCR activation have been also reported in several previous studies [20], [44], [45].
In our simulations, we found that the residue Y982.57 can redirect the flow of water molecules. Keeping a rotamer in “up” position (agonist-bound state) Y982.57 prevents water molecules to enter the area between Y982.57 and W2466.48, but instead allows more water to come near the highly conserved residue D912.50 (Figure S4). In simulations of Apo S1P1 and ML056/S1P1 the number of water molecules within 4 Å distance to D912.50 is much smaller than in agonist-bound complex: there are 3–4 water molecules in Apo state versus about 5–7 molecules in antagonist-bound state and approximately 8–10 molecules in agonist-bound state. Those water molecules form an extensive hydrogen bond network between highly conserved residues N631.50, D912.50 and N3077.49 which can facilitate receptor activation and opening of the cytoplasmic part of the receptor.
During a simulation of agonist-bound receptor the side chain of W2696.48 rotated about 90° between vertical and horizontal positions (Figure 3A–B). This movement facilitated conformational change of adjacent residue F2656.44 located one helix-turn down towards the receptor center in agonist-bound structure. Only after that movement it was possible for the water to enter into the vicinity of D912.50 residue (Figure S4) in ligand-bound state (agonist and antagonist). Final rotamer of W2696.48 is the same as in the crystal structure but its movement facilitated rotameric change of F2656.44 and flow of water (Figure 3C).
Movement of water molecules at inner membrane part of the receptor (close to the NPxxY motif in TM7) can be seen in Figure 4A and 4A′. Large amounts of water accumulate at this position starting at 150 ns in 1st simulation and at 400 ns in 2nd simulation in agonist-bound receptor. At the same time there is much smaller number of water molecules in case of Apo and antagonist-bound receptor (Figure 4A and 4A′). The reason for such behavior is the change of shape of TM7 (Figure 5A). During the MD simulation the kink angle of TM7 with a pivot point at P3087.50 was changing gradually from 155° to 130° with a temporary restoration of initial value between 100 and 200 ns in one simulation (Figure 5B). Such relatively fast movement of intracellular part of TM7 helix is facilitated by short length of that part which consist of two helix turns only. Because of that, a change of TM7 could be the first movement of the transmembrane helix bundle during the activation. Increased volume of this area can accommodate more water molecules (Figure 4B) and make room for the G protein.
As it can be seen from RMSD plots of the receptor backbone (Figure 6A) there is only a small change (about 2 Å) of backbone structure in case of Apo and antagonist-bound receptor. However, in case of the agonist-bound receptor there is a transient and sudden increase of RMSD (up to 4–5 Å) at 200 ns and ending at 600 ns. Then, the RMSD for both simulation with agonist stabilizes at 3 Å. Such an increase may be associated with movement of residues W2696.48 and F2656.44 (Figure 2C and 2C′) being a central part of transmission switch rearranging of the central part of the receptor. Such flexibility of these residues, although finally they assume nearly the same conformations as before, may be necessary for larger movements of cytoplasmic parts of TMs in the next phase of the activation process. Nevertheless, those preliminary movements can be still noticeable in our simulations. We found that conformations of S1P1 receptor during MD simulations can be divided into three major clusters: “inactive”, “intermediate” and “active” (Figure 6B and S5). Such a division was made based on distances between cytoplasmic ends of TM helices (TM7-TM3, TM3-TM6 and TM6-TM7) from MD simulation of agonist-bound receptor structure. In Figure 6B the central structures from each cluster are shown. Those clusters are well separated so one can easily distinguish three different stages of activation. The “active” conformation differs from the “inactive” one primarily through shifts and rotations of intracellular ends of helices TM3-7 (Figure 6B). During the transition from “intermediate” to “active” stage, the intracellular part of TM7 also rotates while moving away from TM3 and TM6 and an angle at pivot point of TM7 (P3087.50) diminish by 25° i.e. the kink of TM7 increases. Although most likely the full activation of the protein was not achieved in our simulation the obtained directions of TMs movements agree well with activated states of other GPCRs: adenosine receptor A2AR [46], β1- and β2-adrenergic receptors [47], [48], and opsin [49].
Parrill et al. [50] studied effect of S1P1 receptor mutations on binding its natural substrate sphingosine 1-phosphate (S1P). Based on experiments: radioligand binding, ligand-induced [35S]GTPγS binding, and receptor internalization assays, they suggested that three amino acids R1203.28, E1213.29 and R2927.34 were involved in the ligand binding. They illustrated their findings with a model of the ligand-receptor complex constructed on early rhodopsin model based on distance geometry calculations with hydrogen bonding constraints [51]. Those three residues were also shown as binding S1P in S1P1 binding site in more recent paper of the same group [52]. The crystal structure of S1P1 receptor with antagonist ML056 can verify to some extent those findings. The residues R1203.28 and E1213.29 are directly interacting with ligand while R2927.34 is neither interacting nor even being a part of a binding site since its side chain is located outside of a receptor. In our simulations the residue R2927.34 is far from antagonist ML056 but also from agonist S1P. Although not interacting directly with the agonist bound in orthosteric binding site this residue may be required as a selectivity filter on the ligand entry pathway.
Loenen et al. [53] determined differences in ligand-induced S1P1 receptor activation using an in silico guided site-directed mutagenesis. They mutated three residues, Y982.57, R1203.28, and F1253.33, and probed mutants with a chemically diverse set of agonists including S1P. Mutation of residue R1203.28 resulted in a reduction of the potency of all ligands, measured as an inhibition of forskolin-induced cAMP accumulation. For all compounds the effects observed for the R1203.28A mutation were larger than those observed for the R1203.28K, however an effect of subtle mutation R1203.28K was the biggest in case of reducing potency of the endogenous agonist S1P. Mutation of Y982.57F did not significantly affect S1P1 agonist potency for any of the ligands tested, however, a mutation of this bulky residue into alanine affected the potency of S1P by almost 80-fold. Also a mutation F1253.33Y did not significantly affect the potency of S1P. The above results are in agreement with our simulations: the agonist S1P formed a tight contact with residue R1203.28 while residues Y982.57 and F1253.33 were located on both sides of the ligand and contributed to hydrophobic interactions so exchanging them into alanine could result in reduced binding.
Recently, Satsu et al. [54] described a selective allosteric agonist of S1P2 receptor. Mutation of receptor residues responsible for binding to the zwitterionic head group of natural agonist S1P abolished activation of the receptor by S1P, but not activation by synthetic ligand CYM-5520. Competitive binding experiments with radiolabeled S1P demonstrated that CYM-5520 was an allosteric agonist which did not displace the native ligand. Computational modeling, based on the crystal structure of S1P1 receptor, suggested that CYM-5520 could bind beneath the orthosteric binding pocket, so that co-binding of S1P could not be affected. Possibly, the similar allosteric agonists can be found for S1P1 receptor.
The proposition of activation mechanism of S1P1 receptor based on our simulations is illustrated in Figure 7. After binding of agonist S1P to the binding site of S1P1, the movement of acyl tail of S1P leads to the flipping of W2696.48 (step 1). Such rotameric change alters the conformation of side chain of F2656.44 which is located next to W2696.48 in the same helix TM6 (step 2). These residues form a core of a transmission switch which involves rearrangement of centrally located residues including N631.50, D912.50, S3047.46 and N3077.49. They facilitate a redirected flow of water molecules inside a receptor (step 3). The influx of water molecules at intracellular part of the receptor leads to limited motions of cytoplasmic ends of TM helices, with the largest movement associated with TM7 (step 4), which is a prerequisite for larger motions of the cytoplasmic parts of transmembrane helices. These movements lead to opening the protein structure to make room for binding a G protein. The mutations of S1P1 receptor analyzed so far were located close to the orthosteric binding site of native agonist S1P. However, finding of the allosteric agonist not having charged functional groups implicated its different binding mode. Possible binding site of this compound close to residue W6.48 in S1P2 receptor may have a direct influence on action of the transmission switch. Investigations of residues close to this region could shed some light on activation processes of S1P1 receptor and maybe discriminate effects of allosteric from orthosteric binding. Studying mutations of R2927.34 and nearby residues is required to analyze how ligands can enter the receptor binding sites both orthosteric and allosteric. The residues found to be important in our simulations for the transmission switch, including D912.50, Y982.57, F2656.44, W2696.48, N3037.45 and S3047.46 are forming a cluster in the central part of S1P1 receptor. Mutagenesis studies of those residues may be important to elucidate the details of transmission switch and also to discover the receptor structures hampered at different stages of activation during action of this complex switch. Additional simulations of wild type and mutated S1P1 receptor complexes with different ligands, including those bound in allosteric sites, will be extremely helpful to visualize or guide the site directed mutagenesis experiments and also to explain the exact role of particular residues in receptor activation.
|
10.1371/journal.pcbi.1003994 | Developmental Self-Construction and -Configuration of Functional Neocortical Neuronal Networks | The prenatal development of neural circuits must provide sufficient configuration to support at least a set of core postnatal behaviors. Although knowledge of various genetic and cellular aspects of development is accumulating rapidly, there is less systematic understanding of how these various processes play together in order to construct such functional networks. Here we make some steps toward such understanding by demonstrating through detailed simulations how a competitive co-operative (‘winner-take-all’, WTA) network architecture can arise by development from a single precursor cell. This precursor is granted a simplified gene regulatory network that directs cell mitosis, differentiation, migration, neurite outgrowth and synaptogenesis. Once initial axonal connection patterns are established, their synaptic weights undergo homeostatic unsupervised learning that is shaped by wave-like input patterns. We demonstrate how this autonomous genetically directed developmental sequence can give rise to self-calibrated WTA networks, and compare our simulation results with biological data.
| Models of learning in artificial neural networks generally assume that the neurons and approximate network are given, and then learning tunes the synaptic weights. By contrast, we address the question of how an entire functional neuronal network containing many differentiated neurons and connections can develop from only a single progenitor cell. We chose a winner-take-all network as the developmental target, because it is a computationally powerful circuit, and a candidate motif of neocortical networks. The key aspect of this challenge is that the developmental mechanisms must be locally autonomous as in Biology: They cannot depend on global knowledge or supervision. We have explored this developmental process by simulating in physical detail the fundamental biological behaviors, such as cell proliferation, neurite growth and synapse formation that give rise to the structural connectivity observed in the superficial layers of the neocortex. These differentiated, approximately connected neurons then adapt their synaptic weights homeostatically to obtain a uniform electrical signaling activity before going on to organize themselves according to the fundamental correlations embedded in a noisy wave-like input signal. In this way the precursor expands itself through development and unsupervised learning into winner-take-all functionality and orientation selectivity in a biologically plausible manner.
| In this paper we address the question of how progenitor cells of the neocortical subplate can give rise to large functional neuronal sub-networks in the developed cortex. We choose winner-take-all (WTA) [1], [2] connectivity as the target of this self-construction and -configuration process because these sub-networks are consistent with the observed physiology [3], [4] and connectivity [5], [6] of neurons in the superficial layers of neocortex, and because they are powerful elements of computation [7], [8]. WTA networks actively select the strongest of multiple input signals, while suppressing the weaker ones. This fundamental characteristic is applicable in various contexts, and so many studies modeling cortical function are based on WTA modules [8]–[15].
The idealized WTA network architecture is shown in Fig. 1A. Excitatory neurons are recurrently connected to each other and also with one or more inhibitory neurons, which project back to the excitatory neurons. This architecture does not in itself guarantee WTA functionality. The degree of recurrent excitation, excitation of inhibitory neurons, and inhibition of excitatory neurons need all to lie within preferred ranges [8] in order for the network to exhibit effective WTA behavior. The appropriate neural architecture must be grown, and then the weights of the many synapses must be tuned to fall within the necessary ranges. Such neuronal growth and synapse formation are subject to variability (1B,C), for which the homeostatic learning mechanisms must compensate.
The behavior of a WTA network depends on the ratios of the effects of its various excitatory and inhibitory connection paths. In its high excitatory gain regime a WTA network will report only the strongest of its feed-forward inputs, and suppress the remainder of the excitatory neurons, which are weakly activated. In a more relaxed regime (soft-WTA, sWTA) the network will return a pattern of winners that best conforms to its input. In this sense the sWTA performs a pattern based signal restoration, which is a crucial mechanism for resisting degradation of processing in neural systems across their many computational steps. In this paper we choose to have the developmental process grow and tune these sWTA networks.
Our goal is to demonstrate how plausible genetic developmental mechanisms can combine with homeostatic synaptic tuning to bring networks of neurons into sWTA functionality (Fig. 1). Our demonstration is based on simulations of the development and growth of neural tissue in 3D physical space using Cx3D [16]. The simulation begins with a single precursor cell. This cell encodes gene-like instructions that are sequentially and conditionally expressed through a gene regulatory network (GRN). By controlling the expression of different genes, this GRN gives rise to pools of differentiated excitatory and inhibitory neurons. These neurons, which are placed randomly in 3D space, extend axons and dendrites and make synapses according to a proximity rule. This process results in a synaptically connected network that matches well experimentally obtained connectivity statistics. During this neurite outgrowth, the synaptic weights calibrate themselves homeostatically using experimentally established synaptic scaling [17] and BCM learning rules [18]. This synaptic learning is conditioned by coarsely patterned neuronal activity similar to that of retinal waves or cortico-thalamic loops [19]–[23].
We compare these grown networks with biological data, and demonstrate WTA functionality. This comparison is done also in the context of cortical functionality, such as orientation selectivity. Importantly, the overall behavior stems solely from local processes, which are instantiated from internally encoded and developmental primitives [24]. Hence, we provide a model that explains the developmental self-construction and -configuration of a neocortical WTA network in a biologically plausible way.
Cell proliferation and differentiation into different cell types is specified implicitly in the genetic code of a single precursor cell. This code determines how a given number of excitatory and inhibitory neurons is produced. During the unfolding process of this code, each cell contains the same genetic code, but because of its local environment can follow different developmental trajectories.
We model the molecular mechanisms that regulate cell differentiation by a dynamical gene regulatory network (GRN). This GRN is defined by a set of 5 variables (, , , , and ) that represent substance concentrations, where each substance is the expression of a gene. Importantly, all cells have their own instantiations of these variables. The secretion, interaction, and decay of substances, is regulated by the laws of kinetics. The differential equations specifying these dynamics are shown in Methods.
During the evolution of the substance concentrations, also cell growth and division is simulated. The cell cycle time and model parameters of the differential equations are fixed and independent of the substance dynamics.
Initially, all concentrations are set to zero. At this stage, only the “starter” substance is produced, which reaches high concentration levels in the first time step, and triggers the production of a second gene . is produced according to a prespecified intrinsic production constant . This value determines how many cell divisions will occur until the concentration of reaches a value of . When this threshold is reached, a probabilistic decision is induced: or , responsible for activating the excitatory and inhibitory cell phenotypes, are triggered with probability or , respectively.
Such a GRN network configuration would enable us to generate cells, where is the number of symmetric divisions. However, the target number of cells might not be an exponential of 2. Therefore, we have introduced a second gene that is (probabilistically) activated by high concentrations of , and that leads to a second round of symmetric division. As for , activates or in a probabilistic manner. The probability to enter into this secondary cell cycle is given by , which is computed based on the target number of cells. The evolution of the GRN across cell types is depicted in Fig. 2.
By setting the production rate constant of gene and the probabilistic activation of , we can control the final number of cells produced. The equations for computing the probabilities for either differentiating into neurons by induction () or by induction (), depending on the target number of cells, are shown in Methods.
Overall, the GRN is designed so that a desired total number of cells is reached, and that the distribution of excitatory vs. inhibitory cells follows the approximate 4∶1 ratio observed in cortex [25]–[27] (S1 Figure). Fig. 3(A-D) shows the evolution of an initial cell giving rise to a number of cells which eventually grow out neurites based solely on their genetic encoding.
Neurite growth and arborization is caused by growth cone traction and bifurcation. The growth cone is able to sense the presence and gradient of morphogens and other signal molecules, and also able to actively explore the local extracellular space. Importantly, neurite growth is steered via a growth cone model instantiated at the tip of the axon or dendrite, and so is a local process.
Diffusable signal molecules are secreted by the cell somata. In these simulations excitatory and inhibitory neurons secrete two characteristic signals, that enable excitatory and inhibitory axons to find inhibitory and excitatory neurons, respectively. The axonal growth cones initially grow out of the somata in random directions. However, they retract whenever the concentration they sense falls below a threshold. The retraction stops and growth recommences when a second higher threshold is exceeded. In this way the axons remain close to substance secreting sources. Retraction is an efficient strategy for establishing connections because axons grow only into regions containing a potential target, and is commonly observed in developing neurons [28]–[31]. A video of a developing neural network with axonal retraction (simulated in Cx3D) is included in the Supporting Information (S1 Video) and on Youtube (http://www.youtube.com/watch?v=il2uc-ZUZQ4).
Axons deploy boutons. Whenever these boutons are sufficiently close to a potential post-synaptic site on a dendrite a synapse is created between them. Consequently, the final synaptic network connectivity depends on the nearly stochastic arrangement of regions of spatial proximity of the outgrowing axons and dendrites.
We adapted the parameters of the neurite outgrowth (see Table 1) so that the connectivity of the simulated neuronal growth matched our experimental observations in layers II/III of cat visual cortex [5], [32] (see Fig. 4A). Overall, we found that connectivity was robust to reasonable variation of the growth parameters and the random location of somata. The absolute numbers of synapses simulated here are smaller than observed in biology, due to constraints on computational resources. However, there is no inherent restriction on scalability using our methods, and so we expect that realistic numbers of cells and synapses could if necessary be simulated using supercomputers.
Fig. 4B shows the distribution of the percentage of excitatory input synapses to the neurons, across the whole population. The average percentage of excitatory inputs to a neuron in this network is 84%, which is in good agreement with the experimental data. This result is consistent with observations across species and cortical areas that some 15% of all the synapses are GABAergic [5], [33]–[35], irrespective of neuronal densities. Importantly, this good agreement arises naturally out of the growth model, and did not require extensive tuning of the model parameters.
The self-configuration of electrophysiological processing depends on the tuning of network synaptic weights and neuronal activity. In order to simulate this aspect of the developing networks, we must model also the electrical activity of neurons. However, the time scales of morphological growth and electrophysiological dynamics are many orders of magnitude different, and this difference makes for substantial technical problems in simulation.
For simplicity, and for minimizing computational demands we have used a rate-based approach to modeling neuronal activity. We approximate the neuronal activation by a linear-threshold function [36] that describes the output action potential discharge rate of the neuron as a function of its input. This type of neuronal activation function is a good approximation to experimental observations of the adapted current discharge relation of neurons [37], [38] and has been used in a wide range of modeling works [8], [39]–[41].
The linear-threshold activation function is:(1)where denotes the firing rate of a neuron with index i, is the neuronal time constant, is the spontaneous activity, is the feed-forward input to neuron i, is the weight of the connection from neuron j to neuron i (can be positive or negative, depending on the presynaptic neuron's type), and is the neuron's threshold. For simplicity, and are set to 1 and 0. Exploratory simulations where yielded very similar results.
For computational efficiency, the electrophysiology simulator is implemented as a global process that acts on the total weight matrix of the neuronal network, rather than performing these frequent computations locally. We chose this global methodology because it leads to a significant speed-up compared with a local version that had been used initially. The total weight matrix is obtained by summation of the weights of all synapses in the Cx3D simulation. Using these connection weights, neuronal activity is computed as described in Eq. 1. Connection weight changes resulting from the learning and adaptation (explained below) are computed based on this summed weight matrix and the activities of the two respective connected neurons, which are saved at each electrophysiology time step. The same connection weights (and neuronal activities) would be computed if only local processes at the synapses were simulated, because the synaptic learning and adaptation dynamics (Eq. 2 and 3) are dependent on the (locally available) neuronal activities, and linearly dependent on the synaptic weight. Hence, the dynamics of the summed synaptic weights match the sum of the individual synapse weight changes.
For reasons of biological plausibility, the electrophysiology simulator incorporates a maximum connection weight. This maximum weight for the functional connection strength between two neurons is determined by counting the number of synapses involved. This number, multiplied by the maximal weight of a single synapse, is defined as the maximum of the total connection weight. Hence, neurons that are connected by few synapses can not establish a strong functional link.
In our model, self-configuration of the weights towards sWTA functionality occurs during sequential developmental phases. Sequential phases of electrical adaptation and learning during development have been observed experimentally [42], [43], and have also been applied in previous models [44], [45].
During the first, homeostatic phase neurons adapt the synaptic weights of their own input in order to maintain a target output activity. The effect of this phase is to bring the neuronal firing rates into a balanced regime, and so allow for a reliable synaptic learning without interference by unresponsive neurons or run-away excitation. During the second, specification phase the neurons structure their individual responses by correlation-based learning on their inputs.
We investigated whether our developmental model can account for experimental findings on orientation selectivity in visual cortex; for example, differences in tuning between excitatory and inhibitory neurons. In order to address this question, we assumed that the hills of activity in the input layer correspond to oriented stimuli (e.g. bars), which are smoothly and periodically rotating between 0 and 180 degrees. As anticipated from the previous results, excitatory neurons become highly orientation selective (Fig. 9), in contrast to inhibitory neurons. These results are in line with biological data. For example, [63] have analyzed orientation selectivity of excitatory and inhibitory neurons in mouse visual cortex. They report inhibitory neurons to be more broadly tuned and hence less selective than excitatory, pyramidal neurons. Similar findings were reported by [64]–[68].
We also quantified the orientation tuning based on the orientation selectivity index (OSI), which specifies the degree to which a neuron is selective for orientation. The value of this index lies between 0 (non-selective) and 1 (selective to a single, specific orientation). Fig. 9B shows the distribution of the OSI for excitatory and inhibitory neurons in a WTA network, demonstrating the discrepancy of orientation selectivity also on a population level. We conducted additional simulations, which demonstrated that when inhibitory neurons follow the same learning rule as excitatory neurons, they exhibit more narrowly tuned orientation selectivity (Fig. 9C). Hence, experimental findings of orientation selective inhibitory neurons in cat visual cortex [69]–[72] can also be accounted for by our model.
We have analyzed the consequences of our model on the nature of the inhibition of excitatory neurons. As mentioned above, inhibitory synapses onto excitatory neurons are subject to the BCM learning rule (Eq. 3).
The competition between excitatory neurons depends on the common input that they all receive from inhibitory neurons. This common input must reflect the overall activity of the network, so that the competition is suitably normalized. However, the inhibition of the excitatory neurons stems from multiple inhibitory neurons, which should partition their common inhibitory task amongst each other in a self-organizing way. We investigated this partitioning, and how an excitatory neuron is inhibited during stimulation.
In order to quantify the impact of a neuron j on another neuron i for a given stimulus, we calculate a value that we will call the recursively effective exertion (REE). It is obtained by multiplying the activity of neuron j (under a given stimulus ) with the total connection weight from neuron j to i:(5)
The REE value is therefore stimulus-dependent, and dependent on the recurrent network connectivity. Fig. 10 shows that inhibition is distributed non-uniformly: A few inhibitory neurons dominate the suppression of an excitatory neuron. This dominance is due to the BCM learning by inhibitory synapses: Strongly and weakly correlated inhibitory connections to excitatory neurons are strengthened or weakened, respectively. These inhibitory connection strengths converge because of the homeostatic activity regulation, which is part of the BCM learning rule.
The nature of inhibition of excitatory neurons is interesting in the context of the anatomy of inhibitory basket cells. These neurons predominantly target locations close to the soma or the proximal dendrites, where they can strongly influence the excitatory neuron [73]. Therefore, it is plausible that the recruitment of a small number of inhibitory neurons is sufficient to inhibit an excitatory neuron. Electrophysiological experiments could in principle validate this hypothesis by showing that only a small proportion of the inhibitory neurons projecting to a pyramidal neuron are predominantly responsible for its suppression.
In this paper we have demonstrated by simulation of physical development in a 3D space, how an autonomous gene regulatory network can orchestrate the self-construction and -calibration of a field of soft-WTA neural networks, able to perform pattern restoration and classification on their input signals. The importance of this result is that it demonstrates in a systematic and principled way how genetic information contained in a single precursor cell can unfold into a functional network of neurons with highly organized connections and synaptic weights.
The principles of morphological and functional development captured in our model are necessarily simplified with respect to the boundless detail of biology. Nevertheless, these principles are both strongly supported by experimental data, and sufficiently rich in their collective expression to explain coherently the complex process of expansion of a genotype to a functional phenotypic neuronal circuit. In this way our work offers a significant advance over previous biological and modeling studies which have focused either on elements of neuronal development, or on learning in networks whose initial connectivity is given. Therefore we expect that methods and results of the kind reported here will be of interest both to developmental biologists seeking a modeling approach to exploring system level processes, as well as to neuronal learning theorists who usually neglect the genetic-developmental and homeostatic aspects of detailed learning in favor of an initial network that serves as a basic scaffold for subsequent learning [74]–[76].
It is relatively easy to express a well-characterized biological process through an explicit simulation. That is, one in which the simulation simply recapitulates the process by expanding some data through a simple model, without regard for physical and mechanistic constraints. By contrast, the simulation methods [16] that we have used here are strictly committed to physical realities such as 3D space, forces, diffusion, gene-expression networks, cellular growth mechanisms, etc. Our methods are also committed to local agency: All active processes are localized to cells, can only have local actions, and have access to only local signals. There is no global controller with global knowledge, able simply to paint the developmental picture into a 3D space. Instead, the ability of a precursor cell to expand to a functional network is the result of collective interaction between localized cellular processes. And overall, the developmental process is the expression of an organization that is encoded only implicitly, rather than explicity, in the GRN of the precursor cell. Thus, our GRN encodes constraints and methods rather than explicit behaviors.
In previous work [77], [78] we have shown how this approach can be used to explain the development of neocortical lamination and connectivity. In that case we did not consider also the electrophysiological signaling between cells and so the self-configuration of their computational roles, as we have done here. However, the incorporation of electrophysiological signaling into the growth model brings substantial technical difficulties, such as those arising out of the large differences in spatio-temporal scales between cellular developmental and electrophysiological signaling processes, as well as the supply and management of sufficient computational resources. Therefore we have chosen to keep these problems tractable in this first functional study, by restricting our question to a sub-domain of cortical development: How could neuronal precursors expand into functional circuits, at all. Even then, we must be satisfied for the moment with a rate based model of neuronal activity, rather than a fully spiking one.
The emphasis of this paper is on the process whereby a precursor expands to some useful network function. The particular function is less relevant, and in any case the functional/computational details of cortical circuits are as yet not fully understood. We have chosen to induce WTA-like function because our previous work has been focused on the likely similarity between the WTA motif and the neuronal types and their inter-connectivity in the superficial layers of cortex [6]. Moreover these WTA networks are intriguing from both the biological, and computational perspective [3], [6]–[15], [41]. The strong recurrent excitation available in the superficial layers of cortex, and their critical dependence on feedback inhibition has been clearly demonstrated by intracellular recordings in the presence of ionophoretic manipulation of GABA agonists and antagonists [4]. These relationships are crucial for WTA-like processing, because they offer the network induced gain that is crucial for providing the signal restoration, signal selection, and process control that support systematic computation. Recent optogenetic studies appear to confirm the presence of circuit induced gain, in the input layers of mouse visual cortex [79], [80]. Taken together these experimental and theoretical results support the hypothesis that at least some fundamental WTA functionality is embedded in the processing architecture of superficial neuronal circuits, and so makes the WTA motif a worth target of the developmental process that we have described here.
Our model predicts that neurons form specific subgroups, or cell assemblies [81], [82]. There is indeed strong evidence from biological data for this clustered connectivity [83]–[85], which (as in our simulations), appears to be grounded in the similarity of functional selectivity [86].
We did not allow dynamic rearrangement of synapses in these first simulations. However, it is plausible that weak synapses are pruned away, freeing synaptic resource to explore for more correlated partners.
Peters' rule [87]–[89] proposes that connectivity can be estimated by the product of the random overlap of pre- and postsynaptic sites. This rule may be true for average connectivity, but specific functionality obviously calls for more specific low level connectivity within the average. One opinion is that such specificity is explicitly genetic, and so accounts for example for the diversity of cortical interneurons [90], [91]. Instead, our result speaks for an implicit rather than explicit genetic specificity. That is, the apparently specific wiring of the WTA network arises by neurons collectively satisfying genetically expressed constraints. This concept is in stark contrast to the view that network functionality emerges from individual processes that do not coordinate with potential interaction partners. In our simulations, a neuron's morphology and the functional strengths of its synapses depend on the collective behavior of the other neurons. Hence, the structure and function of a neuron grown in isolation is different from a neuron with the same genetic code, but that interacts and coordinates with its environment during development.
Our learning rule requires that input projections are ordered in such a way that their collective input patterns provide at least a coarsely structured signal against which the presumptive WTA layer of neurons can successfully deploy a BCM-like learning mechanism. This ordering is not a stringent requirement. For example, provided that there is some degree of coherent axonal mapping of axons from input neurons of the subplate or thalamus into the target WTA layer, then even metabolically induced travelling waves of activity across the developing input population could provide a sufficiently structured signal for learning.
Traditionally, many modeling studies have been based on the assumption that the limited lateral extent of the neuronal axonal and dendritic tree naturally leads to a properly configured 2D neighborhood topology [92]–[94]. However, it is unclear how more realistic anatomical properties (anisotropy, variation of neurite extent, irregular locations of somata etc.) affect these topologies. Our work addresses this problem by demonstrating how neurons can self-calibrate in a stimulus-induced way, within a non-uniform and irregular neuronal setting. Hence, our work provides a better understanding of how developmental mechanisms can generate a neighborhood topology, and so is complementary to the classical approach.
As development of input neurons proceeds, the degree of structuring is likely to improve also, so that input neurons projecting to the same targets share similar features (for example, their ON- and OFF-subfields). This is in line with studies on thalamo-cortical projections [95], as well as cortico-cortical projections from layer IV to II/III [96]. However, it should be noted that this input specificity does not play onto inhibitory targets, which is in accordance with our work. Since the input to the neurons shapes the functional connectivity in the network, it follows from our model that neurons which receive common input are more likely to connect with each other (assuming that structural connectivity is adjusting to functional connectivity). The studies of [96] and [97] provide evidence for this input-dependent intra-network specificity.
Our results predict that only a few inhibitory neurons provide the major part of WTA-relevant inhibition, i.e. a relatively small subset of all the inhibitory basket cells projecting to a single pyramidal cell is responsible for its WTA suppression. These results suggest that WTA inhibition might not be very redundant, so that de-activation of only a few inhibitory neurons could result in very different electrophysiological behavior of pyramidal cells.
Our networks employ Hebbian-type learning for both excitatory and inhibitory synapses onto excitatory postsynaptic neurons. It is known that inhibitory synapses can undergo long-term potentiation (LTP) as well as long-term depression (LTD) [52]–[54], and learning by inhibitory synapses has been used in previous modeling studies [45], [98].
Non-Hebbian synaptic scaling of synapses onto inhibitory neurons results in orientation-nonselective inhibitory neurons. This distinction with respect to pyramidal neurons has been observed in mouse visual cortex, where the tuning of inhibitory neurons is broader than that of excitatory neurons [63]–[67], [99]–[101]. There is evidence for broadly tuned thalamo-cortical input to inhibitory neurons [95], as well as cortico-cortical input to those of layer II/III of mouse cortex [100]. Therefore we propose that at least some types of inhibitory neurons (e.g. fast-spiking (FS), PV-expressing interneurons) do not selectively adjust their inputs, but uniformly adapt the electrophysiological properties of their inputs for homeostasis.
Orientation-selective inhibitory neurons are found in cat visual cortex [69]–[72]. Since we do not model orientation maps, our findings are not directly applicable to the cat. However, we argue that it is the spatial location in the orientation map that determines the tuning curve of inhibitory neurons. Most cortical interneurons have a small horizontal dendritic extent [56], and so they likely receive inputs from similarly tuned excitatory neurons within an orientation map. Inhibitory neurons located close to orientation pinwheels are expected to have relatively broad orientation tuning, as reported in the above studies. The unbiased pooling of surrounding activity by inhibitory neurons is also supported by experimental results across species and sensory modalities [63], [101]. By contrast, we have shown that inhibitory neurons become orientation-selective when they follow the same (BCM) learning rule as excitatory neurons.
Our learning model provides a computational explanation for why most interneurons are smooth, i.e. have very few dendritic spines. It is believed that spines, by compartmentalizing biochemical signals, provide the molecular isolation required for independent synaptic learning [102]–[104]. The nonspecific and homogeneous adaptation of inhibitory neurons, which in our model are homogeneously scaling the input efficacies, is therefore well in line with this suggested function of dendritic spines. This model also provides an explanation for the finding that inhibitory, but not excitatory neurons exhibit structural remodeling of dendrites in the adult rat [105]. Changes in excitatory morphology at the level of dendritic branches (rather then spines) could have detrimental effects on already consolidated memories. Inhibitory neurons may retain their potential for dendritic restructuring, because their homeostatic adaptation does not interfere with learning of sensory experience.
We believe our findings to be robust also with respect to models incorporating spikes, because the main features of the adaptation and learning behavior have been demonstrated also on this more detailed level of electrophysiology. Along these lines, the studies of [106], [107] have explored spike-based WTA network functionality. Spike-dependent plasticity (STDP) is a Hebbian learning rule [108] and can yield synaptic homeostasis [109]. In particular, the BCM learning rule has been related to STDP mechanisms [109]–[112].
The robust self-organization of the WTA network is remarkable in that it arises out a single precursor cell, by simple genetically encoded rules. In future, this genetic developmental approach to functional circuit construction could be extended to larger networks composed of multiple WTA networks. For example, it has been hypothesized that by cooperation of multiple WTA circuits, the superficial layers of cortex could perform context-dependent processing [8]. Along these lines, [78] provide a model for the development of long-range projections connecting multiple columns, arranged on an hexagonal grid, as is observed in the superficial patch system [113]–[116]. It also remains to integrate these computational aspects into the context of a laminated cortical structure, which has already been simulated in Cx3D [24], [77].
The GRN is defined by a set of variables that represent genes and the corresponding substance concentrations. Changes in substance concentration are described by the rate equation:(6)
where is the concentration of a protein encoded by the gene (i.e. or ), and the corresponding concentration vector. The function expresses how the synthesis rate of the protein encoded by gene depends on the cooperative binding of all the substances, and , represent the production and degradation rates (, ). is a vector of Hill functions, which compute the binding probability of a substance to a regulatory region given the affinity constant , cooperativity and binding bias :(7)
Gene substances can regulate gene expression by binding to specific sites in the genomic cis-regulatory regions. Substances that regulate each others' transcription are called transcription factors. Many genes are controlled by a number of different transcription factors and different arrangements of binding sites can compute logic operations on multiple inputs. Here, the function takes the form of a logical combination of interacting substances and is defined by the elementary operations:(8)(9)(10)More information on this description of GRN dynamics can be found in [117], [118]. Although abstract, this formalism can be directly translated into the corresponding mechanistic, kinetic differential equations. For our computational model based on 5 genes, we have used the following equations:(11)(12)(13)(14)(15)
with:(16)
The probabilities of either differentiating into neurons by induction () or by induction () are computed as follows:(17)(18)(19)(20)
where is the number of divisions in the first division cycle, is the difference between the target number of neurons () and the number of neurons resulting from the first division cycle, and denotes the floor function for rounding to integers.
The intrinsic production constant determines the number of cell divisions until differentiation into excitatory and inhibitory neurons can occur. The higher it is, the faster the gene reaches the threshold of 0.99. was adjusted manually in order for divisions to occur in the cycle.
Initially, neuronal cell bodies are assigned uniformly random positions in 3D unprepared space. In Cx3D, these cell bodies are modeled as physical spheres. The neuronal cell density was in agreement with values derived from experimental data, i.e. in the range of 40'000 to 86'900 per mm [119]–[121]. We found 250 neurons (200 excitatory and 50 inhibitory) to be appropriate for the available computer resources. For the establishment of neuronal connectivity, the somata were placed randomly in a cube with side length 160 m. A smaller network of 150 neurons in a cube with side length 140 m was used for simulations where the second developmental phase was simulated, in order to decrease simulation time. 3 of these 150 neurons did not get inhibitory inputs after the initial outgrowth and were not included for the simulation of learning, such that the analyzed network consisted of 117 excitatory and 30 inhibitory neurons. Standard Cx3D parameters for the physical properties of the cells (e.g. mass or adherence) were used [16]. The somatic diameters were set to 8 m. Variation of these parameters had only minimal effects on the simulation results.
Axonal and dendritic growth were encoded with the instruction language G-code [24]. We used the following mechanisms, which are executed by such G-code “modules” located in the growth cone, for axonal and dendritic growth, as well as synapse formation:
The computation of the electrical activity was implemented in Java, to allow a direct interfacing with Cx3D. All the synaptic weights in the Cx3D simulation are summed up, which yields a weight matrix. Based on this weight matrix, the input activity and the spontaneous activity, the firing rate of a neuron is computed according to Eq. 1. The numerical solution of the differential equations was computed using the explicit Euler integration method. The network's activity is computed with 3000 iterations and integration step . The maximal firing rate is set to 250 Hz.
Analyses of the simulated networks were performed with MATLAB (Mathworks Inc.). In order to assess WTA functionality, electrical activity was computed in the same way as in the Java implementation of the Cx3Dp simulation, namely using the rate-based model (Eq. 1) and the explicit Euler method. The integration step was decreased to for minimizing integration errors.
The ordering of neurons for visualization, such as for Fig. 6A, was done using the genetic algorithm “ga.m” from the Global Optimization Toolbox of MATLAB. The energy to be minimized was defined as the sum of weighted topological distances between neurons, i.e. , where are the summed synaptic weights from neuron j to neuron i. The topological distances are inferred from a discrete 1-dimensional position vector of the neurons, which is initialized randomly and optimized. The ordering for the matrix visualization is then given by the locations of the neurons in this vector (i.e. neighbors in this vector are also neighbors in the matrix ordering). Note that the topological position is unrelated to the physical position of the neurons, and is only used for the optimization process. The visualization of the clustering was done with CytoScape [126], an open-source framework that is downloadable from http://www.cytoscape.org/. We used the “dynnetwork” plugin implemented by Sabina Pfister, which clusters weighted networks based on the Kamada-Kawai algorithm [127].
The neurite outgrowth has several parameters, which depend on the neuronal type (excitatory/inhibitory) and also on the neurite type (axon or dendrite). Table 1 lists all these parameters. The 2 substances which are secreted by the cell bodies and used by the axons as guidance cues both have a diffusion coefficient of 50 and a degradation constant of 5.
|
10.1371/journal.pgen.1004163 | A Long-Chain Flavodoxin Protects Pseudomonas aeruginosa from Oxidative Stress and Host Bacterial Clearance | Long-chain flavodoxins, ubiquitous electron shuttles containing flavin mononucleotide (FMN) as prosthetic group, play an important protective role against reactive oxygen species (ROS) in various microorganisms. Pseudomonas aeruginosa is an opportunistic pathogen which frequently has to face ROS toxicity in the environment as well as within the host. We identified a single ORF, hereafter referred to as fldP (for flavodoxin from P. aeruginosa), displaying the highest similarity in length, sequence identity and predicted secondary structure with typical long-chain flavodoxins. The gene was cloned and expressed in Escherichia coli. The recombinant product (FldP) could bind FMN and exhibited flavodoxin activity in vitro. Expression of fldP in P. aeruginosa was induced by oxidative stress conditions through an OxyR-independent mechanism, and an fldP-null mutant accumulated higher intracellular ROS levels and exhibited decreased tolerance to H2O2 toxicity compared to wild-type siblings. The mutant phenotype could be complemented by expression of a cyanobacterial flavodoxin. Overexpression of FldP in a mutT-deficient P. aeruginosa strain decreased H2O2-induced cell death and the hypermutability caused by DNA oxidative damage. FldP contributed to the survival of P. aeruginosa within cultured mammalian macrophages and in infected Drosophila melanogaster, which led in turn to accelerated death of the flies. Interestingly, the fldP gene is present in some but not all P. aeruginosa strains, constituting a component of the P. aeruginosa accessory genome. It is located in a genomic island as part of a self-regulated polycistronic operon containing a suite of stress-associated genes. The collected results indicate that the fldP gene encodes a long-chain flavodoxin, which protects the cell from oxidative stress, thereby expanding the capabilities of P. aeruginosa to thrive in hostile environments.
| Coping with toxic reactive oxygen species (ROS) generated as by-products of aerobic metabolism is a major challenge for O2-thriving organisms, which deploy multilevel responses to prevent ROS-triggered damage, including membrane modifications, induction of antioxidant and repair systems and/or replacement of ROS-sensitive targets by resistant isofunctional versions, among others. The opportunistic pathogen Pseudomonas aeruginosa is frequently exposed to ROS in the environment as well as within the host, and we describe herein a new response by which this microorganism can deal with oxidative stress. This pathway depends on a previously uncharacterized gene that we named fldP (for flavodoxin from P. aeruginosa), which encodes a flavoprotein that belongs to the family of long-chain flavodoxins. FldP exhibited a protective role against ROS-dependent physiological and mutational damage, and contributed to the survival of P. aeruginosa during in vivo infection of flies as well as within mammalian macrophagic cells. Thus, fldP increases the adaptive repertoire of P. aeruginosa to face oxidative stress.
| Microorganisms living in aerobic environments are constantly exposed to the harmful effects of reactive oxygen species (ROS), including H2O2 and the superoxide radical, which are generated as unavoidable by-products of oxygen utilization [1]. In addition, commensal and pathogenic bacteria have to face the host oxidative response, such as H2O2 production from phagocytes [1], [2]. Aerobic organisms have evolved multigenic responses to prevent and/or repair the cellular damage potentially inflicted by these toxic compounds. Whenever defenses are overcome by the amounts of ROS produced, cells are afflicted by a condition called oxidative stress [1], [2]. Protective mechanisms deployed by stressed organisms include regulation of membrane permeability, antioxidant and repair systems, and replacement of ROS-sensitive targets by resistant isofunctional versions. In microorganisms as distantly related as enterobacteria and cyanobacteria, induction of the mobile electron shuttle flavodoxin (Fld) appears to be a common feature of the antioxidant response [3], [4]. Flds contain flavin mononucleotide (FMN) as prosthetic group and are largely isofunctional with the ubiquitous electron carrier ferredoxin (Fd), exchanging reducing equivalents with a promiscuous lot of donors and acceptors. Fld induction is assumed to act as a backup for Fd, which harbors a ROS-sensitive iron-sulfur cluster as redox-active cofactor and whose levels are down-regulated under conditions of environmental stress or iron starvation [5], [6]. Accordingly, Fld overexpression has been shown to confer augmented tolerance toward various sources of oxidative stress in organisms with very different lifestyles, such as Escherichia coli [7], rhizobia [8] and plants [9]. Unlike Fds, which are present in all major kingdoms, Flds are restricted to various groups of prokaryotes and some oceanic algae [10]. From sequence alignments and structural considerations, they can be divided into two classes, short-chain and long-chain Flds, which differ by the presence of a 20-amino acid loop of a so far unknown function [11]. Phylogenetic analyses indicate that the two lineages have diverged only once [12].
Pseudomonas aeruginosa is a free-living bacterium commonly found in soil, water, moist locations, and most man-made environments throughout the Earth. P. aeruginosa has a wide metabolic versatility as the reflection of a large and flexible genome with a substantial number of genes, which facilitates its adaptability to thrive in different habitats, and allows a quick response to diverse environmental stimuli and challenges [13]. This remarkable versatility enables P. aeruginosa to infect damaged animal tissues or immunocompromised individuals, where it constitutes an important opportunistic pathogen highly prevalent in nosocomial infections [14], [15]. Indeed, this bacterium is of particular concern to patients with cystic fibrosis (CF) who are highly susceptible to P. aeruginosa and suffer severe and often fatal chronic airway infections [16].
The present research seeks to identify and characterize putative long-chain Flds in P. aeruginosa which could play a protective role under environmental stress conditions. Both P. aeruginosa and its relative Pseudomonas putida contain a single-copy gene encoding a short-chain Fld [17], [18], annotated as mioC due to the homology of the product with the homonymous Fld from E. coli, but no long-chain Fld orthologs have been so far reported in pseudomonads. We identified a gene, PA14_22540 (hereafter referred to as fldP, for flavodoxin from P. aeruginosa), which displays low but significant sequence homology to the Anabaena and E. coli Fld genes (named isiB and fldA, respectively), and whose recombinant product was able to display Fld activity in vitro. Expression of the fldP gene in P. aeruginosa was induced by H2O2 treatment via an OxyR-independent pathway, whereas its disruption increased H2O2-induced killing and accumulation of intracellular ROS. The mutant phenotype could be complemented by transformation with the isiB gene from Anabaena. Overexpression of fldP mitigated H2O2-induced cell death in a mutT-deficient P. aeruginosa strain as well as the hypermutability caused by DNA oxidative damage. The presence of a functional fldP gene contributed to P. aeruginosa survival in two model systems of infection: cultured mammalian macrophages and Drosophila melanogaster. Improved bacterial endurance in Drosophila resulted in higher death tolls of the infected flies. In line with its presumptive adaptive role, the fldP gene was found to be part of a self-regulated operon belonging to the P. aeruginosa accessory genome, a collection of strain-specific gene clusters which are acquired en bloc and expand the genomic repertoire to fit the needs for survival in adverse environments. Then, the collected results indicate that the fldP gene encodes a long-chain flavodoxin which is induced when P. aeruginosa is under oxidative stress to exert a protective role against the physiological and mutational damage caused by ROS.
We performed an in silico survey of putative Flds in the genome of P. aeruginosa PA14 and compared the retrieved sequences with those reported for Anabaena (IsiB) and E. coli (FldA) long-chain Flds, whose protective roles against oxidative stress have been extensively documented [12]. To find orthologs, the IsiB and FldA sequences were compared to the P. aeruginosa PA14 genome using the Domain Enhanced Lookup Time Accelerated Basic Local Alignment Search Tool (DELTA-BLAST), which is sensitive in detecting remote protein homologs [19]. Four unique open reading frames (ORFs) displaying both sequence homology and similar domain organization were retrieved. They were a putative oxidoreductase with a covalently-linked Fld-like domain (PA14_58560), the repressor binding protein WrbA (PA14_51990), a short-chain Fld similar to E. coli MioC (PA14_19660), and an ORF (PA14_22540, tentatively referred to as FldP), which displayed the highest similarity in length and sequence with the long-chain Flds used as baits. Analysis of ORF PA14_22540 indicated that it would encode a 184-amino acid protein with a molecular mass of ∼20 kDa, which is in the range of those typically observed for long-chain Flds (170–185 amino acids). Multiple-sequence alignment between FldA, IsiB and FldP showed that while FldA and IsiB display 47% identity and 67% similarity, FldP exhibits 23% identity with both flavodoxins, and 50% and 41% similarity with IsiB and FldA, respectively.
By using the JPred3 software we next performed a prediction of the secondary structure of FldP and compared it with those of IsiB (Accession number P0A3E0) and FldA (Accession number P61949). According to this prediction, FldP displays a significant similarity, at the level of the secondary structures, with both FldA and IsiB (Figure 1). The three proteins fit the common scheme of five β-sheets intercalated with five α-helices, as well as the loosely structured region dividing β5 into β5a and β5b (between residues 130 and 150 of FldP), which is typical of the long-chain class of Flds [11], [12].
To determine if the product of the fldP gene is a functional Fld, the coding sequence was expressed in E. coli under the control of an inducible promoter (see Materials and Methods). The recombinant protein accumulated to high levels in the bacterial host, but unlike IsiB, which was readily soluble and assembled its prosthetic group in the E. coli cytosol [20], the P. aeruginosa protein was recovered largely as insoluble inclusion bodies (Figure S1A). Only after the simultaneous expression of a suite of molecular chaperones, a minor but significant amount of the protein could be solubilized and purified by affinity chromatography on Ni-NTA columns (Figure S1B).
The purified protein showed a typical flavoprotein spectrum with absorption maxima at 374 and 446 nm, close to those of free FMN in aqueous solution (Figure 2A). The 446-nm peak, which corresponds to transition I of the flavin, is strongly red-shifted in the flavodoxins from Anabaena and E. coli (Figure 2A), indicating that the environment of the prosthetic group is more hydrophobic and/or less solvent-exposed in these flavoproteins. The flavin moiety of IsiB is sandwiched between the aromatic side-chains of Trp58 and Tyr95, in a coplanar conformation, and the π–π interaction thus established has been regarded as the major cause for the spectral red shift [12], [21]. The two amino acids are conserved in FldA, but not in FldP, where the tryptophan is replaced by a tyrosine and the tyrosine by a leucine (Figure 1). The substitutions will prevent aromatic stacking of the isoalloxazine ring system, and introduce conformational changes in its immediate environment, since both residues are located in flexible loops between 3β-3α and 4β-4α (Figure 1). Then, the absence of red-shifted peaks in the visible absorption spectrum of FldP likely results from a combination of higher solvent exposure and decreased aromatic stacking on the flavin.
FldP was assayed in vitro as a substrate of cyanobacterial ferredoxin-NADP(H) reductase (FNR). Figure 2B shows that purified FldP was able to mediate FNR-driven cytochrome c reduction in a concentration-dependent manner, with an apparent KM of 1.3±0.2 µM and a kcat of 22.0±1.8 min−1. Under similar conditions, IsiB displayed a kcat of 48.5±3.8 min−1 (data not shown). The collected results indicate that the product of the fldP gene displayed the structural and functional properties of a bona fide flavodoxin.
In order to determine the functional role of FldP in P. aeruginosa, we tested the tolerance exhibited by a fldP-deficient mutant strain to H2O2 toxicity. In the absence of stress, the viabilities of the wt and fldP mutant strains were similar (Figure S2), but fldP cells were ∼5-fold more sensitive than their wt siblings to the H2O2 treatment (P = 0.024, Figure 3A). Complementation of this mutant with the fldP gene cloned in plasmid p2 (p2-fldP) led to an approximately 1.4-fold increase in the percentage of surviving cells respect to that observed with the wt strain, although the difference was not statistically significant (P = 0.3929, Figure 3A). Noteworthy, expression of IsiB from the p2-isiB plasmid provided even higher levels of protection against the deleterious effects of H2O2, with a rise of 2.2-fold relative to the wt strain (P = 0.125, Figure 3A).
Taking into account these observations, we further tested whether FldP may be involved in the control of intracellular ROS build-up, presumably the cause of H2O2-induced killing. ROS accumulation was quantified in bacterial extracts from the wt, the fldP and the complemented mutant strains after exposure to H2O2. Cells from the wt strain displayed a small increase of their ROS (-OOH) levels as the H2O2 concentration was raised (Figure 3B). Although this increase was not statistically significant, it was consistently observed in a number of experiments. On the other hand, lack of a functional FldP led to ROS build-up in the fldP mutant, whereas expression of FldP from a plasmid in the complemented cells decreased the total -OOH levels to those observed in untreated wt bacteria (Figure 3B).
ROS accumulation was also detected in whole P. aeruginosa cells by using the fluorogenic dye 2′,7′-dichlorofluorescein diacetate (DCFDA), and visualized by confocal microscopy. Figure 3C shows that the fraction of labelled cells above the detection threshold was significantly higher in the fldP mutant compared to their wt siblings (33% vs. 14%), but decreased to 2% after complementation with the p2-fldP plasmid, presumably due to the effect of increased genic doses provided by the plasmid.
It has been recently reported that P. aeruginosa strains deficient in the 8-oxodeoxiguanine system (GO) are particularly vulnerable to oxidative stress [22], [23]. Specifically, mutT-deficient cells showed to be the most susceptible to oxidants such as H2O2 or methyl viologen. We therefore used this strain to further characterize the protective activity of FldP against ROS. A mutT P. aeruginosa strain was transformed with either p2-fldP or p2-isiB, and the resulting transformants were tested for their susceptibility to H2O2. Parallel controls were carried out by using the wt and mutT strains harboring an empty p2 plasmid.
As shown in Figure 4, the mutT strain showed a 20-fold decrease in survival after exposure to H2O2, being significantly more susceptible than the wt strain (P = 0.0119). However, when fldP or isiB were overexpressed in the mutT mutant, cell survival increased dramatically (8- and 16-fold, respectively, P = 0.0119), relative to mutT transformed with the empty p2 vector (Figure 4). The results suggest that the antioxidant role of FldP and IsiB can partially compensate the increased susceptibility of mutT-deficient cells to oxidative stress.
DNA damage produced by ROS can lead to increased mutation frequencies [24] and emergence of adaptive phenotypes [25], [26]. Particularly in mutT-deficient mutants, in which the damage produced by oxidative stress cannot be avoided, mutation frequencies can increase 100- to 1000-fold, leading to hypermutator phenotypes [22]. We therefore investigated whether fldP and isiB could display an antimutator effect and alleviate the hypermutability that is typically observed in a mutT-deficient background. Thus, we exposed a mutT strain, which overexpressed fldP or isiB to H2O2 and measured the mutation frequency by determining the emergence of mutants resistant to streptomycin. The mutT strain showed a ∼1200-fold increase in the H2O2-induced mutation frequency (3.44×10−6), relative to the wt (2.83×10−9). Expression of either fldP or isiB in the mutT strain diminished the mutation frequencies to 4.49×10−7 and 2.63×10−7, which represent a 13% (P = 0.05) and 8% (P = 0.0286), respectively, of the mutation frequency showed by mutT cells transformed with p2 (Figure 5). This last result indicates that overexpression of a functional Fld could avoid the majority (∼90%), but not all H2O2-induced lesions produced as a consequence of mutT deficiency. The protective effect presumably results from the antioxidant properties of these flavoproteins. Accordingly, the antimutator effect conferred by both fldP and isiB was not observed when the spontaneous mutation frequency was tested (data not shown).
To gain further insight into the role played by FldP in oxidative stress tolerance, we studied the transcriptional induction of the fldP gene by H2O2 treatment. The expression of fldP was monitored by semi-quantitative, two-step, reverse transcription-PCR (RT-PCR), using expression of the constitutive housekeeping rpoD gene as control for equal amounts of cDNA in each reaction. Figure 6 shows that expression of fldP in the wt strain was clearly induced (3.9±0.7-fold relative to untreated cultures, P = 0.004), after H2O2 treatment.
We further investigated whether this increased expression of fldP was dependent on OxyR, the main oxidative stress response regulator of P. aeruginosa [27]. An oxyR deletion mutant strain of P. aeruginosa was constructed and tested for fldP expression in response to H2O2. Interestingly, no differences were observed in the oxyR strain compared to its isogenic wt, showing a 3.0±0.7-fold induction in the expression of fldP relative to untreated cultures (P = 0.0045, Figure 6). Importantly, we also evaluated expression of fldP in a second oxyR strain (ID54029) from the PA14 insertion mutant library [28], which yielded equivalent results (data not shown). These observations indicate that although fldP is a stress-responsive gene in P. aeruginosa, as in other bacterial species [3], [4], this response is triggered by an OxyR-independent mechanism.
The use of ROS to kill bacterial pathogens, such as H2O2 production from phagocytes, is a common feature of the innate immune response of eukaryotic organisms. Considering that FldP sheltered P. aeruginosa from oxidative stress under in vitro conditions, we tried to determine if this protective role of FldP could also provide an advantage to cope with the host immune defenses.
As a first approach, we evaluated the capacity of the different P. aeruginosa strains to survive in the intracellular milieu of phagocytes by using monolayers of the macrophagic cell line RAW 264.7, which were inoculated with wt P. aeruginosa or its isogenic fldP-deficient strain, complemented or not with p2-fldP. Then, with the addition of antibiotics to kill extracellular bacteria we were able to compare the proportion of bacterial cells of each strain that survived during a 3-h period inside the phagocytes (see Materials and Methods) by lysing the cell monolayer and plating the lysates on LB agar. Figure 7A shows that inactivation of fldP produced a moderate but significant decrease of ∼24% in the intracellular survival of P. aeruginosa in phagocytic cells (P = 0.0103). Importantly, this decrease could be reverted by complementation with p2-fldP, even surpassing the wt values. This result indicates that FldP is contributing to the intracellular survival of P. aeruginosa in macrophagic cells, probably by enhancing bacterial resistance to the ROS produced by phagocytes.
To evaluate if this protective role of FldP could also be observed in the context of a whole organism infection, which is a more complex system than cultured cells, we used the insect-infection model of D. melanogaster, as previously validated to evaluate virulence traits of P. aeruginosa [29], [30]. Thus, flies were fed with the same strains of P. aeruginosa mentioned above and left for periods of 1, 3 or 5 days, after which the ability of each P. aeruginosa strain to survive within the host was scored by counting colony forming units (CFU). Interestingly, the capacity of the fldP mutant to survive within the host dropped to 28 to 36% (P<0.05) at the three time-periods assayed, which could recover to wt values after complementation with p2-fldP (Figure 7B).
These results suggest that increased bacterial loads could have an impact on host survival. To assess this, D. melanogaster flies were starved for 3 h and then continuously fed with P. aeruginosa. The number of surviving flies was monitored every day until all flies died. The results showed that the flies that had been fed with the wt strain of P. aeruginosa died faster than those fed with the fldP mutant (Figure 7C), which is in agreement with the better capacity of the wt P. aeruginosa strain to survive within the host. Complementation of the mutant bacteria with p2-fldP increased the mortality rate to wt values (Figure 7C).
Taken together, these findings suggest that FldP could be playing a role in P. aeruginosa pathogenesis by increasing its resistance to the ROS-dependent clearance carried out by the host immune system, which in turn would result in a more lethal infection due to a higher load of viable bacterial cells in the host.
The P. aeruginosa genome is made up by a mosaic of “core” and “accessory and variable” gene clusters [31], [32]. While the gene composition of the core genome is conserved in almost every strain of P. aeruginosa, genes belonging to the accessory genome are found in discrete patches, referred to as regions of genome plasticity (RGP), which can vary in occurrence and location among strains [33]. By using genome sequence information available online, we evaluated the occurrence of the fldP gene in thirteen P. aeruginosa strains (namely PAO1, PA14, 2192, 39016, C3719, LESB58, PACS2, M18, NCGM2.S1, B136-33, RP73, DK2 and PA7), whose genomes have been completely sequenced and made available at the Pseudomonas Genome Database (www.pseudomonas.com) [34]. The genome of the environmental strain Hex1T [35], which has been recently sequenced (Feliziani et al., unpublished), was also included in the analysis. The survey showed that only six of the fourteen strains (PA14, 39016, NCGM2.S1, B136-33, PA7 and Hex1T) contained the fldP gene, suggesting that it is a component of the P. aeruginosa accessory genome. Indeed, fldP is located in the region of genome plasticity RGP32 [33]. With the exception of the taxonomic outlier PA7, all other strains showed a conserved synteny of RGP32, with fldP (PA14_22540) being the fifth gene in a six-gene cluster (PA14_22500 to PA14_22550 in the PA14 genome) (Figure 8). The DNA sequence of RGP32 is highly conserved among strains PA14, 39016, NCGM2.S1, B136-33 and Hex1T, displaying up to 97% identity. The five ORFs accompanying fldP in RGP32 have been assigned putative functions, based on similarity with known genes (Table S1). Interestingly, RGP32 is flanked by two palindromic sequences (inverted sequence repeats) which could have played a role in the acquisition of this gene cluster. On the other hand, the genome of PA7 only showed orthologs for fldP (PSPA7_2449) with 62% identity, and for the two flanking genes of RGP32 (PSPA7_2618 and PSPA7_2620), both with 61% identity respect to their PA14 orthologs. Importantly, none of these genes share the genomic location observed for RGP32 in the other five strains.
To further investigate the prevalence of fldP, we carried out PCR analysis using conserved primers (Table S2) on a collection of clinical and environmental isolates of P. aeruginosa, which were selected on the basis of being clonally different [36]. While highly divergent fldP versions could have been overlooked by this experimental approach, the high degree of sequence conservation of fldP observed among the characterized P. aeruginosa genomes makes this possibility rather improbable. The analysis showed that ∼23% of clones amplified a fragment corresponding to the fldP gene (Figure S3). Considering this low prevalence, the origin of RGP32 in P. aeruginosa is better explained as a consequence of DNA acquisition through one or more horizontal gene transfer processes rather than a genetic loss in those isolates which do not harbor RGP32.
We subsequently analyzed the transcriptional organization of RGP32 in order to elucidate whether the fldP gene behaved as a single transcriptional unit or showed co-expression with other genes of the same variable region, thereby constituting an operon. We carried out PCRs using cDNA as template and primers designed to amplify fragments containing regions of two neighboring genes. We measured co-expression of those RGP32 genes which shared the same orientation and were located downstream from fldP (fldP to PA14_22500). As shown in Figure 9A, we were able to amplify fragments between fldP and PA14_22530, PA14_22530 and PA14_22520, and PA14_22520 and PA14_22510. In contrast, no PCR fragments could be amplified between PA14_22510 and PA14_22500. Importantly, amplicons were obtained among all adjacent genes, even between PA14_22510 and PA14_22500, when genomic DNA templates were used as positive controls (Figure 9A). These results depict a transcriptional organization of RGP32 structured in one polycistronic operon containing fldP, PA14_22530, PA14_22520 and PA14_22510, and two monocistronic transcriptional units for genes PA14_22500 and PA14_22550. To strengthen further these observations, we performed RT-PCRs to measure transcripts of PA14_22530 and PA14_22500 in cultures which were treated or not with H2O2. Interestingly, expression of PA14_22530 exhibited a 3-fold H2O2-dependent induction, similar to that previously observed for fldP, whereas transcripts of PA14_22500 showed only a small increase (Figure 9B). This result is in line with the transcriptional organization of RGP32 described above, and raises the question as to how this fldP-containing operon could be regulated. It has been previously observed that genomic islands often possess their own regulatory elements. In this sense, gene PA14_22550, which is predicted to encode a LysR family transcriptional regulator, appears as a promising candidate to fulfill this role. We investigated this possibility by using a PA14_22550 null mutant strain (ID53714) [28], and compared the expression of fldP and its response to H2O2 with those previously observed for the isogenic wt strain. Surprisingly, RT-PCR analyses showed that inactivation of PA14_22550 produced a 3.4-fold increase in the expression of fldP (P = 0.0285), even in untreated cells (Figure 6). In fact, treatment with H2O2 only raised this difference to 4.1-fold (P = 0.004). Thus, following the increase observed due to PA14_22550 inactivation, exposure to H2O2 did not produce any further rise of fldP expression (P = 0.4558). Then, the collected results indicate that fldP is part of a polycistronic operon which is repressed by the LysR-like transcriptional regulator present in the same RGP.
Even under optimal growth conditions a low percentage of the electrons involved in cellular redox pathways are diverted to oxygen with concomitant ROS generation. This fraction can be dramatically increased under adverse environmental situations, leading to a condition known as oxidative stress. Moreover, cellular auto-oxidations are not the only source of ROS and oxidative stress; they can also be generated by external redox processes, such as phagocytes and other eukaryotic cells, which douse invading pathogens with H2O2 as a strategy to prevent infection (reviewed in [1], [2]).
A common strategy for coping with ROS damage and oxidative stress is the use of electron shuttles to relieve the excess of reducing equivalents and redox-active compounds. One of the most conspicuous among them is the electron carrier flavoprotein Fld, which has been consistently associated with stress protection in a number of organisms. The long-chain Flds from Anabaena and E. coli (IsiB and FldA, respectively) are encoded by oxidant-inducible genes, and confer increased tolerance to environmental and nutritional hardships [3], [4].
Despite the physiological importance of Fld as adaptive resource, little had been described about these flavoproteins in the versatile opportunistic human pathogen P. aeruginosa. In this work, we report the presence of a long-chain Fld (FldP) in P. aeruginosa and describe its role in the defense of the bacterial cell against oxidative conditions. FldP is encoded by the PA14_22540 gene as a 184-amino acid product sharing overall sequence similarity and common structural signatures with well-characterized flavodoxins such as FldA and IsiB (Figure 1). The fldP gene was cloned, expressed in E. coli and the resulting product purified to homogeneity, displaying the spectral properties and activity of a functional Fld (Figure 2).
We deemed it necessary to confirm that FldP was a functional flavodoxin because short-chain Flds have been already identified in both P. aeruginosa and P. putida [17],[18]. These genes have been annotated as mioC by comparison with their ortholog in E. coli. Pseudomonas MioC behaves as a functional Fld, mediating FNR-catalyzed electron transfer to cytochrome c with a kcat of about 6 min−1 [18], comparable to that displayed by FldP (22 min−1, Figure 2B), but its physiological role remains yet to be determined. Null mutants in mioC did not display growth phenotypes in E. coli [37], [38] or P. aeruginosa [17]. In the latter organism, however, mioC mutations exhibited other pleiotropic effects, including altered response to iron stress, increased production of the extracellular pigments pyocyanin and pyoverdine, and modified resistance to antibiotics [17]. In contrast to the enhanced sensitivity of fldP mutants toward H2O2 toxicity (Figure 3A), P. aeruginosa cells deficient in mioC were more tolerant than the wt to oxidative stress caused by methyl viologen or H2O2 [17].
The presence of different non-redundant Flds in a single organism is not uncommon. The E. coli genome, for instance, contains at least four genes predicted to encode flavodoxins: fldA, fldB, mioC, and yqcA [38]. E. coli Flds engage in different cellular pathways, and in general they cannot be functionally exchanged [38], [39]. Moreover, the fldA gene is specifically induced by H2O2 and redox-cycling oxidants [40], [41], whereas mioC is not [37]. Likewise, FldP and MioC, while displaying essentially the same activity in vitro, appear to play different roles in P. aeruginosa, with FldP contributing to the protection against oxidative challenges, and MioC in the response to iron stress [17].
A search of fldP orthologs in the Pseudomonas genomes available in the Pseudomonas Genome Project (http://www.pseudomonas.com) [34] and the environmental Hex1T strain (Feliziani et al., unpublished) revealed their presence in just six out of fourteen strains. The P. aeruginosa genome is made up of a conserved core component disrupted by numerous strain-specific RGPs, the sum of which conform the accessory and variable genome [31], [33]. The fldP gene is part of the region of genome plasticity RGP32, belonging to the accessory genome (Figure 8). Synteny and genome location are highly conserved in the PA14, PA39016, NCGM2.S1, B136-33 and Hex1T strains, with flanking genes being present in PAO1, suggesting that RGP32 is the result of a common insertion event. In this sense, we identified conserved palindromic inverted repeats flanking RGP32 in all these genomes, which provide circumstantial evidences on the origin and acquisition of this genome block.
While the conserved repertoire of the core genome codes for central functions required for survival and reproduction in any habitat [31], [42], the variable regions may play customized roles in adaptation to particular environments [33], [43]. This singular genome architecture, combining conserved and variable components, bestows P. aeruginosa pangenome upon its capability to handle a broad metabolic potential in order to adapt to the widest range of environmental niches. In this context, fldP, as part of the accessory genome, may provide specialized oxidative stress-related functions that benefit survival under stressful conditions, conferring RGP32 a role as adaptive island.
Indeed, several lines of evidence indicate that FldP is involved in protection against oxidative stress, including (i) the strong induction of the fldP gene in response to H2O2 (Figure 6), (ii) the enhanced ROS build-up and lower survival of fldP null mutants exposed to H2O2 (Figure 3), and (iii) the partial protection conferred by FldP overexpression to P. aeruginosa cells deficient in mutT against the deleterious effects (Figure 4) and increased mutational burden (Figure 5) caused by H2O2 treatment.
P. aeruginosa houses a multifaceted antioxidant response, comprising scavenging enzymes such as catalases and superoxide dismutases, thioredoxins and glutaredoxins, as well as small antioxidant molecules such as glutathione and melanine [44]. OxyR can be considered the master P. aeruginosa oxidative stress adaptive response [27]. Furthermore, recent studies have extended the implication of OxyR in other P. aeruginosa important responses, such as quorum sensing regulation, iron homeostasis and oxidative phosphorylation, also identifying a large number of target genes and revealing a far more complex cellular response than previously envisaged [45]. It was therefore tempting to speculate that fldP could be just another effector gene of this transcriptional regulator. Analysis of fldP expression in two oxyR loss-of-function mutants ruled out this possibility (Figure 6), indicating that induction of fldP in response to oxidative stress proceeds through an OxyR-independent pathway.
Inspection of the gene content and organization of RGP32 suggested another attractive possibility. The fldP gene is the first in a row of five ORFs displaying the same transcriptional orientation and extending up to the left end of RGP32 (Figure 8). This accessory genomic region contains still another ORF (PA14_22550), which is divergently transcribed and encodes a putative protein related to the LysR family of transcriptional regulators (Figure 8). LysR is the prototype for the most extended family of transcription factors in the bacterial world. Although originally described as activators of divergently transcribed genes, subsequent research placed LysR proteins as global transcriptional regulators, acting as either activators or repressors of single or operonic genes (reviewed in [46]). We found that four of the five divergent ORFs (including fldP) were co-transcribed in response to oxidative stress (Figure 9A), therefore constituting an operon. Moreover, inactivation of PA14_22550 led to full induction of fldP in the absence of oxidants, and to H2O2 insensitivity (Figure 6), indicating that the LysR homologue acts as a repressor of fldP and presumably the entire operon. The sixth ORF of RGP32, whose presumptive product is homologous to protein-disulfide isomerases (Table S1), is not part of the operon.
Stress-associated traits are often encoded by loci adjacent to those of other defensive products, especially when they are co-regulated [31], [42]. In RGP32, the ORF immediately downstream from fldP encodes a putative glutathione S-transferase (Table S1). Involvement of this superfamily of conjugative enzymes in the protection against oxidative stress has been extensively documented in all types of organisms (see, for instance [47]). On the other hand, nucleoside-disphosphate-sugar epimerases as that presumably encoded by PA14_22520 have been consistently associated with stress responses in plants [48], fungi [49] and bacteria [50]. Finally, PA14_22510 encodes a putative H protein from the glycine cleavage system (Table S1). These are lipoate-containing carrier proteins which provide reducing power (in the form of thiols) to the multi-enzymatic complex involved in glycine decarboxylation. The gene encoding this protein has been shown to be strongly up-regulated by H2O2 in streptococci [51], and lipoic acid is known to participate in the cellular response against oxidative stress in different organisms [52]. Then, all members of the operon have the potential to contribute to the defense against oxidative stress at different levels. We therefore propose that RGP32 represents a stress-inducible, self-regulated genetic element which confers increased tolerance to oxidative challenges. Under normal growth conditions, expression of RGP32 genes should be repressed by the LysR-type regulator encoded by PA14_22550. Oxidants somehow inactivate this transcription factor allowing induction of fldP and other components of the operon. The mechanism by which oxidants modulate the activity of the LysR-like protein of RGP32 is at present unknown and deserves further investigation.
The protective effect displayed in vitro by FldP against oxidative stress prompted us to evaluate whether this tolerance could be advantageous to P. aeruginosa cells exposed to the oxidative assault of cells belonging to the mammalian immune system. Indeed, wt and complemented P. aeruginosa strains expressing FldP from the chromosome or a plasmid exhibited better survival within infected macrophages relative to an isogenic mutant lacking this flavoprotein (Figure 7A). Cifani et al. [53] have recently shown that the oxidative burst produced by P. aeruginosa-infected macrophages plays a key role in the short-term killing of intracellular bacteria following invasion, strongly suggesting that the protective effect of FldP stems from its antioxidant function as it occurs in vitro.
We also used the insect model system of D. melanogaster (in which the PA14 strains has been shown to be particularly aggressive [29]), to further investigate the importance of this protective effect on a real infection process. In good agreement with the differential sensitivity observed in macrophages, fldP mutant bacteria accumulated to lower levels in D. melanogaster, as compared to the wt or the complemented strains (Figure 7B). Noteworthy, the mortality kinetics was delayed for ∼24 h in the mutant-infected flies, whereas the death rate of the flies infected with the fldP-deficient bacteria complemented with p2-fldP was equivalent to that of the wt strain (Figure 7C). Thus, the oxidative killing of P. aeruginosa within Drosophila hemolymph may involve mechanisms similar to those utilized by mammalian hosts. While the FldP effect suggests that the electron shuttle could be involved in P. aeruginosa virulence, it is more likely that the different death rates observed in Figure 7C are a consequence of longer persistence of FldP-containing bacteria in the fly due to the adaptive advantages conferred by the flavoprotein. In line with this proposal, mutation of the P. aeruginosa oxyR gene had similar effects on bacterial survival and Drosophila killing as those reported here [54].
Then, our data identify an oxidant-responsive long-chain flavodoxin in P. aeruginosa, which participates in the defense against ROS and contributes to the bacterial tolerance to the oxidative onslaught elicited by the host immune system. We anticipate that studies on this direction will lead to a more comprehensive panorama of the mechanisms allowing this opportunistic pathogen to adapt and persist in stressful and dynamic environments.
P. aeruginosa PA14 and its isogenic strains ID38939 (PA14_22540 mutant, fldP), ID53714 (PA14_22550 mutant) and ID54029 (PA14_70560 mutant, oxyR) were kindly provided by Dr Eliana Drenkard and Dr Jonathan Urbach from the Massachusetts General Hospital, Boston, USA [28], whereas P. aeruginosa MPAO1 and its isogenic mutT strain were provided by Dr Michael Jacobs from the University of Washington Genome Center, USA [55]. Insertion of the MAR2xT7 mini-transposon in mutant ID38939 is unlikely to have polar effects on the expression of downstream genes since the consecutive aacC1 promoter is oriented in the same direction. To prepare inocula, bacteria were routinely cultured on Luria-Bertani (LB) agar plates from frozen stocks and subcultured overnight in LB liquid medium at 37°C with shaking at 220 r.p.m. Antibiotics were used at the following concentrations: 30 µg ml−1 gentamicin (Gm); 250 µg ml−1 kanamycin (Km).
To find Fld homologs in P. aeruginosa PA14, the Anabaena IsiB and E. coli FldA sequences were compared against the entire PA14 genome using the DELTA-BLAST Search Tool [19]. Multiple DNA sequence alignments were performed by using ClustalW (http://www.clustal.org). Secondary structures were predicted using the Jpred3 software provided by the Dundee Scotland University (http://www.compbio.dundee.ac.uk/www-jpred).
A DNA fragment containing the entire coding region of the fldP gene (PA14_22540) was amplified by PCR from PA14 genomic DNA using oligonucleotides FldP-F and FldP-R (Table S2), containing BamHI and HindIII restriction sites, respectively. The PCR product was ligated to the pGem-T Easy vector (Promega) and subsequently cloned into the broad-host-range plasmid pBBR1MCS2 (p2), which harbors a Km resistance marker [56], to generate p2-fldP. A similar strategy was employed to prepare p2-isiB containing the Fld-encoding gene from Anabaena PCC7119, originally cloned in pEMBL8-isiB [20], using oligonucleotides IsiB-F and IsiB-R as forward and reverse primers, respectively (Table S2). The resulting plasmids (p2-fldP and p2-isiB) and the empty p2 vector (as control), were introduced in the different P. aeruginosa strains via electroporation [57].
Complete deletion of the oxyR gene was carried out as previously described [57]. All primer sequences are described in Table S2. Briefly, a first round of three PCR reactions was performed in which the 5′ and 3′ flanking regions of oxyR, as well as a Gm resistance cassette were amplified from plasmid pPS856 [58] using four gene-specific primers (Oxy-UpF-GWL, Oxy-UpR-Gm, Oxy-DnF-Gm and Oxy-DnR-GWR) and the common Gm-specific primers (Gm-F and Gm-R). This generated three fragments with partial overlaps either to each other or the attB1 and attB2 recombination sites. The purified fragments were then assembled in vitro by overlap extension during the second round PCR using the common primers GW-attB1 and GW-attB2. This resulted in an oxyR-deletion-mutant PCR fragment which was subsequently cloned into pDONR221 (Invitrogen) via the BP clonase reaction to create pDONR221-oxyR::Gm. This construct served as the substrate for LR clonase-mediated recombination into the destination vector pEX18ApGW. The resulting suicide vector pEX18ApGW-oxyR::Gm was then transferred to P. aeruginosa and the plasmid-borne oxyR-deletion mutation was exchanged with the chromosome via homologous recombination to generate the chromosomal deletion mutant.
For expression in E. coli, a 568-bp fragment encoding the complete sequence of the fldP gene was obtained by PCR amplification, using p2-fldP as template, and primers Rec-FldP-F and Rec-FldP-R, which contain restriction sites for NdeI and HindIII, respectively (Table S2). The amplified fragment was digested with the corresponding enzymes, cloned into compatible sites of pET-TEV (Novagen) under the control of the T7 promoter, and fused in-frame to an N-terminal His-tag. To improve solubility, BL21 E. coli cells were co-transformed with this vector and plasmid pG-Tf2 (Takara Bio Inc) expressing E. coli molecular chaperones (GroEL, GroES and Trigger Factor). After induction with 0.2 mM isopropyl-β-D-thiogalactoside (IPTG), the soluble flavoprotein was purified from cleared lysates in a Ni-NTA column by elution with 500 mM imidazole. Expression and purification of IsiB were carried out according to Fillat et al. [20]. UV-visible spectra of the purified recombinant proteins and FMN were recorded in 50 mM Tris-HCl pH 8.5.
The ability of both electron shuttles, FldP and IsiB, to mediate the cytochrome c reductase activity of Anabaena FNR was assayed according to Shin [59]. The reaction mixture contained 3 mM glucose 6-phosphate, 0.3 mM NADP+ and 1 unit ml−1 glucose 6-phosphate dehydrogenase (G6PDH, to generate NADPH), 0.5 µM FNR, 50 µM equine heart cytochrome c and various amounts of Fld in 50 mM Tris-HCl pH 8.5. Cytochrome c reduction was followed at 30°C by the increase in absorbance at 550 nm (ε550 = 19 mM−1 cm−1).
A modified version of the FOX II assay [60] was used to quantify the presence of peroxides in bacterial extracts. Cultures of the parental PA14 and the fldP mutant strains transformed with either p2 or p2-fldP were grown aerobically in LB broth at 37°C for 5 h with the appropriate antibiotics. Then, H2O2 was added to final concentrations of 0, 25 and 50 mM, and bacterial suspensions were incubated for 30 min with vigorous shaking. Cultures were split into two equal portions; one of them was used to measure protein concentration, while the other was centrifuged, washed with 0.9% (w/v) NaCl and finally resuspended in 1 ml of an 80∶20 ethanol/water solution containing 0.01% (w/v) butylated hydroxytoluene (BHT). Samples were disrupted by sonic oscillation (10 times for 10 sec each, 30% amplitude), centrifuged at 10,000 g for 10 min, and 250 µl of the supernatants were combined with 250 µl of 10 mM Tris-phenyl phosphine in methanol (TPP, a -OOH reducing agent), or with 250 µl of methanol, to measure total oxidants. Mixtures were incubated for 30 min to allow complete -OOH reduction by TPP. Five hundred µl of FOX reagent (100 µM xylenol orange, 4 mM BHT, 250 µM ferrous ammonium sulphate and 25 mM H2SO4 in 90% (v/v) methanol) were then added to each sample, and the absorbance at 560 nm was recorded 10 min after reagent addition. The absorbance differences between equivalent samples with and without TPP indicate the amounts of -OOH, which were calculated using a 0–20 µM H2O2 standard curve. Protein concentrations were estimated in cleared lysates in 50 mM Tris-HCl pH 8.0 by a dye binding method [61], using bovine serum albumin as standard.
For observation at the confocal microscope, DCFDA was introduced into P. aeruginosa cells by electroporation. Briefly, 10-ml overnight cultures of the various strains were collected by centrifugation at 12,000 g for 2 min, washed three times with 1 ml of 0.3 M sucrose and finally resuspended in 100 µl of the same solution. DCFDA (in dimethyl sulfoxide) was added to a final concentration of 500 µM and the suspension transferred to electroporation cuvettes of 1-mm width. Cells were electroporated in an Electro Cell Manipulator 600, BTX electroporation System at 25 µF, 2.5 kV and 200 Ω for 5 ms, diluted with 300 µl of 0.3 M sucrose and divided into two equal fractions. One of them was incubated with 25 mM H2O2 at 37°C for 15 min and the other kept under the same conditions without the oxidant. Seven µl of the suspensions were mixed with 3 µl of the membrane-staining dye FM 4-64 (0.1 mg ml−1) on the surface of a glass plate and the lid was sealed with colorless nail paint. Cells were observed at excitation/emission maxima of ∼515/640 nm in a Nikon TE-2000-E2 confocal microscope. The number of fluorescent spots was determined for each dye by using the Image J 1.46 software.
For H2O2 susceptibility assays, cells were grown to mid-exponential phase (OD600∼0.8), collected by centrifugation and washed with 0.9% (w/v) NaCl solution. Then, bacteria were normalized to ∼108 cells and treated with 50 mM H2O2 for 30 min at 37°C. Cells were centrifuged again, washed twice with 0.9% (w/v) NaCl and finally resuspended in fresh LB medium. Serial dilutions were spotted on LB agar plates and incubated overnight at 37°C to determine viability. Non-treated controls were included. Bacterial susceptibility was expressed as the ratio between the survivals of treated to untreated cells.
For estimation of spontaneous-mutation frequencies, five bacterial colonies of the different P. aeruginosa strains were cultured overnight in 10 ml LB medium at 37°C with shaking at 220 r.p.m. Appropriate dilutions of the cultures were plated on LB agar to determine the total number of viable cells, or on LB agar supplemented with 500 µg ml−1 streptomycin to count the number of streptomycin-resistant cells, after overnight incubation at 37°C. Spontaneous-mutation frequency was determined as the ratio between the number of streptomycin–resistant cells and the number of viable cells.
Determination of H2O2-induced mutant frequency was carried out according to previously described protocols [62], with some modifications. Briefly, five independent bacterial cultures for each strain were grown in LB to mid-exponential phase, collected by centrifugation and washed with 0.9% (w/v) NaCl. Cells were normalized to ∼108 and subsequently treated with 50 mM H2O2 for 30 min at 37°C, centrifuged again and washed twice with 1 ml of 0.9% (w/v) NaCl. An aliquot of treated cells (0.5 ml) was diluted 10-fold in fresh LB broth and cultured overnight at 37°C with shaking at 220 r.p.m. Then, 100 µl of each overnight culture were plated onto LB agar supplemented with 500 µg ml−1 streptomycin, whereas aliquots from appropriate dilutions were plated on LB agar without antibiotic to measure cell viability. The H2O2-induced mutation frequency was calculated as above.
P. aeruginosa PA14 mid-exponential phase cultures (OD600∼0.8) were treated with 50 mM H2O2 for 15 min, centrifuged and subsequently washed twice with 0.9% (w/v) NaCl. Treated and untreated cultures were used to extract total RNA using the RNA Purification Kit (Fermentas). RNA was quantified by UV spectrophotometry, and its integrity was checked by electrophoresis in 1.5% (w/v) agarose gels. Then, 1 µg of total RNA was reverse-transcribed using the QuantiTect Reverse Transcription Kit (QIAGEN). PCR primers were manually designed with the assistance of the Netprimer software (PREMIER Biosoft International, Palo Alto, CA) and evaluated for their specificity with the BLAST program at the NCBI Web site. Specific transcripts were semi-quantitatively measured by RT-PCR using primers FldP-RT-F and FldP-RT-R (for fldP), 22530-RT-F and 22530-RT-R (for PA14_22530), and 22500-RT-F and 22500-RT-R (for PA14_22500). Transcripts of the rpoD gene were amplified with primers RpoD-RT-F and RpoD-RT-R and served as housekeeping controls. All primer sequences are described in Table S2. The optimal number of cycles was determined in advance to evaluate expression in the exponential phase of amplification. Final cycling conditions included a hot start at 95°C for 10 min, followed by 32 cycles at 94°C for 30 sec, 60°C for 30 sec and 72°C for 30 sec, and a final extension at 72°C for 5 min. Specificity was verified by agarose gel electrophoresis. Fold change in gene expression was calculated by measuring band intensities with the Gel-Pro Analyzer Software. No amplification was observed in PCR reactions containing water or non-reverse transcribed RNA as template.
Primers were designed in order to amplify PCR products containing regions of two neighboring genes for those genes of RGP32 which share the same orientation and are located downstream from fldP (Figure 9A). The cDNA to be used as a template for PCR was obtained by reverse transcription of purified total RNA as described above. Thus, primers FldP-F and FldP-30-R were employed to determine the co-expression of genes fldP and PA14_22530 (Table S2). In the same way, co-expression of downstream contiguous genes was analyzed with the following primers: a) 30-20-F and 30-20-R to analyze PA14_22530 and PA14_22520 co-expression; b) 20-10-F and 20-10-R for PA14_22520 and PA14_22510; and c) 10-00-F and 10-00-R for PA14_22510 and PA14_22500 (Table S2). Genomic DNA was used as template for positive controls of the PCR reactions. No amplification was observed in PCR reactions containing water or non-reverse transcribed RNA as template.
Cells of the macrophagic-derived line RAW 264.7 were grown on 12-mm-diameter wells until they reached 95–100% confluence. To prepare inocula, overnight cultures of P. aeruginosa were washed with phosphate buffered saline (PBS) and suspended in Dulbecco's Modified Eagle's Medium (DMEM) at a final concentration of 107 cells ml−1. Bacterial antibiotic protection assays were conducted as previously described with few modifications [63]. Briefly, macrophagic cells were inoculated with the various P. aeruginosa strains by the addition of 500 µl of the bacterial suspension, corresponding to a multiplicity of infection (MOI) of about 20, followed by centrifugation for 10 min at 1000 r.p.m. to facilitate cell contact. After a 2-h incubation period at 37°C, the supernatant was removed and the cells were washed three times with PBS to remove non-associated bacteria. To count intracellular P. aeruginosa, fresh DMEM with streptomycin (1 mg ml−1) and carbenicillin (400 µg ml−1) was added to the cell monolayer and incubated for 2 h to kill extracellular bacteria. Antibiotic bactericidal activity was confirmed by plating 100-µl aliquots from the wells directly on LB-agar. Cells were again washed to remove antibiotics, then lysed by incubation with 0.1% (v/v) Triton X-100 in PBS. At this initial stage (Ti) intracellular bacteria were enumerated by plating serial dilutions of cell lysates on LB agar and counting CFU. To estimate P. aeruginosa intracellular survival, an equivalent set of wells were incubated with antibiotics for an additional 3-h period (Tf), cells were lysed and surviving bacteria enumerated as described above. Intracellular survival was calculated as the ratio between the final and the initial (Tf/Ti) CFU counts. Three wells of cells were used for each strain in each experiment, and all experiments were repeated three times. To rule out the pre-existence of resistant bacteria in the overnight cultures, 10× inocula (5×107 cells) were plated on LB agar containing streptomycin (1 mg ml−1) and carbenicillin (400 µg ml−1), with not a single CFU observed even after 72 h of incubation.
Phagocytes viability was monitored at every time point of the experiment by measuring release of LDH (CellTiter 96 Aqueous Non-Radioactive Cell Proliferation Assay, Promega), which showed an initial decrease of ∼50% in the first 2 h, but remained unaltered after the antibiotic treatment.
P. aeruginosa cells of the various strains were grown overnight, washed and diluted in PBS, 10% (w/v) sucrose, 250 µg ml−1 Km to OD600 = 0.025. D. melanogaster strain W1118 flies (4–5 days old) were starved for 3 h and then fed continuously at 25°C on cottons plugs, which had been previously embedded in the bacterial solution. For each time point, 3 groups of 5 flies were pestle homogenized in 200 µl of PBS and the homogenate was serially diluted in PBS and plated on LB agar containing 250 µg ml−1 Km, to quantify the number of bacterial colonies.
Overnight cultures of each P. aeruginosa strain were diluted in fresh LB to an OD600 = 0.25. This solution was then diluted 10-fold in PBS, 10% (w/v) sucrose, 250 µg ml−1 Km. Groups of 15 D. melanogaster W1118 male flies were starved during 3 h and then placed in vials with sterile cotton plugs, which had been previously embedded in 3 ml of the bacterial suspension. Flies were kept at 25°C and survival was monitored daily. Three groups of 15 flies were used for each condition in three independent experiments.
Statistical analyses were performed using one-tailed Mann-Whitney test appropriate for nonparametric adjustment. When appropriate, one- or two-tailed Student's t test was applied. In all cases, P values less than or equal to 0.05 were considered statistically significant.
The collection of P. aeruginosa clinical isolates used in this work has already been published by Feliziani et al. [36]. In this previous study, the informed consent as well as the approval from an Institutional Research Committee were appropriately evaluated and fulfilled the PLOS ethical standards.
|
10.1371/journal.ppat.1000800 | Fatal Transmissible Amyloid Encephalopathy: A New Type of Prion Disease Associated with Lack of Prion Protein Membrane Anchoring | Prion diseases are fatal neurodegenerative diseases of humans and animals characterized by gray matter spongiosis and accumulation of aggregated, misfolded, protease-resistant prion protein (PrPres). PrPres can be deposited in brain in an amyloid-form and/or non-amyloid form, and is derived from host-encoded protease-sensitive PrP (PrPsen), a protein normally anchored to the plasma membrane by glycosylphosphatidylinositol (GPI). Previously, using heterozygous transgenic mice expressing only anchorless PrP, we found that PrP anchoring to the cell membrane was required for typical clinical scrapie. However, in the present experiments, using homozygous transgenic mice expressing two-fold more anchorless PrP, scrapie infection induced a new fatal disease with unique clinical signs and altered neuropathology, compared to non-transgenic mice expressing only anchored PrP. Brain tissue of transgenic mice had high amounts of infectivity, and histopathology showed dense amyloid PrPres plaque deposits without gray matter spongiosis. In contrast, infected non-transgenic mice had diffuse non-amyloid PrPres deposits with significant gray matter spongiosis. Brain graft studies suggested that anchored PrPsen expression was required for gray matter spongiosis during prion infection. Furthermore, electron and light microscopic studies in infected transgenic mice demonstrated several pathogenic processes not seen in typical prion disease, including cerebral amyloid angiopathy and ultrastructural alterations in perivascular neuropil. These findings were similar to certain human familial prion diseases as well as to non-prion human neurodegenerative diseases, such as Alzheimer's disease.
| Prion diseases, also known as transmissible spongiform encephalopathies, are infectious fatal neurodegenerative diseases of humans and animals. A major feature of prion diseases is the refolding and aggregation of a normal host protein, prion protein (PrP), into a disease-associated form which may contribute to brain damage. In uninfected individuals, normal PrP is anchored to the outer cell membrane by a sugar-phosphate-lipid linker molecule. In the present report we show that prion infection of mice expressing PrP lacking the anchor can result in a new type of fatal neurodegenerative disease. This disease displays mechanisms of damage to brain cells and brain blood vessels found in Alzheimer's disease and in familial amyloid brain diseases. In contrast, the typical sponge-like brain damage seen in prion diseases was not observed. These results suggest that presence or absence of PrP membrane anchoring can influence the type of neurodegeneration seen after prion infection.
| Transmissible spongiform encephalopathies (TSE diseases) or prion diseases are fatal neurodegenerative diseases of humans and animals. These diseases include scrapie in sheep, bovine spongiform encephalopathy (BSE) in cattle, and chronic wasting disease (CWD) in cervids, as well as several human diseases including kuru, Gerstmann-Sträussler-Scheinker syndrome(GSS), and sporadic, familial and variant forms of Creutzfeldt-Jakob disease (CJD) (see [1] for review). TSE diseases are transmissible within a species, but can also cross to new species in some cases. For example, variant CJD appears to be a form of BSE transmitted to humans. In addition, experimental transmission to rodents such as mice, hamsters and, more recently, bank voles [2],[3], has provided numerous models for laboratory research.
In prion diseases brain pathology is characterized by spongiform degeneration of the gray matter together with neuronal loss and gliosis. During disease there is an accumulation in brain of an abnormal partially protease-resistant form of prion protein (PrPres) derived from host-encoded protease-sensitive prion protein (PrPsen). PrPres can be detected by immunoblot or immunohistochemistry, and this detection is often used as an important diagnostic feature of prion disease. PrPres can be deposited in brain either as large fibrillar amyloid plaques and/or as small diffuse punctate deposits of non-amyloid aggregated protein. The diffuse non-amyloid PrPres form is prevalent in many human sCJD cases and most prion disease animal models [4]–[7]. However, both amyloid and non-amyloid forms of PrPres coexist in some human and animal prion diseases [6] [8]–[13], and both forms may contribute to prion disease pathogenesis.
A variety of proteins are capable of forming amyloid deposits in nervous system tissues as well as other organs. Amyloid deposits often displace organ structure resulting in dysfunction and cell death. In cerebral amyloid angiopathy (CAA), associated with Alzheimer's disease (AD) and several genetic CNS amyloid diseases, vascular amyloid deposits can damage the structure of blood vessel walls leading to hemorrhage or thrombosis [14],[15]. However, in AD, oligomeric pre-amyloid Aß aggregates are also thought to have important neuropathogenic effects. For therapy of diseases such as AD and prion diseases, where both amyloid and non-amyloid may be pathogenic, it will be important to understand the contribution of both types of abnormal protein aggregates to the various pathogenic processes present in these complex diseases. Therefore, we focused on the pathogenic effects of PrPres amyloid versus non-amyloid induced by prion infection in mice.
In uninfected animals PrPsen is anchored to the plasma membrane by a glycosylphosphatidylinositol (GPI) moiety [16]. In prion disease PrPres is found on plasma membranes of neurons and other brain cells, where it is associated with morphological membrane changes that are common to different animal TSEs [7] [17],[18]. Membrane attachment of PrP may have an important influence on the prion disease process. To study the role of PrP membrane linkage on pathogenesis of prion disease we previously generated 2 lines of transgenic mice (tg44+/− and tg23+/−), which express PrP lacking the GPI anchor at similar levels and do not express GPI membrane–anchored PrP. Anchorless PrP in these mice is secreted by cells and is not attached to the plasma membrane [19]. After scrapie infection, both lines of transgenic mice developed high titers of prion infectivity and extensive PrPres amyloid deposits in brain at late times after infection; however, typical scrapie clinical signs and gray matter spongiosis characteristic of prion diseases were not seen [19].
In the present paper we studied homozygous anchorless PrP transgenic mice, which expressed two-fold more anchorless PrPsen than the above mentioned heterozygous transgenic mice. In these experiments scrapie infection of homozygous mice produced a fatal clinical disease. However, this disease differed in incubation period, clinical signs, and neuropathology from typical prion disease seen in non-transgenic mice, which express anchored PrP, and thus appeared to be a distinct pathogenic process. Therefore, depending on the presence or absence of anchored PrP, scrapie infection could induce two different fatal brain diseases: PrPres amyloidosis without gray matter spongiosis in anchorless PrP transgenic mice, and diffuse non-amyloid PrPres with gray matter spongiosis in mice with anchored PrP.
Since PrPsen expression is known to influence scrapie incubation period [20] [21], it is possible that low PrP expression might account in part for the lack of clinical scrapie disease in previous experiments using heterozygous tg44+/− and tg23+/− mice [19]. Therefore, in the present study we generated homozygous anchorless PrP transgenic mice from both lines 44 and 23. These mice each expressed two anchorless PrP transgene alleles and no normal mouse PrP alleles.
PrPsen levels were analyzed by immunoblotting in brain homogenates of uninfected transgenic and non-transgenic mice. Homozygous tg44+/+ mice expressed 2-fold higher levels compared to tg44+/− mice (Figure 1A). Non-transgenic C57BL/10SnJ mice homozygous for the PrP gene (Prnp+/+) were used as controls, and these mice expressed 2-fold higher PrP levels than did Prnp+/− mice (generated by crossing Prnp+/+ mice to Prnp-null mice also on the C57BL/10SnJ background (see methods)) (Figure 1A). Quantitative comparisons between transgenic and non-transgenic mice were difficult due to PrP glycosylation differences (Figure 1A). Therefore we compared these mice using PrPsen deglycosylated with PNGase F (Figure 1B). In these experiments Prnp+/− mice had approximately four-fold higher PrPsen levels than tg44+/+ mice (compare lanes 2 vs. 4 and lanes 7 vs. 10). Given the 2-fold difference between Prnp+/+ and +/− mice, tg44+/+ expressed 8-fold lower levels of brain PrPsen than did Prnp+/+ mice. Brain PrPsen levels in tg44+/+ and tg23+/+ mice were indistinguishable (data not shown).
Scrapie-induced clinical disease was analyzed in transgenic and non-transgenic mice using two different scrapie strains, RML and 22L. After intracerebral (IC) inoculation of scrapie strain 22L, Prnp+/+ and +/− mice developed clinical scrapie at 150–165 dpi and 245–260 dpi respectively (Figure 2A). Similar results were also seen using the RML scrapie strain in Prnp+/+ mice (Figure 2B). Prnp+/− mice were not tested with the RML strain. In contrast to experiments with non-transgenic mice, all tg44+/+ and tg23+/+ mice infected with strains 22L or RML developed neurological signs and required euthanasia from 300 to 480 dpi (Figure 2). Homozygous transgenic mice differed from non-transgenic mice in incubation period, duration and progression of clinical signs, as well as gait and postural abnormalities (Table 1). The most obvious signs in homozygous transgenic mice were the presence of a wide-based gait, rear extremity weakness with low posture, and lack of kyphosis. The signs seen in homozygous transgenic mice differed from the narrow-based tippy-toed gait and frequent kyphosis seen in infected Prnp+/+ and +/− mice (Table 1). The gait and postural differences probably reflected different patterns of neurological damage. Overall the differences between non-transgenic mice and homozygous transgenic mice suggested that there might be different pathogenic mechanisms operating in these two scrapie-induced disease models.
In our earlier studies, heterozygous tg44+/− and tg23+/− mice did not manifest the usual clinical signs of scrapie during 600 days of observation after infection with scrapie strains 22L or RML (Table 1) [19]. However, in the present experiments with the experience of observing the clinical signs described above in homozygous transgenic mice, we noted similar clinical signs in the infected heterozygous transgenic mice starting around 480 dpi. Clinical diagnosis in these mice was difficult due the erratic presence of signs in the initial stages, the longer duration of signs, and the possibility of confusion with signs of old age. These mice were euthanized between 480 and 700 dpi, but the indication for euthanasia was primarily debilitation (weight loss, dermatitis, bladder distention, cancer and infections), rather than neuromuscular dysfunction.
PrPres deposition in brain is a major hallmark of prion diseases, and PrPres is often associated with areas of brain degeneration. Therefore brain PrPres levels in scrapie-infected mice were analyzed by immunoblot. As shown in Figure 3, tg44+/+ mice with clinical neurological disease had higher amounts of PrPres at 348–408 dpi than tg44+/− mice had at 567–594 dpi, which was when these mice had to be euthanized due to debilitating signs as described in Table 1. A similar difference was seen between tg23+/+ and tg23+/− mice (data not shown). The difference between tg44+/+ and tg44+/− mice in timing and levels of PrPres correlated with the higher PrPsen expression level seen in homozygous mice (Figure 1A), and appeared to explain the earlier onset and more prominent clinical signs seen in homozygous transgenic mice.
The PrPres detected in these immunoblots had a molecular weight of 19 kD. In our earlier study this band was found to react with an anti-PrP peptide antibody (R20) directed at the C-terminal region of PrP (residues 218–232) [19]. Therefore, there was no evidence for loss of these C-terminal residues as often occurs in human GSS [6],[22].
Interestingly tg44+/+ and tg44+/− mice had higher PrPres levels than did non-transgenic Prnp+/+ or +/− mice, but the non-transgenic mice died earlier than the tg44+/+ mice (Figure 2). This suggested either that amyloid PrPres in tg44+/+ mice might be less pathogenic than the non-amyloid PrPres in non-transgenic mice, or that transgenic mice might be less susceptible to the pathogenic effects of PrPres amyloid due to the absence of membrane-anchored PrP.
Previously we showed that scrapie-infected tg44+/− mice lacked signs of clinical scrapie but had infectivity titers in brain as high as 4.6×108 ID50/gram brain at 120–286 dpi as measured by end-point titration in C57BL/10 mice [19]. In the present experiments we passaged brain from infected tg44+/− (passage 1) mice into other tg44+/− mice (passage 2), and at 512 and 531 dpi we found brain infectivity titers of 1.1–1.3×1010 ID50/gram brain (Table 2). These mice also had brain PrPres levels similar to those shown in other tg44+/− mice (Figure 3). Similar high titers were detected in passage 1 homozygous tg44+/+ mice at 384 dpi (Table 2). Thus in the transgenic anchorless PrP model, scrapie infectivity was present in brain at very high titers. Furthermore, the agent did not appear to develop new strain-like properties selective for tg44+/− mice as it could passage easily from tg44+/− mice to either C57BL/10 or tg44+/− mice.
We compared the PrPres deposition and neuropathology after scrapie infection in transgenic tg44+/+ mice and non-transgenic C57BL/10 control mice (Table 3). Following scrapie infection in C57BL/10 mice typical TSE-specific diffuse deposits of PrPres were found in many brain areas (Figure 4A, 4B). This PrPres did not stain with the amyloid stain, Thioflavin S [19]. In many brain regions by H&E staining we observed gray matter spongiosis (Figure 4C), which is an important feature of TSE/prion diseases.
In scrapie-infected tg44+/+ mice PrPres accumulated as large dense plaque-like deposits, usually in a perivascular location around capillaries, veins and arteries in numerous brain regions, including leptomeninges, cerebral cortex, corpus callosum, forebrain, hippocampus, thalamus, hypothalamus, midbrain, colliculi, brainstem, and spinal cord (Figure 4D, 4E, 5A, 5B) After infection with the 22L scrapie strain, the cerebellar molecular layer and granular layer were also involved [19], but this was not seen after infection with the RML strain (Figure 4D). These deposits were Thioflavin S-positive [19], and no areas of diffuse non-amyloid PrPres were observed. The most distinguishing histopathological feature in tg44+/+ mice at the time of clinical signs was distortion of brain structures adjacent to large amyloid plaques in many areas (Figures 4D, E, H–L). These areas had intense micro- and astrogliosis (Figures 4G). Small blood vessels showed occasional micro-hemorrhages, or perivascular haemosiderin accumulation, but no lymphocyte infiltration of blood vessel walls was detected. Marked neuronal loss was seen around edges of some gray matter plaques (Figure 4H, I); however, no gray matter spongiosis typical of prion diseases was seen (Figure 4C vs 4F). In addition, tg44+/+ mice had focal areas of abnormal staining of amyloid precursor protein (APP) (Figure 4K), non-phosphorylated neurofilament protein (NFP) (Figure 4L), and phosphorylated NFP (not shown), all of which indicated a process of severe axonal dystrophy. These latter effects were rarely seen in scrapie-infected C57BL/10 mice (Table 3). These results suggested that scrapie-infected anchorless PrP transgenic mice had a different pathogenic process compared to non-transgenic C57BL/10 mice.
For higher resolution of details, scrapie-infected tg44+/+ mice were also studied using immunohistochemistry on 1 micron thick plastic-embedded sections as well as immunogold labeling at the ultrastructural level. Light microscopy on thin sections showed both perivascular and vascular PrPres labeling (Figure 5A), as well as occlusion of vessels in some cases (Figure 5B). By electron microscopy abundant PrPres labeling of blood vessels was seen most predominantly at basement membranes. In some larger vessels smooth muscle cells of the media were atrophied and replaced by extensive PrPres accumulation (Figure 5C). Vascular and plaque PrPres accumulation could be seen to be of a fibrillar amyloid nature at high magnification (Figure 5D). In smaller vessels PrPres was seen at both endothelial and pericyte basement membranes (Figure 5E). PrPres was also observed within the extracellular space along the borders of swollen astroglial and neurite processes in the absence of visible fibrillar amyloid (Figure 5F and G). No PrPres labeling was seen in uninfected control mice.
Using staining with uranyl acetate/lead citrate large areas of distended swollen processes could be seen (Figure 6A), which were similar to the areas of immunogold-labeled PrPres shown above (Figure 5G). At higher magnification swollen perivascular glial processes were often seen (Figure 6B), and fibrils were visible in the endothelial basement membrane (Figure 6C) and/or pericyte basement membrane (Figure 6D). The initial site of aggregation into fibrils was in the ablumenal basement membranes (Figure 5D). Dystrophic neurites were also frequently noted in gray matter (Figure 6E). These were most conspicuous surrounding perivascular amyloid plaques and corresponded to sites of APP labeling. In white matter we observed degeneration of axons, including empty distended myelin sheaths (Figure 6F) (Table 3) which could be seen as white matter vacuoles by light microscopy (Figure 4F).
In contrast to transgenic mice, infected C57BL/10 mice at the time of clinical disease had numerous TSE vacuoles with broken or “hanging” membranes (not shown) [23] [24]. Such vacuoles were never seen in infected transgenic mice (Table 3). In C57BL/10 mice other ultrastructural hallmarks specific for classical prion diseases including membrane accumulation of disease-specific PrPres and TSE-specific membrane alterations were also seen, as reported previously in other prion disease models [17] [7],[18],[24] (Table 3). However, none of these prion disease-specific features was seen in infected tg44+/+ mice. The ultrastructural differences between scrapie-infected C57BL/10 and tg44+/+ mice supported the conclusion that the pathogenesis of disease in these transgenic mice was not typical TSE/prion disease.
The reasons for the different types of scrapie-induced pathogenesis in C57BL/10 mice and anchorless PrP transgenic mice are not known. Two possibilities include: first, PrPres amyloid and diffuse non-amyloid PrPres might have different neurotoxic effects; second, PrPsen anchoring might influence neurotoxicity induced by infection.
To test whether PrPres derived from GPI-anchored PrPsen could induce gray matter vacuoles in tissue expressing anchorless PrPsen, brain tissue from C57BL/6 mice at embryonic day E12–E14, which expressed green fluorescent protein constitutively in all tissues [25], was grafted into the brain of adult tg44+/− mice or PrP null mice as controls [26]. One month after grafting, mice were infected IC with scrapie, and at 132–511 dpi the brain tissue was examined by histopathology. Recipients had from 1–6 detectable grafts per mouse (Table 4). Representative grafts are shown in Figure 7.
At 261 dpi in the control PrPnull recipient, C57BL/6 graft tissue, identified by presence of green fluorescent protein (GFP) (Figure 7A), had easily detectable PrPres (Figure 7B) present both within the graft and at the interface between the graft and the host tissue, but PrPres did not appear to spread extensively into the PrPnull tissue. TSE gray matter vacuolation was seen only within the graft tissue (Figure 7B and 7C). This was similar to a previous report [26].
In tg44+/− recipient mice receiving C57BL/6 grafts, PrPres and gray matter vacuolation was also seen in the graft (Figure 7F and 7I). In the adjacent host tissue expressing anchorless PrP, amyloid PrPres and white matter vacuoles were noted; however, the C57BL/6 PrPres present at the edges of the graft appeared to be unable to induce gray matter vacuoles in the adjacent transgenic tissue (Figure 7F). In some cases the graft cells were not well-demarcated from the host (Figure 7G), and it was not clear whether the PrPres and vacuoles were in the graft or the host (Figure 7H and 7I). These results were representative of observations in 25 grafts in tg44+/− recipients where PrPres was detected in the graft (Table 4). In summary, we found no grafts where expression of anchored PrPres from the C57BL/6 graft could be associated with gray matter spongiosis in adjacent transgenic host tissue. This result suggested that expression of anchored PrPsen in gray matter might be a fundamental requirement for the induction of the typical TSE/prion disease pathogenic process.
In the present experiments scrapie infection of transgenic mice expressing anchorless PrP resulted in a slow fatal brain disease. These results demonstrated new mechanisms of prion-induced pathogenesis associated with the presence of PrPres amyloid and the absence of GPI-anchored PrP. This disease lacked gray matter spongiosis and differed in this respect from scrapie infection in non-transgenic mice, where the disease is characterized by extensive gray matter spongiosis and non-amyloid PrPres deposition.
The current results raised the question of how lack of GPI-linked membrane anchoring of PrP might facilitate formation of PrPres amyloid. GPI anchorless PrP has a longer biological half-life [27] and is secreted by the cell. Both of these attributes might allow more effective and extensive interactions between soluble PrP molecules. In addition, the minimal amount of carbohydrates and the absence of the GPI group on anchorless PrP might favor amyloidogenic hydrophobic protein-protein interactions, particularly at a time of partial protein unfolding during PrP conversion. These features of anchorless PrP are likely to contribute to its enhanced tendency to form amyloid during conversion to PrPres. Anchorless protease-resistant PrP, cleaved at residue 228, comprises 15% of the PrPres in hamster scrapie brain extracts [28], but it is unclear whether this material contributes to the amyloid PrP seen in this model.
Our results differed from those of two interesting mouse prion disease models where PrPres was also found almost entirely in an amyloid form. In the GSS PrP-8kd model [29] and the G3-ME7 model [30], which both used PrP mutant mice, PrP amyloid was seen primarily in the corpus callosum, but did not spread significantly to other brain regions. There was no clinical disease in these models, and transmission experiments suggested very low infectivity titers in the GSS PrP-8kd model. Compared to these two models, the three main distinguishing features of the anchorless PrP model are the ability of the PrPres amyloid to accumulate widely throughout the brain (Figure 4), the resulting fatal brain disease (Figure 2), and the high titer of transmissible agent (Table 2) [31].
The association of amyloid deposition without gray matter spongiosis in our system is reminiscent of the neuropathology seen in certain human familial prion diseases. For example, GSS patients with PrP mutations Y145Stop and Y163Stop had both CAA and parenchymal perivascular amyloid without gray matter spongiosis [32] [33]. Both these mutations result in C-terminally truncated PrP lacking the GPI anchor. Parenchymal amyloid deposition without gray matter spongiosis has also been seen in GSS patients with several other PrP mutations including P102L, P105L, A117V and F198S [6]. Recently two human GSS patients with new PrP mutations producing nonsense codons at positions 226 and 227 were described [34]. Both patients had widespread PrPres amyloid deposition in the absence of gray matter spongiosis, and one had CAA. These patients expressed a nearly full-length form of PrP lacking 6–7 C-terminal residues and the GPI anchor, which was quite similar to the PrP expressed in our anchorless PrP tg mice.
In many GSS patients, amyloid PrPres purified from brain was truncated resulting in a 7–11 kDa protease-resistant fragment from the central region of PrP (approximately residues 81–150)[6],[22],[34]. Interestingly, presence of this truncation has been correlated with the lack of gray matter spongiosis [8],[9]. In contrast, based on previous immunoblot studies, the proteinase K-resistant PrPres amyloid in our model appeared to contain residues 88–231 [19], which was similar to the PrPres found in human and animal prion diseases with extensive gray matter spongiosis. Furthermore, PrPres in tissue sections could be stained with anti-PrP serum R24, specific to residues 23–37 (data not shown) suggesting that there was no significant truncation at the N-terminus beyond the signal peptide. Thus, lack of spongiosis in our model appeared dependent on the absence of GPI-anchoring rather than truncation of the PrPres.
Two possibilities might explain the correlation between lack of GPI- anchored PrP and lack of gray matter spongiosis in our infected transgenic mice: (1) anchorless amyloid PrPres might be less neurotoxic than diffuse PrPres, and/or (2) anchored PrPsen might be required for PrPres-mediated neurotoxic membrane interactions. The former explanation could not be proven or excluded by our results. However, the latter interpretation was supported by data from brain graft experiments. After scrapie infection of tg44+/− mice grafted with C57BL/6 brain expressing normal anchored PrPsen, we observed gray matter spongiosis and non-amyloid PrPres deposition in C57BL/6 grafts, but not in adjacent host tissue expressing only anchorless PrPsen. Tissue expressing only anchorless PrPsen appeared to be unable to respond to the presence of GPI-anchored PrPres produced in the nearby grafts, and no gray matter spongiosis was produced. Therefore, lack of anchored PrPsen might by itself explain the lack of gray matter spongiosis in transgenic mice.
However, even in the absence of anchored PrPsen, the amyloid PrPres was able to induce additional pathogenic processes capable of causing fatal neurological disease. By both light and electron microscopy we observed evidence for three distinct pathogenic processes not seen in typical prion disease in C57BL/10 mice (Box 1):
(1) Brain damage caused by tissue distortion by large amyloid plaques. These plaques were associated with neuronal loss, axonal pathology and gliosis (Figure 4E, H–L Figure 6A,E,F). The more rapid accumulation of PrPres in tg44+/+ mice compared to tg44+/− mice (Figure 3) suggested a faster growth of large space-occupying plaques which might explain in part the clinical neurological signs leading to death of tg44+/+ and tg23+/+ mice (Figure 2).
(2) A second pathogenic process in scrapie-infected transgenic mice was suggested by ultrastructural studies finding that the early aggregation of PrPres into fibrillar amyloid was located at or within vascular basement membranes (Figures 5C, 5D, 6C, 6D). This was associated with vascular damage including occlusion (Figure 5B), amyloid replacement of basement membrane and tunica media, and occasional micro-hemorrhages. This pathology was similar to that observed in CAA seen in Alzheimer's disease and several familial amyloid diseases including two prion diseases [14],[15],[32],[33].
(3) Evidence of a third pathogenic process in the transgenic mice was suggested by finding of small deposits of immunogold-labeled PrPres at the ultrastructural level in the extracellular spaces between glial and neuritic processes in gray matter (Figures 5D, 5E, 5F). These PrPres deposits were small, and there was no distortion of the extracellular space or visible aggregation into amyloid fibrils. However, the adjacent processes were often highly dystrophic (Figure 6E) or swollen and devoid of organelles, and they appeared to coalesce to form empty spaces larger than the original processes (Figures 5G, 6A, 6B). These abnormal areas, which were also noted in heterozygous tg44+/− and tg23+/− mice, appeared to represent a form of damage related to small, rather than large, PrPres deposits, and they did not require the presence of anchored PrPsen for their formation.
The early localization of PrPres at basement membranes (Figures 5A, C, D), suggested that the PrP conversion process might initiate at these sites, and implied that basement membrane molecules might facilitate PrP conversion. For example, basement membrane might filter or trap soluble PrPsen molecules or small PrPres oligomers from the extracellular interstitial fluid of brain increasing their local concentration, thus favoring conversion to larger PrPres amyloid aggregates. Serum amyloid P-component which binds to all amyloids and is a constituent of basement membranes might also contribute to local PrP conversion [35]. In addition, collagen, laminin and heparin sulfate-containing proteoglycans are major components of basement membranes, and PrP can bind to both the laminin receptor and heparan sulfate which can associate directly or indirectly with PrP [36]–[38]. Heparan sulfate and other glycosaminoglycan (GAG) moieties can delay scrapie disease in vivo [39]–[42] [43],[44], and some GAG molecules can alter PrP conversion in vitro [45]. A scaffolding mechanism might account for this effect. For example, soluble anchorless PrPsen monomers might be held in place by GAG polymers to increase local concentration and facilitate conversion by PrPres, analogous to the tethering of anchored PrPsen on cell membranes [46],[47]. In addition, attachment of small mobile PrPres oligomers to GAG polymers might assist conversion at the basement membrane. Subsequently newly formed larger less mobile PrPres could serve as an efficient scaffold for further conversion allowing the process to extend out into the brain parenchyma. Eventually this process might form very large PrPres amyloid plaques with blood vessels at the center as we observed (Figure 4E and 5A).
The vascular amyloid pathology seen in our scrapie-infected transgenic mice (Figures 4E, 5A–D, 6B–D) was similar to CAA seen in Alzheimer's disease as well as several familial amyloid diseases [14], including two forms of familial prion disease [32] [14]. In Alzheimer's disease, amyloid fibrils within vascular basement membranes are thought to impede interstitial fluid drainage leading to an increase in Aβ concentrations within the extracellular space. Such increased soluble Aβ and oligomeric proto-amyloid fragments are considered a likely contributory factor in the cognitive decline of Alzheimer's disease patients [15]. Similar processes might contribute to the clinical disease seen in the anchorless PrP scrapie model. Since all these diseases with CAA show amyloid localization with basement membranes, drugs capable of blocking amyloid-basement membrane interactions might be effective treatments for some of these diseases. In the case of prion diseases, one such compound, pentosan polysulfate, a small GAG oligomer, was effective in blocking PrPres generation in an infected cell line [48] and delayed onset of clinical scrapie in vivo [41] [42] [44]. Similarly a decoy molecule preventing PrP interaction with the laminin receptor (LRP/LR) reduced PrPres levels and delayed disease in vivo [37]. Determining the precise glycans and proteins involved in the protein interactions leading to amyloid deposition in all the CAA diseases might be important in designing new therapeutic approaches.
Ethics statement: All mice were housed at the Rocky Mountain Laboratories (RML) in an AAALAC-accredited facility, and research protocols and experimentation were approved by the NIH RML Animal Care and Use Committee.
C57BL/10SnJ mice (Prnp+/+) were obtained from Jackson Laboratories (Bar Harbor, Maine). C57BL/10SnJ PrP−/− mice were created at RML by crossing 129/Ola PrP−/− mice [49] with C57BL/10SnJ mice, followed by nine serial backcrosses to C57BL/10SnJ with selection for the Prnp+/− genotype using previously described PCR reactions to detect both the Prnp+ and Prnp null alleles [19]. One intercross was then done, and C57BL/10SnJ Prnp−/− (PrP−/−) mice were selected and interbred. Heterozygous Prnp+/− mice were obtained by intercrossing C57BL/10 (Prnp+/+) mice with C57BL/10 Prnp−/− mice.
Transgenic GPI anchorless PrP mice (tg44+/− and tg23+/−) were made as described previously [19] and then backcrossed to C57BL/10SnJ-Prnp−/− mice for six to nine generations with selection for the Prnp−/− genotype and the tg44 or 23+/− genotype. Thus these mice contained one anchorless PrP transgene allele and did not express any normal anchored mouse PrP allele. Heterozygous transgene lines tg23+/− and tg44+/− were each interbred to create homozygous lines (tg44+/+ and tg23+/+). Offspring were tested for transgene zygosity using real-time DNA PCR on an ABI Prism 7900 HT Sequence detection system and SDS 2.2.2 software. The following probes and primers were designed to amplify the mouse Prnp sequence: probe (moPrPlower418T): (5′-CGGTCCTCCCAGTCGTTGCCAAA), forward primer (moPrP-396F): (5′-CGTGAGCAGGCCCATGATC), reverse primer (moPrP-465R): (5′GCGGTACATGTTTTCACGGTAGT). Individual mice identified by rtPCR as transgene homozygous were then bred to Prnp−/− mice to confirm homozygosity. Homozygous mice were then interbred to create additional mice for experimentation. Both tg44 and tg23 lines were used in the present experiments to demonstrate that the observed findings were consistent with transgene expression rather than a result of an integration site artifact.
Four to six week old mice were inoculated intracerebrally with 50 µl of a 1% brain homogenate of 22L or RML scrapie containing 0.7–1.0×106 ID50. One ID50 is the dose causing infection in 50% of C57BL/10 mice. Animals were observed daily for onset and progression of scrapie. Mice were euthanized when clinical signs were consistent and progressive. Signs differed somewhat in C57BL/10 and tg44+/+ and tg23+/+ mice (Table 1). In heterozygous tg44+/− and tg23+/− mice many mice developed signs of debilitation such as weight loss, dermatitis and infections requiring euthanasia prior to severe neurological signs.
For detection of PrPsen from uninfected brains, tissues were homogenized (20% w/v) using a bead beater in ice-cold 0.01 M Tris-HCl pH 7.6 containing protease inhibitors (10 µM leupeptin, 1 µM pepstatin, and 1 µM aprotinin). Each sample was vortexed for 1 minute followed by sonication for 1 minute. Insoluble debris was removed by centrifugation at 2700 g for 10 minutes at 4°C. Samples were mixed 1∶1 with 2X SDS-PAGE sample buffer and boiled for 3–5 minutes. PNGase F reactions were done using 4.4 mg tissue equivalents in a total volume of 20 µl SDS-PAGE sample buffer [50]. Samples were serially diluted two-fold in sample buffer to give the amount of brain tissue (mg brain equivalents) indicated for each lane. Immunoblots were probed by using monoclonal anti-PrP D13 at a dilution of 1∶5000 (InPro Biotechnology, South San Francisco, CA), followed by secondary antibody sheep anti-human Ig (dilution 1∶5000) (GE Healthcare, formerly Amersham Biosciences, Piscataway, NJ) and enhanced chemiluminescence according to the manufacturers instructions (Amersham-Pharmacia, Uppsala, Sweden).
For detection of PrPres either with or without PNGase F, samples were prepared as described [51]. Blots were probed as described above.
Embryonic brain tissue was obtained from E12–E14 C57BL/6 embryos which expressed green fluorescent protein (GFP) in all tissues [25]. Mice were purchased originally from Jackson laboratories and were bred at Rocky Mountain Laboratories by Dr. Kim Hasenkrug. Pregnant mothers were euthanized and embryos dissected with forceps in media under a dissecting microscope to obtain the mesencephalon and telencephalon. Tissue was partially disrupted by pipetting to generate small fragments. This suspension (30 µl) was inoculated intracerebrally through the skull into the parietal brain region of 3–4 week old PrPnull mice or tg44+/− mice. One month later recipient mice were infected intracerebrally with 22L scrapie as described above. At various times thereafter mice were euthanized and brain tissue was examined histologically for GFP and PrPres by specific immunohistochemistry and for typical scrapie-induced gray matter spongiosis by H&E staining.
Mice were euthanized and brains were placed in 3.7% phosphate-buffered formalin for 3 to 5 days before dehydration and embedding in paraffin. Serial 4 µm sections were cut using a standard Leica microtome, placed on positively charged glass slides and dried overnight at 56°C. Slides were stained with a standard protocol of hematoxylin and eosin (H&E) for observation of overall pathology. For PrPres detection, slides were rehydrated in 0.1 M citrate buffer, pH 6.0 and then heated at 120°C, 20 psi for 20 minutes in a decloaking chamber (Biocare, Walnut Creek, CA). Immunohistochemical staining was performed using the Ventana automated Nexus stainer (Ventana, Tucson, AZ). Staining for PrP used a standard avidin-biotin complex immunoperoxidase protocol using anti-PrP antibody D13 (In-Pro Biotechnology, South San Francisco, CA) at a dilution of 1∶500 and incubated at 4°C for 16 hours. Biotinylated goat anti-human IgG (Jackson Immuno Research, West Grove, PA) was used at a 1∶500 dilution as the secondary antibody. Detection was performed with Ventana streptavidin-alkaline phosphatase with Fast Red chromogen. Tissue sections for microglia staining were pretreated and stained with anti-Iba1 as described [52] except that detection was done using the Ventana Fast Red chromagen as above. Astroglia were stained with anti-GFAP as described [52], and detection was completed with Ventana streptavidin-alkaline phosphatase using Fast Red. Tissue sections for staining with anti-amyloid precursor protein (APP) were pretreated as described for anti-PrP antibody D13. Anti-APP (Zymed Laboratories, San Francisco, CA) was used at a 1∶500 dilution followed by a 1∶250 dilution of biotinylated-goat anti-rabbit IgG (Vector Laboratories, Burlington, CA), and detection with Ventana streptavidin-horseradish peroxidase plus amino ethyl carbazol (AEC) chromagen. Staining of phosphorylated neurofilament proteins was performed using a monoclonal antibody cocktail pan-axonal neurofilament marker SMI-312 (Covance, Princeton, NJ) at a 1∶250 dilution. Monoclonal antibody to nonphosphorylated neurofilament proteins was also used (SMI-311). Primary antibodies were followed by biotinylated horse anti-mouse IgG secondary antibody at a 1∶250 dilution. Ventana AEC reagent was used for detection. Green fluorescent protein (GFP) was detected using a mixture of two mouse anti-GFP monoclonal antibodies (clones 7.1 and 13.1) at dilution of 1∶200 (Roche Applied Science, Indianapolis, IN), followed by biotinylated horse anti-mouse IgG (Vector Laboratories, Burlington, CA) at a dilution of 1∶250 and detected with AEC chromogen (Ventana) as described above. All histopathology slides were read using an Olympus BX51 microscope and images were obtained using Microsuite FIVE software.
Mice were perfused with fixative containing 3% paraformaldehyde and 1% glutaraldehyde in PBS. Excised tissues were then immersed in this fixative and held overnight at 4 degrees C. Tissue pieces were processed further using a Lynx® automated tissue processor with agitation as follows: one wash in PBS for 3 hr at 20 degrees, one wash in 0.1 M sodium phosphate buffer pH 7.2 at 20 degrees for 4 hr, post-fix in 2% osmium tetroxide in phosphate buffer at 20 degrees for 6 hr, one wash in phosphate buffer at 20 degrees for 3 hr, three washes in water at 20 degrees for 3 hr each, in-block staining with 1% uranyl acetate in water at 20 degrees for 6 hr, 3 washes in water at 20 degrees for 3 hr each, dehydration in 70%, 100%, and 100% acetone at 10 degrees for 3 hr each, and infiltration at 20 degrees in Araldite resin (Structure Probe, Inc., West Chester, PA) at 50% for 8 hr, 75% for 12 hr, and two changes of 100% for 20 hr each. Further tissue blocks were processed using a Leica EM TP processor using the procedure above with the omission of the uranyl acetate. Tissue blocks were then transferred to fresh resin in molds and polymerized at 65 degrees for 24 to 48 hr.
Thick (1 µm) sections were stained by toluidine blue or were immunolabelled using the avidin-biotin technique. Sections were deplasticized with saturated sodium ethoxide for up to 30 minutes. Endogenous peroxidase was blocked and sections were de-osmicated with 6% hydrogen peroxide for 10 minutes, followed by pre-treatment with neat formic acid for 5 minutes. Normal serum was then applied for 1 hour to block non-specific labeling. 1A8 anti-PrP serum [53] at a dilution of 1∶6000, or pre-immune serum were then applied for 15 hours and reaction product developed using 3-3′ diaminobenzidine.
For routine electron microscopy areas were selected from 1 µm thick toluidine blue stained sections and counterstained with uranyl acetate and lead citrate. For ultrastructural immunohistochemistry, serial 65 nm sections were taken from blocks previously identified from immuno-labeled 1 µm thick sections as described above. The 65 nm sections were placed on 600 mesh gold grids and etched in sodium periodate for 60 minutes. Endogenous peroxidase was blocked and sections de-osmicated with 6% hydrogen peroxide in water for 10 minutes followed by enhancement of antigen expression with formic acid for 10 minutes. Residual aldehyde groups were quenched with 0.2 M glycine in PBS, pH 7.4 for 3 minutes. Preimmune serum or anti-PrP primary antibody 1A8 [53] or R30 [54] at a 1∶500 or 1∶1500 dilution respectively in incubation buffer were then applied for 15 hours. After rinsing extensively, sections were incubated with Auroprobe 1 nm colloidal gold diluted 1∶50 in incubation buffer for 2 hours. Sections were then post-fixed with 2.5% glutaraldehyde in PBS and labeling enhanced with Goldenhance (Universal Biologicals, Cambridge, UK) for 10 minutes. Grids were counterstained with uranyl acetate and lead citrate.
|
10.1371/journal.pbio.1000512 | Fission Yeast Cells Undergo Nuclear Division in the Absence of Spindle Microtubules | Mitosis in eukaryotic cells employs spindle microtubules to drive accurate chromosome segregation at cell division. Cells lacking spindle microtubules arrest in mitosis due to a spindle checkpoint that delays mitotic progression until all chromosomes have achieved stable bipolar attachment to spindle microtubules. In fission yeast, mitosis occurs within an intact nuclear membrane with the mitotic spindle elongating between the spindle pole bodies. We show here that in fission yeast interference with mitotic spindle formation delays mitosis only briefly and cells proceed to an unusual nuclear division process we term nuclear fission, during which cells perform some chromosome segregation and efficiently enter S-phase of the next cell cycle. Nuclear fission is blocked if spindle pole body maturation or sister chromatid separation cannot take place or if actin polymerization is inhibited. We suggest that this process exhibits vestiges of a primitive nuclear division process independent of spindle microtubules, possibly reflecting an evolutionary intermediate state between bacterial and Archeal chromosome segregation where the nucleoid divides without a spindle and a microtubule spindle-based eukaryotic mitosis.
| The process of cell division, mitosis, ensures that chromosomes are accurately segregated to generate two daughter cells, each with a complete genome. Eukaryotic cells use a microtubule-based mitotic spindle to ensure proper chromosome segregation. In the fission yeast Schizosaccharomyces pombe, mitosis is “closed”: that is, the nuclear envelope does not break down, and the mitotic spindle forms within the nucleus. Unexpectedly we have found that in certain circumstances division of the fission yeast nucleus and progression into the next cell cycle can take place without the mitotic spindle. We call this nuclear division process “nuclear fission” because the nucleus separates into two bodies. We show that nuclear fission requires filamentous actin and functional spindle pole bodies, which are the fission yeast equivalent of the centrosome in other organisms. We also show that nuclear fission requires sister chromatid separation and is accompanied by some level of chromosome segregation. We propose that nuclear fission is a vestige of a primitive nuclear division process and might reflect an evolutionary intermediate between the mechanism of chromosome segregation that takes place in bacteria and the microtubule-based mitosis of modern eukaryotes.
| Mitosis is a feature of all known eukaryotic cells, essential for the generation of viable progeny. Upon entry into mitosis the duplicated centrosomes that serve as microtubule organizing centers separate and organize a bipolar array of spindle microtubules. Microtubules are essential for chromosome segregation and eukaryotic nuclear division is not known to occur in their absence. Spindle microtubules attach to kinetochores, specialized protein complexes that assemble on centromeres of each chromosome [1]. After sister chromatid cohesion is lost at anaphase, microtubules pull the sister chromatids apart to opposite poles of the spindle. A surveillance system, the spindle checkpoint, monitors mitotic progression and prevents the onset of anaphase until all chromosomes have achieved bipolar attachment and can segregate [2]. The spindle checkpoint monitors kinetochore-microtubule attachment and a single detached or misaligned kinetochore is thought to be sufficient to trigger the checkpoint delaying the onset of anaphase and cytokinesis as well as blocking chromosome replication in the following cell cycle [3]. Defects in the spindle checkpoint result in the premature onset of anaphase and lead to chromosome mis-segregation. Genetic screens aimed at the isolation of mutants hypersensitive to microtubule depolymerizing drugs have identified many components of the spindle checkpoint [4],[5] such as mad1, mad2, mad 3, bub1, and bub3, which are highly conserved from yeast to humans [6]–[8].
The fission yeast Schizosaccharomyces pombe undergoes a closed mitosis with the nuclear membrane remaining intact and a microtubule-based spindle extending within the nucleus between two spindle pole bodies (SPB), the centrosome equivalent [9]. As in other organisms, a spindle checkpoint delays mitotic progression in the presence of microtubule defects which disrupt the spindle. The extent of the mitotic delay due to spindle checkpoint activation is variable and depends on the nature of the mitotic insult. The β-tubulin cold sensitive nda3-KM311 mutant has no spindle microtubules and blocks in pre-prophase with condensed chromosomes [10],[11]; in contrast the temperature sensitive nda3-1828 mutant proceeds through an aberrant mitosis and cytokinesis [12]. Here we show that in fission yeast, although depolymerization of spindle microtubules activates the spindle checkpoint, it delays mitosis only for a short time, especially at high temperature. Also, unexpectedly, in the absence of spindle microtubules, cells undergo an alternative nuclear division process and proceed into the next cell cycle. This process requires SPB separation, sister chromatid separation, and actin polymerization. We suggest that this process might represent a primitive kind of eukaryotic nuclear division.
We assessed the extent of mitotic and cytokinesis delay due to spindle checkpoint activation by treating wild type and spindle checkpoint deficient mad2Δ fission yeast cells [13] with 50 µg/ml of carbendazim (MBC), which disrupts the mitotic spindle by depolymerizing microtubules [14]. Cells that fail to segregate their chromosomes but escape the spindle checkpoint proceed through to cytokinesis and septation without completing mitosis, generating a “cut” phenotype with the septum cutting through the nucleus (Figure 1A) [15]. After MBC addition we observed a delay of cytokinesis of 2 h at 20°C and 45 min at 25°C, while at the higher temperatures of 32°C and 36.5°C no significant delays were observed (Figure 1B). These results indicate that in fission yeast MBC-dependent spindle depolymerization activates the spindle checkpoint and delays cytokinesis only transiently at low temperatures and barely at all at high temperatures. We therefore tested whether at high temperature (36.5°C) MBC treated cells could re-enter the next cell cycle and replicate their DNA. We used a temperature sensitive mutant, defective in septation initiation network signaling, which prevents cytokinesis and thus the cutting of the nucleus by the closing septum. At 36.5°C, cytokinesis defective cdc11-119 cells treated with MBC continued DNA replication at a rate comparable to control DMSO-treated cells (Figure 1C). Similar results were obtained when additional MBC was added every hour to ensure the presence of active drug in the medium (unpublished data) and also when the cytokinesis defective mutants cdc4-8, cdc8-27, cdc12-112 (Figure 1D), as well as cdc7-24 and cdc3-124 (unpublished data) were used [16],[17]. In contrast no DNA replication was observed when cdc25-22 mutant cells blocked in G2 were treated with MBC (Figure 1D and unpublished data) [18],[19]. We conclude that if cytokinesis and septation are blocked, MBC-treated cells can proceed into the next cell cycle and undergo DNA replication. Consistently we observed no difference in DNA replication in cdc11-119 (Figure 1C) and cdc11-119 mad2Δ cells (Figure 1E) treated with MBC at the restrictive temperature.
The spindle checkpoint monitors kinetochore-microtubule attachment and a single unattached kinetochore is sufficient to activate the spindle checkpoint delaying the metaphase to anaphase transition and mitotic exit. We reasoned that given that MBC treated cells re-enter interphase, the spindle checkpoint may be inactive. To test this we determined whether at 36.5°C the checkpoint control was able to detect unattached kinetochores and therefore recruit checkpoint proteins such as Mad2 to the kinetochores. We investigated Mad2-GFP accumulation on kinetochores at 36.5°C and found that 21%±4% of MBC treated cells had discrete Mad2GFP loci compared to only 6%±1% in control DMSO treated cells (Figure 2A), indicating that unattached kinetochores were present and were recognized by the checkpoint machinery. Next we tested whether fission yeast cells were able to block cell cycle progression at 36.5°C in the presence of spindle microtubule defects. We used the kinesin 5-related mutant cut7-446, which fails to form a functional bipolar spindle due to lack of spindle microtubule interdigitation in the central region [20]. As shown in Figure 2B double mutant cdc11-119 cut7-446 cells blocked cell cycle progression and did not replicate their DNA during the 5 h time course. MBC treatment, however, was sufficient to allow cells to proceed through the cell cycle and replicate their DNA (Figure 2B), suggesting that fission yeast cells are competent to activate the spindle checkpoint at high temperature but not in the presence of MBC. As the spindle checkpoint senses unattached kinetochores we reasoned that in MBC treated cells either the checkpoint was never activated or residual undetected microtubules bound to kinetochores inactivated the checkpoint. To distinguish between these two possibilities we utilized a strain bearing a temperature sensitive mutation in the kinetochore protein Nuf2 [21]. The nuf2-2 allele at restrictive temperature abolishes microtubule binding, leaving the kinetochore competent to activate the spindle checkpoint [21]. As shown in Figure 2C, at their restrictive temperature cdc11-119 nuf2-2 cells delayed mitosis and re-entered interphase more slowly than cdc11-119 cells under the same conditions. However, when cdc11-119 nuf2-2 cells were shifted to the restrictive temperature in the presence of MBC, DNA replication occurred more readily than in DMSO treated control cells (Figure 2C). This indicates that failure to arrest in mitosis upon MBC treatment is unlikely to be caused by the inactivation of the spindle checkpoint by residual stable kinetochore microtubules.
We next tested whether MBC treated cells completely lacked mitotic spindles. First, we stained for α-tubulin using antibodies and detected only very short microtubule remnants less than 1µm in length (Figure 3A), consistent with previously published data [14]. Second we could not detect mitotic spindles using a strain bearing a GFP-tagged version of α-tubulin (Atb2-GFP) (Figure 3B); only very short microtubule stubs were occasionally observed in the cytoplasm. However, despite the absence of mitotic spindle microtubules, staining with the DNA dye 4′, 6-diamidino-2-phenylindole (dapi) revealed the presence of multiple DNA masses in MBC treated cdc11-119 cells. After 5 h at 36.5°C, 38% of the cell population had at least two well-separated DNA masses (Figure 3D). Visualization of the nuclear membrane with the marker Cut11-GFP (Figure 3A,C,D) [22] established that these DNA masses represented individual nuclear fragments enclosed by nuclear membrane. Time-lapse videos of cdc11-119 cut11-GFP atb2GFP cells at 36.5°C showed that the nucleus was undergoing a division process. However, unlike a normal mitosis there was no elongation of the nucleus into a dumbbell shape. Instead the nuclear envelope acquired a wobbly ruffled aspect and then eventually pinched into two nuclear masses (Figure 3E and Videos S1, S2). To confirm that the division of the nucleus was occurring without spindle microtubules, we used the double mutant cdc11-119 cut7-446, which fails to undergo mitosis at the restrictive temperature due to formation of monopolar spindles [20]. After 5 h at 36.5°C, 28% of the MBC treated cells had two nuclear masses (Figure 3F). Therefore, a nuclear division process takes place independently of a functional microtubule spindle. Occasional nuclear fragmentation has been reported in blocked nda3-KM311 cells [23], and we found that after 5 h at 19°C, dapi staining of nda3-KM311 cells showed that 30% of cells contained 2–3 nuclear bodies (Figure S1A,B). Membrane ruffling was also observed in these cells, as assayed using the nuclear envelope marker Uch2GFP [24],[25] (Video S3). We conclude that in the absence of spindle microtubules or a functional bipolar spindle, fission yeast cells can undergo an unusual nuclear division associated with ruffling of the nuclear membrane. Because the clearest characteristic of this process is fission of the nucleus, we have called it nuclear fission.
As normal mitotic progression is under surveillance of the spindle checkpoint, we tested whether this control was operative during nuclear fission. We reasoned that if nuclear fission was subject to the spindle checkpoint, then mad2Δ checkpoint deficient cells would undergo division more efficiently. However, we observed no difference in the accumulation of binucleates between control cdc11-119 cells and cdc11-119 mad2Δ cells, suggesting that nuclear fission is independent of the spindle checkpoint control (Figure 3G and Figure S2).
Fission of the nuclear membrane is the final event of mitosis, so we determined whether earlier events of mitosis were also taking place during nuclear fission. At the onset of mitosis, the duplicated SPBs separate and the mitotic spindle forms between them [9],[26]. To assess whether SPB separation occurred during nuclear fission, we used two SPB markers, Cut12-GFP (Figure 4A) [27] and Sad1-DsRed (Figure 5A,B) [28]. After 5 h in MBC 82% of the cdc11-119 cells had at least 2 SPBs (Table 1), and after 6 h cells with up to 8 SPBs were observed (Figure 4A). Significantly, almost all of the dapi-staining DNA masses (98%) were associated with at least one SPB (Figure 4A), suggesting that SPB separation was part of the process of nuclear fission. It has been previously shown that the SPBs facilitate nuclear envelope division during mitosis [25] and if SPB function is also required for nuclear fission, then impairing SPB maturation should block nuclear fission. To investigate this we monitored the appearance of binucleates in a cut11-2 mutant that fails to anchor the SPB in the nuclear envelope and exhibits defective maturation of a new SPB [22]. As shown in Figure 4D, nuclear fission was reduced from 38%±3% in cdc11-119 control cells to 7.5%±4% in cdc11-119 cut11-2 cells, indicating that efficient nuclear fission, like mitosis, requires functional SPBs.
A second important event of mitosis is sister chromatid separation, which is induced by degradation of the cohesin complex component Scc1 at the metaphase-to-anaphase transition [29]. We monitored chromosome I segregation during nuclear fission using a cen1-GFP expressing strain to mark the centromere of chromosome I. Within 5 h, all cells showed at least 2 cen1-GFP marked dots, establishing that separation of chromosome I centromeres was taking place (Figure 4B and Table 1). Centromeres were observed to segregate to different nuclear masses in 73% of the cells, which contained two nuclear masses. Similar results were obtained for chromosome II using a cen2-GFP strain (unpublished data). To monitor chromosome III segregation we used Clp1-GFP, which marks the nucleolus and co-segregates with chromosome III [30], and found that Clp1-GFP also partitioned to different nuclear masses in 70% of the cells with two nuclear masses (Figure 4C). These results indicate that sister chromatid cohesion is lost during nuclear fission, allowing sister chromatids to move away from each other and to segregate within the different nuclear masses. If chromatid separation is required for nuclear fission, then blocking the release of sister chromatid cohesion should reduce fission efficiency. In a separase mutant (cut1-645) [31] sister chromatids retain some cohesion and do not separate completely. After 5 h treatment with MBC, 11.5%±2% cdc11-119 cut1-645 cells contained two nuclear masses compared to 43%±8% in control cdc11-119 cells (Figure 4E), demonstrating that nuclear fission is significantly reduced if chromosome separation does not take place. We conclude that during nuclear fission SPBs and sister chromatids separate in the absence of spindle microtubules, that some level of chromosome segregation can take place, and that for efficient nuclear fission functional SPBs and sister chromatid separation are required.
During interphase, fission yeast centromeres cluster at the nuclear envelope in the vicinity of the SPBs [32] in a microtubule independent fashion [9],[32]. This clustering is lost upon entry into mitosis when kinetochores associate with mitotic spindle microtubules [32]. We considered that kinetochores might remain associated with SPBs in the absence of mitotic spindle microtubules. We therefore monitored centromere clustering at SPBs in MBC treated cells, using a strain bearing a centromere I marked with GFP and SPB tagged with Sad1-DsRed. In contrast to a normal mitosis we did not observe the cen1GFP signal dissociating from Sad1-DsRed labeled SPBs (Figure 5A), indicating that when microtubules are depolymerized by MBC treatment SPB-centromere association persists. We confirmed that the association was maintained for all three fission yeast centromeres using an ndc80-GFP bearing strain to label all three chromosomes simultaneously. In the presence of MBC at 36.5°C, no dissociation of the Ndc80-GFP signal from Sad1-DsRed tagged SPBs was detected (Figure 5B), suggesting that centromere clustering near the SPB persists during nuclear fission. However, it was possible that centromeres transiently dissociate from the SPBs upon mitotic commitment and then are quickly recaptured. To investigate this possibility we performed time lapse imaging of SPBs and centromeres in the presence of MBC. Time-lapse imaging of cdc11-119 sad1DsRED ndc80GFP showed that in the presence of MBC, SPB separation occurred without centromeres declustering (Figure 5C and Video S4). The two SPBs appeared to move apart from each other with their associated set of centromeres. As declustering occurs upon mitotic commitment, we further analyzed kinetochore clustering in a cdc11-119 cut11mcherry ndc80GFP strain. Cut11mcherry accumulates on SPBs from early mitosis through to the metaphase to anaphase transition, and therefore acts as a marker of mitotic commitment. We observed that MBC treated, early mitotic cells (as defined by SPB-Cut11mcherry accumulation) contained Ndc80GFP labeled dots that remained in close proximity to the SPB (Figure 5D and Video S5). The Ndc80GFP labeled dots moved away from each other only after Cut11GFP came off the SPB (Figure 5D 90″). Thus, during nuclear fission, unlike mitosis, centromeres remained clustered around the SPB. If centromere-SPB association is important for nuclear fission, then a mutant, which fails to maintain clustering of kinetochores at the SPBs, should impair nuclear fission. We used an ima1Δ strain, which disrupts kinetochore clustering at SPBs in 75% of cells [33]. We observed that after 6 h at the restrictive temperature 14%±2% of cdc11-119 ima1Δ MBC treated cells underwent nuclear fission compared with 34%±4% in control cdc11-119 cells. MBC treated cdc11-119 ima1Δ cells showed no significant change in nuclear ruffling compared to cdc11-119 cells (Video S6). Thus, failure to maintain the association between centromeres and SPBs significantly reduces the efficiency of nuclear fission.
Given that there are no microtubules present to generate the force necessary for nuclear fission, we ascertained whether nuclear fission required filamentous actin. Cdc11-119 cells were treated with MBC for 2 h, followed by addition of either DMSO or 10 µM latrunculin A (LA), an actin depolymerizing drug. In MBC treated cells, we observed that LA treatment completely blocked nuclear fission (Figure 6A). Nuclei of LA treated cells were rounder than those of control DMSO treated cells (Figure 6B) and time-lapse videos of cdc11-119 cut11GFP cells showed no nuclear membrane ruffling (Figure 6C and Video S7). Mitosis proceeded normally in the presence of 10 µM LA when MBC was not present. As in cdc11-119 cells, LA treatment blocked nuclear fission in nda3-KM311 cells at 19°C (Figure S1C). Despite the dependency of nuclear fission on the actin cytoskeleton we were unable to detect actin structures in or around the nucleus (Figure S3).
We next examined whether SPB separation was also affected in LA treated cells, by blocking cells expressing the SPB marker Sad1-RED for 1 h in the presence of MBC and then treating them either with DMSO or 10 µM LA. After 1 h MBC treatment, 24%±3% of cells had 2 SPBs. After a further 2 h in the presence of LA, 23%±2% of cells had 2 SPBs, compared to control DMSO treated cells, which showed 64%±2% of cells with 2 or more SPBs (Figure 6D). The SPBs in LA treated cells also remained in close proximity to each other compared to control cells (Figure 6E). Similarly we observed no increase in the number of cells with 2 Ndc80-GFP labeled kinetochores in LA treated cells (unpublished data). Thus, we conclude that nuclear fission depends on filamentous actin.
This work describes an unexpected process whereby in the absence of a mitotic spindle the fission yeast nucleus can undergo nuclear division. This process, which we have called nuclear fission, requires SPB maturation and sister chromatid separation and leads to some sister chromatid segregation. We propose the following mechanism for nuclear fission (Figure 7). As cells exit G2, sister chromatids remain clustered at the SPBs. The duplicated SPBs separate slowly by moving within the nuclear membrane, and as the sister chromatids lose cohesion, they move apart as a consequence of their association with SPBs. This mechanism assumes that the two sister chromatids are associated non-randomly with different SPBs. As the chromatids separate through the nucleus, the nuclear membrane deforms around the two DNA masses generating two nuclear bodies. Preventing SPB maturation or maintaining sister chromatid cohesion interferes with the separation of DNA masses and blocks nuclear fission. Although nuclear fission occurs in the context of closed mitosis, it is reminiscent of the formation of multiple nuclei in metazoan cells when a portion of the DNA becomes separated form the bulk of the chromosomes and is encapsulated in a separate nuclear entity. This happens under pathological conditions, for example in cancer cells as a consequence of inappropriate chromosome segregation or chromosome breakage [34],[35], or during oocyte meiosis and early developmental stages in mice deficient for the chromokinesin Kid [36], which show incomplete chromosome compaction. It also occurs in more physiological situations such as the formation of karyomeres in the early embryonic divisions of Xenopus, sea urchin, and polychaetes, where individual chromosomes are separated and engulfed by the nuclear envelope [37],[38]. Important for the formation of separate nuclear entities are the necessity for a minimal distance between DNA masses and for sufficient nuclear membrane to be available. Similarly in fission yeast, nuclear fission occurs only when SPB-chromosomes masses move away from each other and when lipid biosynthesis is up-regulated during the expansion of the nuclear envelope to accommodate the elongating spindle [39]–[41]. Understanding the regulation of nuclei formation during nuclear fission might shed light on the mechanism that controls the formation of a single nucleus around each chromosome complement at the end of mitosis.
Actin polymerization, which is required for nuclear fission, might be involved in the membrane redistribution associated with the increase in nuclear envelope area observed during early anaphase B. Nuclear envelope ruffling could be a consequence of a rapid redistribution of membranes between ER and nuclear envelope. During nuclear fission, ruffling is most obvious after cut11GFP comes off the SPBs at a stage corresponding to anaphase B when spindle elongation takes place during a normal mitosis. Therefore, nuclear envelope ruffling might be expected to be more obvious during nuclear fission when there is no spindle elongation to stretch the nuclear membrane. Consistently, nuclear ruffling in fission yeast is also observed during mitosis if spindle elongation is blocked, as in the kinetochore mutant nuf2-3 (our unpublished observation). Nuclear membrane expansion was also observed in budding yeast cells, blocked in mitosis with nocodazole to depolymerize microtubules. It has been suggested that this expansion is a consequence of the up-regulation of lipid biosynthesis normally taking place during mitosis [42]. Nuclear membrane extensions appear also upon deregulation of phospholypid biosynthesis by spo7 inactivation [41],[42]. However, such extensions are not observed if spo7 inactivation takes place during anaphase probably because of the incorporation of extra membrane into the elongating nucleus [41]. Further studies will be required to clarify whether phospholipid biosynthesis has a role in nuclear fission or if actin is involved in the nuclear expansion observed during a normal mitosis. However, there could be other roles for the involvement of actin in nuclear fission. Actin could generate either a pushing force causing nuclear membrane distortion as is seen during lamellipodia protrusion [43] or a pulling force separating SPBs and their cargo of chromosomes. In this context it is of interest that bacterial chromosome segregation is driven by polymerization of the actin-like MreB/ParM protein [44],[45] and also that in vertebrates the driving force for centrosome separation is provided by actin filaments [46]–[48].
Differently from a normal mitosis, nuclear fission appears not to be under spindle checkpoint control and takes place irrespective of checkpoint engagement. We observed nuclear fission both in nda3-KM311 cells, which activate the spindle checkpoint blocking cells in pre-prophase, and in MBC treated cdc11-119 cells, which do not delay mitotic exit. Interestingly, in both circumstances cells accumulate Mad2 on kinetochores, suggesting that unattached kinetochores have been detected. However in MBC treated cells, where microtubules are almost completely depolymerized, no mitotic delay is observed and cells re-enter interphase. This result suggests that Mad2 accumulation on kinetochores, although necessary for checkpoint activation, might not be sufficient to maintain the mitotic block. Alternatively the checkpoint might be activated normally but only transiently. As a monopolar spindle does activate the checkpoint (cut7-446) and we have excluded the possibility that the binding of residual kinetochore microtubules inactivates the checkpoint (nuf2-2), something other than microtubules would have to be involved in the inactivation. It has been shown that in budding yeast activation of the APC inhibits the mitotic checkpoint through APC-mediated degradation of the checkpoint kinase Mps1 [49]. This feedback between APC and the spindle checkpoint implies that if enough APC is activated the checkpoint is turned off. In S. pombe, such a mutual inhibition would mean that if the nuclear fission process were to generate enough active APC, then the checkpoint block would be overcome. Alternatively, maintenance of the SPB-kinetochore interaction might cause kinetochores to be less readily available for full strength checkpoint activation, leaving sufficient levels of active APC in the nucleus to turn off the checkpoint. Further studies will be required to understand how MBC-mediated microtubule depolymerization inactivates the spindle checkpoint in fission yeast.
We speculate that nuclear fission might be a vestige of an ancient mechanism of nuclear division that predates the appearance of a mitotic spindle. It might reflect an evolutionary intermediate state between the mechanism of chromosome segregation seen in bacteria and Archea [44],[50] and that seen during mitosis in eukaryotic cells. In the intermediate evolutionary state, the replicated sister chromatids would remain attached to the centrosomes and become segregated by movement of the divided centrosomes within the nuclear membrane. Later in evolution, addition of a mitotic spindle between the centrosomes would have increased the efficiency of centrosome separation and the microtubule-based polar movement of sister chromatids would have led to a more efficient and accurate conventional mitosis. It has been observed that in the absence of the tubulin homologue FtsZ, the L-form of Bacillus subtilis acquires an unusual mode of proliferation with cells undergoing membrane ruffling prior to the formation of protrusions, which then resolve into independent round bodies [51]. It has been suggested that this extrusion-resolution mode could either be driven by force generated by the actin homologue MreB or by active chromosome segregation followed by collapse and resealing of the membrane. These mechanisms may be relevant to the nuclear fission process we have observed in fission yeast in the absence of spindle microtubules. A spindle independent mechanism (SIM) has also been reported for nucleolar segregation during Aspergillus nidulans mitosis [52]. The mechanism underlying SIM in A. nidulans is not yet clear, but the nuclear envelope is believed to play a critical role to generate the force necessary for nucleolar separation. In summary, we suggest that nuclear fission represents the vestiges of a primitive nuclear division process that existed early in the eukaryotic lineage prior to the evolution of a mitotic spindle and mitosis as known today.
All S. pombe strains used in this study are listed in Table S1. Standard methods were used for growth and genetic manipulation [53]. All experiments, unless otherwise stated, were performed in YE4S (yeast extract with added 250 mg/l histidine, adenine, leucine, and uridine).
Cells were grown at 25°C to 1–2×106 cell/ml density before shifting to the restrictive temperature (36.5°C). After 3 h (unless otherwise stated), cultures were split in two and treated with either 50 µg/ml MBC (freshly made in DMSO) or DMSO at 36.5°C, unless otherwise stated. It should be noted that some batches of MBC are more toxic for cells and these were not used in this study. For lat A treatment, following a 2 h block at 36.5°C in the presence of either 50 µg/ml MBC or DMSO, cells were treated with 12.5 µM lat A.
For immunofluorescence, cells were collected by filtration and then fixed. Cells were fixed in −80°C methanol for 1 h and then processed as previously described [54]. For microtubule detection, TAT1 antibody (a-tubulin antibody; a kind gift of Prof. K. Gull) was used at 1∶200 dilution and Alexa fluor 546-linked anti-mouse (Molecular Probes) at 1∶1000 dilution as secondary antibody.
For actin staining, cells were fixed by adding formaldehyde (final concentration 3.7%) to the medium for 25 min, washed twice with PEM (100 mM Pipes, 1 mM EGTA, 1 mM MgSO4 pH = 6.9), permeabilized with 1% Triton X-100, washed twice with PEM, and stained with rhodamine phalloidin (Molecular Probes).
For dapi staining, cells were either heat fixed (70°C) or fixed in 70% cold ethanol, then re-hydrated in distilled water, and stained with 2 µl of 50% glycerol, 0.1 M Tris pH 8 containing 1 µg/ml dapi. For immunofluorescence, cells were fixed in cold methanol at −80°C overnight and then processed as previously described [54]. Images were taken using a Deltavision microscope. The percentages are averages of 3–8 experiments and the standard errors were calculated and reported.
For live imaging cells were attached to coverslips using soya bean lectin (100 µg/ml) and imaged in minimal medium either containing DMSO or 50 µg/ml MBC at 36.5°C, using a Deltavision microscope supplied with a temperature controlled chamber.
|
10.1371/journal.ppat.1002173 | Autocrine Regulation of Pulmonary Inflammation by Effector T-Cell Derived IL-10 during Infection with Respiratory Syncytial Virus | Respiratory syncytial virus (RSV) infection is the leading viral cause of severe lower respiratory tract illness in young infants. Clinical studies have documented that certain polymorphisms in the gene encoding the regulatory cytokine IL-10 are associated with the development of severe bronchiolitis in RSV infected infants. Here, we examined the role of IL-10 in a murine model of primary RSV infection and found that high levels of IL-10 are produced in the respiratory tract by anti-viral effector T cells at the onset of the adaptive immune response. We demonstrated that the function of the effector T cell -derived IL-10 in vivo is to limit the excess pulmonary inflammation and thereby to maintain critical lung function. We further identify a novel mechanism by which effector T cell-derived IL-10 controls excess inflammation by feedback inhibition through engagement of the IL-10 receptor on the antiviral effector T cells. Our findings suggest a potentially critical role of effector T cell-derived IL-10 in controlling disease severity in clinical RSV infection.
| IL-10 is a major anti-inflammatory protein that plays an essential role in regulating the balance between pathogen clearance by the immune response and immune mediated injury resulting from the immune response to pathogen infection. In this report, we demonstrate that anti-viral effector T cells, a critical cell type responsible for respiratory syncytial virus clearance, are able to produce a large quantity of IL-10. The function of IL-10 is to control the immune response in order to avoid the development of excessive pulmonary inflammation associated with the clearance of infectious virus. We further identified a likely mechanism that T cell-derived IL-10 operates to control inflammation and describe a novel potential target of IL-10 action in the RSV infected lungs. Our data thus may lay the ground for the future studies exploring the application of IL-10 in therapeutic approaches to modulate pulmonary inflammation and injury in young infants suffering severe respiratory syncytial virus induced diseases.
| Respiratory syncytial virus (RSV) infection is the leading viral cause of upper and lower respiratory tract illness in young infants. In the USA, nearly 100% of children are infected with RSV by the age of 2–3 [1]. Approximately 1–2% of these infected children develop moderate to severe bronchiolitis [2]. The exact mechanisms underlying the development of severe pulmonary diseases in the small proportion of children remain poorly defined. Nevertheless, in both clinical studies and animal models, severe pulmonary disease induced by RSV infection is typically associated with an exaggerated inflammatory response in the lower respiratory tract, characterized by the overproduction of pro-inflammatory cytokines/chemokines and increased infiltration of inflammatory cells [3], [4], [5]. Furthermore, there is no firm correlation between disease severity and the extent of RSV replication [6], further suggesting a likely important role of the host immune response to RSV in determining disease severity. Consequently, in responding to an infectious agent like RSV with a strong potential to induce immune-mediated pathology, there is a need to finely balance the immuno-protective and immuno-pathological potential of the immune response in order to insure virus clearance without excess inflammatory injury.
IL-10 is a major regulatory cytokine with broad anti-inflammatory properties [7]. Depending on the nature of the pathogenic stimulus, many cell types including neutrophils, NK cells, macrophages, dendritic cells (DC), regulatory and effector T cells have been shown to be capable of producing IL-10 both in vitro and in vivo in response to infection [8], [9]. IL-10 is generally viewed as a negative regulator of the response of both innate and adaptive immune cells during infection particularly during persistent parasitic, bacterial, and viral infections where it can suppress pathogen clearance and/or the inflammatory response triggered by the infectious agent [8]. Recent evidence suggests that IL-10 may play an important regulatory role in acute viral infections of the respiratory tract where it inhibits the development of excess pulmonary injury in the face of normal virus clearance from the respiratory tract [10]. These findings, along with evidence of a link between a polymorphism in the IL-10 locus and the severity of bronchiolitis in infants infected with RSV [11], [12], [13], [14], prompted us to explore the role of IL-10, and in particular of IL-10 produced by virus-specific effector T-cells, in controlling pulmonary injury associated with experimental murine RSV infection.
In this study, we investigated the source and role of IL-10 during primary RSV infection. We found that high levels of this regulatory cytokine are produced simultaneously with effector cytokines in the respiratory tract at the onset of the adaptive immune response. The main cellular sources of IL-10 in the lung are RSV-specific CD4+ and CD8+ effector T-cells. We show that effector T cell-derived IL-10 during RSV infection in vivo acts to inhibit excess inflammation in the respiratory tract and thereby maintain critical pulmonary function in the infected host. We further provide evidence for a novel mechanism of IL-10-mediated inflammation where effector T cell-derived IL-10 acts in an autocrine manner on the effector T cells to suppress excess pulmonary inflammation induced by the anti-viral response of the effector T cells. The implications of these findings for RSV infection are discussed.
Because of IL-10′s documented role in controlling or inhibiting the development of excess inflammation in response to infection and injury, and due to the evidence linking expression of IL-10 to injury severity in human RSV infection, we evaluated the role of IL-10 in virus clearance and the control of pulmonary inflammation in a murine model of experimental RSV infection [8], [14]. To address this question, we first infected BALB/c mice with RSV and measured by ELISA the kinetics of IL-10 release into the bronchoalveolar lavage fluid (BALF) sampled from RSV-infected lungs. As a measure of proinflammatory cytokine release during infection we monitored in parallel the kinetics of IFN-γ production in the infected lungs. We found that minimal levels of either cytokine were detected in the BALF early during infection (day 1–3 p.i.) (Figure 1A). By day 5 p.i., a time at which RSV effector T cells have been shown to begin infiltrating into the respiratory tract [15], both IL-10 and IFN-γ levels in the BALF increased dramatically in a coordinated manner. The levels of the two cytokines then decreased progressively and both returned to background levels by day 9 p.i. (Figure 1A).
The kinetics of IL-10 and IFN-γ release into the BALF and its tight association with the influx of RSV-specific effector T cells into lungs raised the possibility that these 2 cytokines may be products of adaptive immune cells, or at least that their expression is linked to the recruitment of virus-specific adaptive immune cells into the infected lungs. To initially explore the contribution of adaptive immune (T and/or B) cells to the IL-10 response in the infected respiratory tract, we infected WT or Rag1-/- mice (which lack T and B cells) with RSV and measured IL-10 and IFN-γ levels in the BALF of these mice. We found that high levels of both cytokines were released into the airways of WT but not Rag1-/- mice after RSV infection (Figure 1B). Uninfected WT (Figure 1B) or Rag1-/- (data not shown) mice had negligible levels of this regulatory cytokine. These data suggest that the release of both regulatory cytokine IL-10 and effector cytokine IFN-γ following infection is dependent on the adaptive immune T and/or B cells.
To further define the cellular source(s) of IL-10 in vivo during RSV infection, we infected the IL-10/eGFP reporter mice (Vert-X) with RSV [16]. At day 5 p.i. (i.e at the peak of IL-10 detection in the BALF) we harvested lung cells and measured IL-10/eGFP expression by the liberated lung cell populations using flow cytometry. We found that, compared to the background fluorescence observed in lung cells from RSV infected WT mice, a significant percentage of lung cells from infected Vert-X mice express IL-10/eGFP (Figure 1C). More importantly we found that IL-10/eGFP expression in the Vert-X lungs was restricted almost exclusively (>90%) to Thy-1+ cells infiltrating the infected lungs and that the frequency of IL-10/eGFP+ CD4+ T cells among the Thy-1+ lymphocytes roughly equaled that of IL-10/eGFP+ CD8+ T cells (Figure 1C). These data further reinforced the view that CD4+ and CD8+ T lymphocytes are the major cellular sources of IL-10 in the lung during RSV infection. Next, we examined the IL-10/eGFP expression by CD8+ and CD4+ T cells from the uninfected (day 0) Vert-X lungs and from the lungs, the draining mediastinal lymph nodes (MLN) and the spleens of RSV infected Vert-X mice. We found that IL-10/eGFP-expressing cells are highly enriched in the RSV infected lungs, but not in the uninfected lungs nor in the MLN or the spleen of infected mice (Figure 1D, E). We also investigated the kinetics of accumulation of IL-10-expressing cells in the lungs following RSV infection and found that the IL-10-expressing cells were restricted to the acute phase of infection (Figure S1). Notably, even though the absolute number of IL-10/eGFP-expressing (IL-10 mRNA +) CD8+ and CD4+ continued to increase in the lungs from d5 to d7 post infection, the in vivo release of IL-10 protein in the airway peaks at d5 post infection, which coincides with the fall in lung virus titers and so the viral antigen load in the lung. This observation is consistent with our previously reported findings [10], [17] and likely reflects dependence of IL-10 protein synthesis and release on TCR engagement and viral antigen recognition.
The coordinated release of IL-10 and IFN-γ in the respiratory tract suggests that the IL-10 producing T cells in the RSV infected lungs may be anti-viral effector T cells capable of producing IFN-γ as well. To address this possibility, we first examined the expression of several cell surface molecules associated with T-cell activation and/or effector differentiation by the lung T cells from uninfected Vert-X mice (control), along with lung IL-10/eGFP+ and IL-10/eGFP− T cells from RSV infected Vert-X mice. We found that both IL-10/eGFP+ as well as IL-10/eGFP− CD8+ T cells from infected Vert-X lungs express high levels of T cell activation/effector markers such as CD44, CD43, ICOS, and low levels of naïve T cell marker CD62L suggesting that, like the eGFP− CD8+ T cells, the IL-10/eGFP+ CD8+ T cells likely represent activated effector CD8+ T cells (Figure 2A). Likewise, both IL-10/eGFP+ and IL-10/eGFP− CD4+ T cells express higher levels of those activation/effector markers compared to control CD4+ T cells (Figure 2A). We also FACS-sorted IL-10/eGFP+ and IL-10/eGFP− T cells from infected lungs and measured the gene expression of signature molecules of effector cells. We found that both CD8+ IL-10/eGFP+ cells and CD4+ IL-10/eGFP+ cells express high levels of effector molecules such as IFN-γ, Granzyme B and the type 1 effector cell lineage specific transcription factor T-bet [18], suggesting that the IL-10 expressing CD8+ and CD4+ cells are indeed type 1 effector cells (Figure 2B, C). Consistent with this idea, we observed that all IL-10-producing CD8+ and most IL-10-producing CD4+ T cells simultaneously produce IFN-γ in response to mitogenic or antigenic stimulation in vitro in the intracellular staining assay (ICS) (Figure 2D, E and data not shown). Furthermore, we analyzed by ICS assay the expression of T-bet and the regulatory T cell specific transcription factor Foxp-3 [19], in IL-10− IFN-γ+ (IFN-γ single positive, IFN-γSP) CD8+ T cells, IL-10− IFN-γ+ (IFN-γSP) CD4+ T cells , IL-10+ CD8+ T cells and IL-10+ CD4+ T cells from the infected lungs directly ex vivo. Consistent with the mRNA levels (Figure 2B, C), we found that IL-10 positive CD8+ or CD4+ T cells express T-bet at levels as high as the (IL-10−) IFN-γ single-positive CD8+ or CD4+ type 1 like effector cells (Figure 2D, E). Furthermore, most of IL-10+ CD8+ cells and the majority of the IL-10+ CD4+ T cells are Foxp-3 negative (Figure 2F). Collectively, these data suggest that the vast majority of the IL-10-expressing CD8+ and CD4+ T cells are primarily type 1 effector cells. We did however observe that a minor proportion, up to 30%, of the IL-10+ CD4+ T-cells, expressed Foxp-3. These Foxp-3+ CD4+ T cells were uniformly negative for IFN-γ production (Figure 2F and data not shown).
To determine if the effector T cell-derived IL-10 had any impact on the outcome of RSV infection, we examined the effect of blockade of the IL-10 receptor (IL-10R) by in vivo administration of a blocking anti-IL-10Rα (α-IL-10R) mAb on virus clearance, pulmonary function and lung inflammation. We found that the administration of α-IL-10R mAb in vivo significantly enhanced the weight loss of RSV infected mice, particularly at day 5 and thereafter when RSV-specific adaptive immune T cells began infiltrating into the respiratory tract (Figure 3A). We also determined whether three parameters of lung function, as described in Methods, were similarly affected after blocking the action of IL-10 in vivo. As Figure 3B demonstrates, the blockade of IL-10 function in vivo leads to a significant decrease in lung compliance (LC) and increase in pulmonary artery pressure (PAP) in the treated mice at d7 post infection. In addition, although it did not reach statistical significance, airway resistance (AR) was also increased in the mice treated with α-IL-10R mAb compared to the mice administered control mAb (Figure 3B).
We next compared viral clearance in RSV infected mice which were treated with either Rat IgG1 control mAb or α-IL-10R mAb by plaque assay and by quantitative RT-PCR for RSV genome copies. As Figure 3C, and D demonstrate, mice treated with α-IL-10R mAb had cleared virus by day 7 post infection as efficiently as the mice receiving Rat IgG1 control mAb, suggesting that the blockade of IL-10R signaling in vivo does not alter the kinetics of viral clearance in the lung during RSV infection. These results collectively demonstrated that the blockade of IL-10 function in vivo leads to enhanced host diseases without affecting viral clearance. Although we cannot exclude formally the contribution of IL-10 derived from the small fraction of Thy-1−/CD3− IL-10+ cells in the infected lungs (Figure 1C) to the control of inflammation, our results strongly suggest that effector T-cells are the major source and the most important source of this regulatory cytokine.
To determine the role of T cell-derived IL-10 in restraining host morbidity, we transferred naive WT or IL-10-/- Thy1+ cells (including both CD4+ and CD8+ T cells) into Rag1-/- mice and infected the recipient animals with RSV. We found that Rag1-/- mice reconstituted with IL-10-/- T cells had increased weight loss compared to Rag1-/- mice reconstituted with WT T cells following RSV infection (Figure 3E). These data suggested that IL-10 derived from T cells themselves is able to control host morbidity in response to RSV infection.
We next examined whether the blockade of IL-10 action in vivo leads to enhanced pulmonary inflammation following RSV infection. We first determined the impact of IL-10R blockade on the recruitment of innate inflammatory cells to the lung. We found that IL-10R blockade resulted in a substantial increase in the number of proinflammatory monocytic cells and notably neutrophils infiltrating the infected lungs (Figure 4A, B). This increase in inflammatory monocyte and neutrophil infiltration is accompanied by enhanced release of pro-inflammatory cytokines including IL-12/23 p40, IL-6, TNF-α and IFN-γ into the airways of α-IL-10R mAb-treated mice compared to control mAb-treated mice (Figure 4C–F). Taken together, these data suggest that a critical function of T-cell-derived IL-10 in vivo during RSV infection is to prevent the development of excessive pulmonary inflammation in the infected respiratory tract associated with virus infection and the host innate/adaptive immune response and to retain essential lung function in the infected host without inhibiting virus clearance.
Along with their role in virus clearance [20], effector T cells have been shown to significantly contribute to lung inflammation and host morbidity in the murine model of primary RSV infection [3], [20], [21]. The enhanced morbidity, excess pulmonary inflammation/injury and altered pulmonary function observed with IL-10R blockade during RSV infection may reflect a normal function of IL-10 in regulating the induction, expansion and/or effector activity of effector T cells responding in the lungs to infection. We first investigated the impact of IL-10R blockade in vivo on the induction (in the draining MLN) and migration of effector T cells to the lung. Somewhat unexpectedly, the numbers of activated/effector CD8+ or CD4+ T cells in the infected lungs were comparable in mice treated either with control Rat IgG1 mAb or α-IL-10R mAb (Figure S2). Furthermore, effector T cells from either control or α-IL-10R-treated mice had comparable capability to produce IFN-γ and TNF-α in response to mitogenic or antigenic stimulation (Figure 4G and data not shown), suggesting that IL-10 does not affect the differentiation of T cells into RSV specific effector T-cells of the Th1 or Tc1 lineage.
In view of the elevated levels of lung proinflammatory cytokines, in particular IFN-γ, we next explored the impact of IL-10R blockade on the in vivo frequency and cytokine profile of virus-specific T cells responding in the infected respiratory tract. For this purpose we chose IFN-γ and TNF-α as representative effector cytokines and used the the in vivo ICS assay [22] to measure their release by CD8+ and CD4+ T cells in the respiratory tract. We found that blockade of IL-10 function in vivo triggered an increase in the percentage/number of CD4+ and particularly CD8+ T cells producing IFN-γ in vivo (Figure 4H, I). In addition, the in vivo production of TNF-α by CD8+ T cells was also modestly increased following IL-10R blockade during RSV infection (Figure 4H). The blockade of IL-10 function in vivo also leads to increased production of IFN-γ on a per cell basis (Figure S3). These data demonstrate that this effector T cell-derived IL-10 can act in vivo to suppress the production of proinflammatory mediators by effector T cells responding in the respiratory tract of virus infection.
Cells of the myeloid, monocyte/macrophage/dendritic cell lineage are believed to be the major targets of IL-10 [7]. Therefore IL-10 would most likely be expected to diminish the effector activity of effector T cells in vivo by inhibiting the APC function of these inflammatory mononuclear cells infiltrating the infected lungs. Alternatively, the T-cell-derived IL-10 could also act in an autocrine fashion to suppress the activation/stimulation of CD8+ and CD4+ lung effector T cells through the engagement of the IL-10R on these T-cells. To explore this latter possibility, we first isolated CD8+ effector T cells from RSV infected lungs and examined whether they can respond to IL-10. As Figure 5A demonstrates, we found that CD8+ effector T cells isolated from infected lungs are able to respond to IL-10 treatment by phosphorylating STAT-3. Furthermore, IL-10 is able to inhibit the release of IFN-γ by CD8+ effector T cells in response to CD3 stimulation in the absence of antigen presenting cells (Figure 5B). Collectively, these data demonstrated that IL-10 is able to signal to effector T cells and suppress their proinflammatory activity. To directly explore in vivo the possible contribution of an autocrine mechanism of IL-10 action during RSV infection, we examined the response to primary RSV infection of mice with a conditional deletion of the IL-10Rα gene selectively in CD4+ and CD8+ T cells [23]. We confirmed that IL-10Rα was selectively deleted in T cells, particularly CD8+ T cells in the lung but not in lung monocytes/macrophages or NK cells etc (Figure 5C and data not shown). Notably, compared to CD8+ effector T cells, we failed to detect significant IL-10Rα expression in CD4+ effector T cells (Figure 5C). Importantly, following RSV infection, the deletion of IL-10Rα in T cells resulted in dramatically increased release of effector T cells derived cytokine IFN-γ into the airway (Figure 5D). Interestingly, the deletion of IL-10Rα in T cells also resulted in significantly increased infiltration of neutrophils and monocytes (but not T cells) into the infected lungs (Figure 5D, and Figure S4), suggesting a role for enhanced adaptive immune-mediated inflammation in promoting neutrophil and monocyte infiltration to the lung. Consistent with the finding of increased inflammation in the lung, we observed that conditional deletion of the IL-10Rα in T cells resulted in increased weight loss and delayed recovery following RSV infection of the knockout mice (Figure 5E).
In this report, we examined the in vivo production, cellular sources and function of the regulatory cytokine IL-10 during acute pulmonary RSV infection. Importantly, the primary source of IL-10 is the anti-viral CD4+ and CD8+ effector T cells recruited to the RSV infected lungs. We demonstrate that this effector T cell-derived IL-10 plays a critical role in preventing excess inflammation during the innate and particularly the adaptive immune response to RSV. In the absence of IL-10 signaling, an enhanced inflammatory response and a concomitant alteration in pulmonary function ensues. Thus, IL-10 may play a critical role in maintaining lung function during infection. Of note, our findings suggest that effector T cells are not only the major source of IL-10 in the infected lungs but may also serve as important cellular targets for the action of this regulatory cytokine, reflecting a novel autocrine pathway for the action of IL-10 during infection.
Several lines of evidence in the current study implicate effector T cells as the primary source of IL-10 observed in the RSV infected respiratory tract in this model. Similar to our recently reported findings in the murine influenza infection model [10], we noted minimal IL-10 secretion in RSV infected Rag1-/- mice and localized IL-10 production by effector T cells in the infected lungs using the IL-10 reporter Vert-X mice for infection. However, we cannot formally exclude a major contribution of the small percentage (2%–5%) of Thy-1– (negative) cells detected to the amelioration of pathology. In this case, however, we demonstrated that IL-10 derived solely from T-cells are sufficient to control disease development during RSV infection (Figure 3E) which also implicated T-cell-derived IL-10 in the control of excess pulmonary inflammation. The IL-10-producing effector T cells are, as we observed in this report, IL-4 and GATA-3 negative (data not shown) and thus show few features characteristic of type 2 lineage of T cells; but rather are T-bet positive type 1 effectors. Interestingly, we observed a minor fraction of T-bet+ IL-10 producers which co-express Foxp-3, a cell population that has been recently described in other systems [24]. Of note, CD25+ T regulatory cells have been recently demonstrated to modulate RSV specific CD8+ T cell responses and pulmonary inflammation during experimental RSV infection [25], [26], [27]. How these Tregs interact with the IL-10 secreting effector T cells responding in the respiratory tract to RSV infection to control excess inflammation will require further investigation.
It is of interest that RSV infection induced a lower level of IL-10 release into the infected airways relative to the release of effector T cells derived pro-inflammatory effector cytokine IFN-γ than detected in influenza infected lungs. The explanation for this discrepancy is unclear, but this could reflect an as yet unappreciated mechanism by which RSV infection results in an imbalance in the expression of pro-inflammatory (e.g. IFN-γ and regulatory (e.g. IL-10) cytokines by effector T-cells resulting in exaggerated inflammation/injury profile in the infected lungs out of proportion to the degree of RSV replication in the infected lungs.
Multiple lines of evidence both from human studies and murine models implicate myeloid cells of the monocyte/macrophage/dendritic cell lineage as the major targets of IL-10 action in vitro and (in model systems) in vivo [7], [9], [23], [28], [29]. Thus results from the current study suggest that effector T cell-derived IL-10 can act directly on these inflammatory mononuclear cells infiltrating the RSV infected lungs to decrease production of pro-inflammatory cytokines/chemokines by these inflammatory mononuclear cells as well as to modulate the expression of costimulatory ligands (e.g. CD40, 80, 86 etc.) and T-cell stimulatory cytokines (e.g. IL-12 etc.) [28]. The latter effects of IL-10 would be expected to diminish the efficiency of effector T cell triggering in response to contact with these APC populations and thus down regulating the effector activity of T cells. While our results are consistent with such a mechanism, our findings on the impact of conditional deletion of the IL-10Rα selectively in T-cells suggests the novel possibility that the effector T cell-derived IL-10 may also act in an autocrine fashion to regulate effector T cell activity directly. Of note, using animals deficient in IL-10Rβ chain, several recent reports have provided evidence that IL-10 may act directly on and suppress the function of T cells including Treg cells [30] and CD4+ T cells directly in vivo during acute LCMV infection [31] as well as inhibit robust memory CD8+ T cell development [32]. It should be noted, however, that IL-10Rβ is the common subunit for receptors recognizing several other cytokines including IL-22, IL-26 and IL-28 etc [33]. Our results using the conditional deletion of the IL-10 receptor a chain, which is unique to IL-10, firmly establish that effector T cells are able to respond to IL-10 directly in vivo during infection. Furthermore, we described a novel autocirne function of IL-10; that effector T cell-derived IL-10 is able to signal back to effector T cells, especially CD8+ effector T cells, to restrict the proinflammatory cytokine production by these effector T-cells. Importantly, this autocrine regulatory function of IL-10 acting on effector T cells modulates pulmonary inflammation and thereby results in diminished host morbidity. A similar autocrine mechanism has been reported recently for regulation of IL-10 producing macrophages during endotoxin challenge [23], [29]. The mechanisms through which IL-10R engagement on activated effector T cells acts to suppress effector T cell functions at sites of infection are currently under investigation.
IL-10 has been detected both in respiratory tract secretions and in the serum of infants and young children during the acute phase of RSV infection [34]. Although controversial [14], [35], several lines of evidence suggest that the level of IL-10 secretion may correlate inversely with disease severity [12], [13]. This may be particularly evident among children hospitalized for RSV infection where those children with symptoms of severe RSV bronchiolitis requiring mechanical ventilation express lower levels of IL-10 than hospitalized children with less severe disease [13]. Similarly, a recent study reported lower IL-10 levels in stimulated cord blood of children who were hospitalized for RSV infection before 6 months of age than in cord blood of infected infants who were treated as outpatients [36]. Furthermore, homozygosity for certain IL-10 alleles correlates with a higher risk of severe RSV bronchiolitis [12]. Such studies point to the importance of IL-10 in controlling the severity of acute infection with RSV.
The sources of IL-10 during acute primary RSV infection have not been clearly defined. While various cell types have been implicated as the source of this IL-10 [34], [37], [38], [39], [40], our results point to effector T cells as a potential major source of IL-10 produced during the acute phase of RSV infection. In this connection, we did observe a small increase of IL-10 released into the airways (BAL fluid) prior to the dramatic increase in IL-10 production observed in the respiratory tract at the time of effector T cell infiltration at day 5 post RSV infection. Thus, while one or more additional cell types may contribute to the pool of IL-10 observed during acute RSV infection, our results suggest that effector T-cells (both CD4+ and CD8+) may be a major source of IL-10 during human infection. Therefore the contribution of IL-10 from effector T cells should be rigorously evaluated in future studies of human RSV infection. Our analysis also suggests that the RSV specific effector T cells in the infected lungs serve as an important target of IL-10 action. Thus, the development and extent of immune mediated pathology in RSV infection may not only be linked to the level of IL-10 production by effector T cells (and potentially other cell types in the infected lungs) but also dependent upon the effectiveness of IL-10 signaling through the IL-10R on effector T cells leading to modulation of effector T cell function.
In summary, we have discovered a critical role of effector T cell-derived IL-10 in controlling the pulmonary inflammation and function during RSV infection. Furthermore, we established a previously unrecognized autocrine function of IL-10 in controlling proinflammatory activity of anti-viral effector cells. Our findings thus provide a cellular and mechanistic link to earlier clinical studies which implicate IL-10 in the pathogenesis of RSV disease and may provide the groundwork for future studies examining IL-10 as a therapeutic option in the treatment of RSV induced bronchiolitis in young infants.
BALB/c mice were purchased from Taconic Farms and Rag1-/- mice were purchased from The Jackson Laboratories. IL-10-eGFP reporter mice (Vert-X) were obtained from C.L. Karp from Children Hospital of Cincinnati. IL-10-/- mice were obtained from T. A. Wynn from NIAID, NIH. The T cell conditional IL-10Rα knock-out mice were generated through crossing IL-10Rα fl/fl mice to CD4-cre transgenic mice [23]. All mice were housed in a specific pathogen-free environment and all animal experiments were performed in accordance with protocols approved by the University of Virginia Animal Care and Use Committee. The A2 strain of RSV (obtained from P. L. Collins from NIAID, NIH) was grown in HEp-2 cells (ATCC) and titered for infectivity. We infected 10–12 week old BALB/c mice with a dose of ∼1–1.2×107 pfu RSV in serum-free Iscove's medium (Invitrogen) intranasally after anesthesia with methyl ether (Matix Scientific). For the conditional IL-10Rα mice infection (B6 background), we infected 9–12 week transgenic mice with a dose of ∼2×107 pfu RSV in serum-free Iscove's medium after anesthesia with ketamine and xylazine.
This study was carried out in strict accordance with the Animal Welfare Act (Public Law 91-579) and the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health (OLAW/NIH, 2002). The protocol was approved by the University of Virginia Animal Care and Use Committee (ACUC, Protocol Number: 2230).
We purified WT or IL-10-/- Thy1 positive T cells (including both CD4+ and CD8+ cells) from spleen and lymph nodes by MACS separation. 15 million T cells were then transferred into Rag1-/- mice i.v. 1 week after transfer, the recipient mice were infected with RSV.
We obtained BALF by flushing the airway multiple times with a single use of 500 µl sterile PBS through a cannula attached to a syringe. Cells were recovered from the suspension by centrifugation at 4°C. Supernatants were collected and stored at –80°C until use for the cytokine and virus determinations. Cytokines were measured by ELISA (BD Biosciences) according to the manufacturer manuals. The viral titer in the BALF was determined through the plaque assay in serially diluted Hep-2 cell cultures.
To measure RSV-L gene expression, we isolated RNA from the infected lungs via Trizol (Invitrogen) and treated it with DNase I (Invitrogen). We used random primers (Invitrogen) and Superscript II (Invitrogen) to synthesize first-strand complementary DNAs from equivalent amounts of RNA from each sample. We performed real-time RT-PCR in a 7000 Real-Time PCR System (Applied Biosystems) with SYBR Green PCR Master Mix (Applied Biosystems). The sequence of RSV-L gene primers was previously reported [41]. Data were generated by the comparative threshold cycle (ΔCT) method by normalizing to GAPDH. For experiments to measure host genes in effector T cells from infected Vert-X mice. We isolated CD44hi IL-10/eGFP+ or CD44hi IL-10/eGFP- CD8+ T cells and CD44hi IL-10/eGFP+ or CD44hi IL-10/eGFP- CD4+ T cells by FACS-sorting. We then isolated RNA, synthesized cDNA and performed real-time RT-PCR as described above. The sequences of the primers are available upon request.
Lung single cell suspensions were generated as described [42]. Lung cells were subsequently re-stimulated with either PMA (100 ng/ml) and ionomycin (1 µg/ml) in the presence of Golgi-Stop (1 µl/ml) for 5–6 h. Then cells were fixed and permeablized using the Cyto-Fix and Perm-Wash system (BD Biosciences). Cell surface CD4 and CD8 and intracellular IL-10 and IFN-γ were stained accordingly. Measurement of IL-10 and IFN-γ producing cells in vivo was based on a previously described protocol with modifications [22]. Briefly, at indicated days post RSV infection mice were injected i.v. with 500 µl of a PBS solution containing 500 µg Monensin (Sigma-Aldrich) 6 h before harvesting. Lung single cell suspensions were prepared in the presence of monensin. Cells were then fixed and permeablized and intracellular IL-10 and IFN-γ staining was as described [43].
α-IL-10R blocking mAb (clone 1B1.3A) and isotype control Rat IgG1 mAb were obtained from Schering-plough Biopharma and Bio-express. We achieved IL-10 signaling blockade in vivo by injecting α-IL-10R blocking mAb on day 1 (0.75 mg intraperitoneally in 500 µl), day 3 (0.15 mg intranasally in 40 µl) and day 4 (0.75 mg intraperitoneally in 500 µl).
The functional properties of the lung i.e. the airway and vascular mechanics are characterized by different lung function parameters, including the pulmonary compliance, the airway resistance and pulmonary artery pressure. The airway resistance is an index for the resistive forces against the airflow in the airways and depends on the diameter and length of the airways. The airway resistance can be calculated from the relation between transpulmonary pressure and airflow velocity. The airway resistance increases as a consequence of narrowing of the airways due to bronchoconstriction or obstructive processes, e.g. bronchial edema or enhanced mucus deposition. The pulmonary compliance is a marker for the functional stiffness of the lung and can be calculated from the relation between tidal volume and transpulmonary pressure. The pulmonary compliance decreases during restrictive pathological processes e.g. atelectasis, fibrosis, pulmonary edema or disturbed surfactant secretion. The increase in pulmonary artery pressure is an index of vasoconstriction of lung and represents an underlying pathophysiology resulting in pulmonary edema or changes in pulmonary vascular resistance. We measured the lung function using a buffer-perfused mouse lung system (Hugo Sachs Elektronik) as previously described [44]. Briefly, at day 7 post RSV infection, mice were anesthetized with ketamine and xylazine and ventilated with room air at 100 strokes/min with a tidal volume of 7 ml/g body weight with a positive end expiratory pressure of 2 cm H2O using the MINIVENT mouse ventilator (Hugo Sachs Elektronik). The pulmonary artery was cannulated via the right ventricle, and the left ventricle was immediately tube-vented through a small incision at the apex of the heart. The lungs were then perfused at a constant flow of 60 µl·g body wt–1·min–1 with Krebs-Henseleit buffer containing 2% albumin, 0.1% glucose, and 0.3% HEPES. The perfusate buffer and isolated lungs were maintained at 37°C throughout the experiment. Isolated lungs were allowed to equilibrate on the apparatus during a 5-min stabilization period. After equilibration, data were recorded for an additional 10 minutes. Hemodynamic and pulmonary parameters were continuously recorded during this period by the PULMODYN data acquisition system (Hugo Sachs Elektronik).
Effector T cells were purified from RSV infected lung at d7 p.i. and resuspended in complete media. Then T cells were stimulated with 20 ng/ml rIL-10 (eBioscience) in the presence or absence of JAK inhibitor (JAK inhibitor I, EMD Biosciences) and 15 min later the phosphorylation of STAT3 was determined through ICS following previousely described protocols [10]. To measure the inhibition of IFN-γ release by IL-10, we stimulated purified T cells (5×105 cells/ml) with plate-bound α-CD3 (100 µl of 0.1 µg/ml α-CD3 for 4 h at 37°C) overnight. Then the supernatant of the culture were collected and IFN-γ concentration were determined through ELISA.
All FACS antibodies are purchased from BD Biosciences or eBioscience. The dilution of surface staining antibodies was 1 in 200 and the dilution of intracellular staining antibodies was 1 in 100. After antibody staining, we examined cells using a six-color FACS-Canto system (BD Biosciences) and the data were analyzed by FlowJo software (Treestar). We characterized the various cell types according to their phenotypes as follows: neutrophils (Ly6G+CD11b+Ly6c–), monocytic cell lineage (Ly6G–CD11b+Ly6c+), natural killer cells (NK1.1+CD3−), CD8+ T lymphocytes (Thy1+CD8+) and CD4+ T lymphocytes (Thy1+CD4+).
Data are means ± SEM. We used two-tailed Student's t test for statistical analyses. We considered all P values >0.05 not to be significant.
|
10.1371/journal.pcbi.1005284 | An Atlas of Peroxiredoxins Created Using an Active Site Profile-Based Approach to Functionally Relevant Clustering of Proteins | Peroxiredoxins (Prxs or Prdxs) are a large protein superfamily of antioxidant enzymes that rapidly detoxify damaging peroxides and/or affect signal transduction and, thus, have roles in proliferation, differentiation, and apoptosis. Prx superfamily members are widespread across phylogeny and multiple methods have been developed to classify them. Here we present an updated atlas of the Prx superfamily identified using a novel method called MISST (Multi-level Iterative Sequence Searching Technique). MISST is an iterative search process developed to be both agglomerative, to add sequences containing similar functional site features, and divisive, to split groups when functional site features suggest distinct functionally-relevant clusters. Superfamily members need not be identified initially—MISST begins with a minimal representative set of known structures and searches GenBank iteratively. Further, the method’s novelty lies in the manner in which isofunctional groups are selected; rather than use a single or shifting threshold to identify clusters, the groups are deemed isofunctional when they pass a self-identification criterion, such that the group identifies itself and nothing else in a search of GenBank. The method was preliminarily validated on the Prxs, as the Prxs presented challenges of both agglomeration and division. For example, previous sequence analysis clustered the Prx functional families Prx1 and Prx6 into one group. Subsequent expert analysis clearly identified Prx6 as a distinct functionally relevant group. The MISST process distinguishes these two closely related, though functionally distinct, families. Through MISST search iterations, over 38,000 Prx sequences were identified, which the method divided into six isofunctional clusters, consistent with previous expert analysis. The results represent the most complete computational functional analysis of proteins comprising the Prx superfamily. The feasibility of this novel method is demonstrated by the Prx superfamily results, laying the foundation for potential functionally relevant clustering of the universe of protein sequences.
| Peroxiredoxins (Prxs) are a large, ubiquitous superfamily of proteins that are arguably the most important reductants of peroxide in biological systems. These proteins are involved in a diverse array of essential cellular functions, including peroxide reduction, signal transduction, circadian rhythms, chaperone function and apoptosis. Previously, Prxs have been classified multiple ways, based on biological role and evolutionary analysis. A more detailed expertly curated analysis identified six functionally relevant Prx classes and identified over 3500 proteins in these six classes; this set provides a validation for molecular function annotation methods. It is well-known that automated molecular functional annotation for individual protein sequences is difficult without detailed manual curation. In this work, we address this deficiency in available technologies by presenting a novel iterative method, MISST, for agglomeratively identifying superfamily members and clustering them into functionally relevant groups. Using this potentially automatable approach, 38,739 Prx sequences were identified from GenBank. MISST identified six functionally relevant clusters from these sequences, matching those previously identified by experts. Key mechanistic determinants and organismal distribution are explored. This analysis provides a significantly more complete understanding of this biologically important protein superfamily; the method lays a foundation for automated functionally relevant clustering of the protein universe.
| Peroxiredoxins (Prxs) are a large and ubiquitous superfamily of thiol dependent peroxidases, which have long been known to be involved in the reduction of aliphatic and aromatic hydroperoxides and peroxynitrite in biological systems [1–3]. Historically, these proteins have also been called TSA (thiol-specific antioxidant), AhpC (alkyl hydroperoxide reductase), and TPx (thioredoxin peroxidase). Prxs are known to protect cellular components from oxidative damage [4,5]. Indeed, it has been argued that Prxs are one of the most important peroxide scavengers in biological systems [6–9].
In addition to a peroxide scavenger role, Prxs are involved in essential biological processes such as redox signaling, which, because of the Prx reaction efficiency, can occur by one of two mechanisms. In the first mechanism, oxidation of redox-regulated proteins is not caused by H2O2 directly, but is rather mediated by Prxs, such that Prx CP is first oxidized by H2O2, which then reacts directly with the regulated kinase or phosphatase modifying its function. The regulated protein is subsequently regenerated by a cellular reductant. This signal transduction mechanism has been extensively reviewed [10–12]. In the second signaling mechanism, redox-regulated proteins may be directly oxidized by H2O2 [11,13–16]. However, thiol oxidation by H2O2 in redox regulated proteins is typically much slower in cellular proteins than the corresponding H2O2 detoxification by Prxs [17]. Thus, signal propagation occurs by Prx inactivation: Prxs are subject to H2O2 hyperoxidation at the active site cysteine, peroxidatic Cys (CP), which inactivates them (until they are repaired by the enzyme sulfiredoxin) [18,19]. The “floodgate hypothesis” posits that localized Prx inactivation (e.g. via hyperoxidation) serves to promote H2O2-mediated oxidation of redox-regulated proteins [20] and examples of such signaling in cells are emerging [21,22]. Hyperoxidation is also reported to play a role in circadian rhythms [23] and chaperone function [24]. Fine control of the Prx reaction mechanism is clearly essential; thus, understanding molecular function of this large and complex superfamily would provide insight into broader biological mechanisms.
As one would expect, peroxide detoxification and redox regulatory systems can be quite complex. For example, mammalian cells express six Prx isoforms: 2-Cys (PrxI, PrxII, PrxIII, and PrxIV), atypical 2-Cys (PrxV), and 1-Cys (PrxVI) [25]. Chloroplasts contain three Prx isoforms [26]. All Prxs contain CP preceded in the sequence by a conserved Pxxx(T/S)xxCP, a definitive motif for the Prx superfamily. An Arg is also absolutely conserved, but is contributed by a sequence fragment close in structure and distant in sequence. These residues activate the peroxide substrate, catalyze peroxide bond breakage, and catalyze attack of the CP thiolate on the substrate hydroxyl [27–29].
The extent and importance of the Prx proteins has led to several approaches to cluster the superfamily based on active site details. At its most simple, Prxs are classified into typical 2-Cys, atypical 2-Cys, and 1-Cys Prx families based on the presence or absence and position of a resolving Cys, CR [30,31]; however, proteins may have structural features of one of these classes, but mechanistic details of another [32]. Detailed sequence comparison and evolutionary analysis determined that Prxs diverged from an ancestor of the thioredoxin fold family and identified four classes of Prx, which these researchers called Prx1, Prx2, Prx3, and Prx4 [33]. Subsequent work based on detailed sequence analysis divided the Prx superfamily into six isofunctional families: AhpC/Prx1 (abbreviated Prx1), Prx6, Prx5, Tpx (thiol peroxidase), PrxQ/BCP (bacterioferritin comigratory protein, abbreviated PrxQ) and AhpE [28,32,34]. This level of detailed molecular functional annotation is typically lacking in the sequence databases, as we have previously shown [35].
More recently, we have used a bioinformatics approach based on active site profiling [36] to identify sequences in a given isofunctional family based on active site features [35,37]. Active site profiles (ASPs, Fig 1) are used to identify and compare functional site features. The Deacon Active Site Profiler (DASP), a tool that uses ASPs to search databases for sequences containing active site features similar to those in the ASP [37,38] identified many additional Prx members of each expertly identified isofunctional group [35]. Using this single search approach, we identified over 3500 proteins in the six Prx functional subgroups; these sequences are available in the Prx database, PREX [39] and in the Structure Function Linkage Database, SFLD [40,41]. SFLD curators subsequently added sequences to these groups using their hidden Markov model (HMM) approach.
The significant question is: could one automatically identify such isofunctional families within a protein superfamily without expert analysis? Databases such as CATH [42,43], PFAM [44,45] and SCOP [46,47] have clustered large superfamilies of proteins based upon domain characteristics and/or structural and sequence classification. Such approaches capture broad levels of functional similarity. On the other hand, in SFLD, proteins are clustered based on functional similarity [41]. An SFLD superfamily contains proteins that share part (but not all) of their enzyme mechanism. At a more detailed level, SFLD families contain proteins which exhibit the same enzyme mechanism (i.e., are isofunctional). CATH PFAM, and SCOP families are more similar to what is defined as an SFLD superfamily [40]; such broad groups usually contain multiple isofunctional families. Our goal is to develop a method to more automatically identify isofunctional clusters.
Several approaches aim to cluster sequences into isofunctional clusters, including FunFHMMer [48] (an updated version of GeMMA [49]), SCI-PHY [50], and ASMC [51]. These methods start with known superfamily sequences and subdivide that large set using clustering and pattern recognition of full sequences. SCI-PHY starts with a multiple sequence alignment, builds a hierarchical tree using agglomerative clustering, and identifies the point at which to prune the tree. SCI-PHY includes phylogenetic details in the clustering. ASMC starts with a PFAM family, uses modeling and analysis of specificity determining positions (SDPs) to cluster the PFAM family, and structural modelling to create active sites; ultimately structural comparisons are performed to identify functional groups. FunFHMMer starts with and clusters a CATH-Gene3d superfamily. Essentially, FunFHMMer builds weighted HMMs of the identified clusters, so new members of each group can be identified. Both ASMC and FunFHMMer identify SDPs or mechanistic determinants that are weighted heavily in creating profiles. Remaining challenges focus on determining when subdivision is complete and identifying the SDPs more automatically.
The method described here, MISST (Multi-level Iterative Sequence Searching Technique), presents a novel approach to identifying functionally relevant clusters. Previous methods start with the complete superfamily and divisively cluster that superfamily, while the current method begins with a few examples and agglomeratively builds the isofunctional clusters from those representatives. To define groups in MISST, we build on the observation that suggests if a group is isofunctional, a DASP search using that group as the input profile self-identifies its members and no other proteins, while groups that are not isofunctional do not self-identify in this way [52]. That is, a group of proteins is deemed a functionally relevant cluster if a database search (using DASP) returns all proteins in the group at significant scores and no (or few) other proteins at significant scores (within a range of uncertainty). The iterative searches of MISST are built on this observation.
The first step in this approach is to identify the starting set of isofunctional clusters, a process called TuLIP (Two-Level Iterative clustering Procedure), during which proteins of known structure that share common active site features are clustered [52]. This process is also built on the same premise: an isofunctional cluster is one that self-identifies in a DASP search. Briefly, TuLIP starts with all structures from a protein superfamily and iteratively subdivides those into smaller and smaller groups based on active site features. At each iteration, each cluster is used in a DASP search of the sequences in the PDB. For each cluster, if the DASP search self-identifies–that is all proteins in the cluster are identified in the search and nothing else–that cluster is deemed a functionally relevant group. All clusters that do not pass this criterion are further subdivided and searched again. Results on the enolase superfamily demonstrate that TuLIP does identify the functionally relevant subgroups and families [52].
In this work, a comprehensive atlas of the Prx superfamily is identified through application of the TuLIP and MISST processes. Four functionally relevant clusters were identified by TuLIP from the known Prx structures. Through MISST iterations, sequences are added to the groups and the four clusters are subdivided into six clusters which correspond to the six expertly identified functionally relevant groups, even though this expert information of six groups was not input into the process. Because TuLIP and MISST involve iterative DASP searches, a modified process, DASP2, was used in this work. DASP2 database search results are essentially identical to DASP search results, however DASP2 is significantly more efficient than DASP [53].
This agglomerative and divisive approach allowed us to assign molecular functional detail to over 38,000 sequences, many of which were previously uncharacterized or annotated as a general Prx (or one of its synonyms). The current work suggests the feasibility of automation of MISST. Though more testing and validation is required, the MISST process should be generally adaptable for the analysis of other protein superfamilies to produce high-quality molecular function annotation and identification of isofunctional clusters within the protein universe.
Identification of functionally relevant clusters among proteins of known structure is the first step in our process and is accomplished using TuLIP, a two-stage approach to clustering structures based on active site features [52] (see Methods for details). TuLIP identifies four functionally relevant clusters from 47 non-redundant peroxiredoxin (Prx) structures: three clusters (Sct2, Sct3, and Sct4) during the first stage and one (Rlx6) during the second stage (Fig 2A).
A good, though not perfect, correspondence is observed between expertly-identified subgroups, as deposited in SFLD, and TuLIP-identified groups (Fig 3A). Prx5 maps one-to-one to TuLIP group Sct3. TuLIP group Sct4 contains all proteins in three Prx subgroups: AhpE, Prx1, and Prx6, a result suggesting similar active site features, which is, indeed, observed (S1 Fig). Prx1 and Prx6 had previously been identified as being closely evolutionarily related, as well [33]. All Tpx proteins are identified in TuLIP group Sct2; Sct2 also contains four PrxQ proteins (Fig 3A). The two other PrxQ structures were grouped into their own cluster, Rlx6. This subdivision of the PrxQs of known structure was previously observed in hierarchical clustering of active site signatures [35]. Hierarchical clustering based on the canonical Prx active site motif (S1 Fig) suggests that residue differences at the PrxQ active site of proteins of known structure are driving this subdivision.
The TuLIP clustering results are not unexpected from the limited dataset of known structures and what is known about functional similarities. However, the results do present a challenge for the agglomerative and iterative process of searching sequence space: an ideal process would subdivide Sct4 into the expertly identified functionally relevant clusters and would recombine the PrxQ subgroup.
MISST is an iterative search process developed to be both agglomerative and divisive. That is, the process was developed to add (agglomerate) sequences containing similar functional site features to each TuLIP group and to subdivide TuLIP groups when functional site features suggest distinct clusters. As an illustration, MISST should identify the two groups represented in the ASP in Fig 1B without curator intervention. The MISST process is outlined in Fig 4 and described in detail in Methods. Briefly, the process involves iterative DASP2 searches of GenBank, each followed by evaluation for cluster division, combination, and self-identification. DASP2 is a more efficient version of the DASP sequence-searching method that focuses not on the complete protein sequence, but rather only on a protein’s functional site features [37,38,55]. Groups defined by MISST should, thus, be identified and subdivided based on their mechanistic differences. Notably, no step in the MISST process requires human evaluation—the process should be automatable, although adjustment of two parameters may be needed once the process is automated.
ASPs were created from sequences in each of the four TuLIP-identified groups: Sct2, Sct3, Sct4, and Rlx6 (profiles are provided in S1 File). Each ASP was used as input into an iterative process of DASP2 GenBank searches (see MISST flow chart, Fig 4A). Following each iteration, each group was evaluated for self-identification (Fig 4A) and need for subdivision (Fig 4C). If a group self-identifies, it is removed from the iterative process and set aside for final analysis. For all other groups, a new ASP is created from functional site pseudo-signatures (see Methods) of sequences identified at scores ≤1e-12 (Search0) or ≤1e-14 (subsequent search iterations; see Supplemental Methods in S3 File for justification, validation, and broader applicability of these score thresholds).
Five search iterations (Search0 through Search4) were performed (Fig 2B). All groups satisfied self-identification criteria after Search3 except Rlx6_PrxQ which satisfied the criteria after Search4. Through the iterations, sequences were added to each group and the four original TuLIP groups were divided into six. The process of adding sequences and splitting groups is represented in the dendrogram in Fig 2B; proteins found in the final groups are visually represented as networks in Fig 2C. Qualitatively, the six groups correspond almost perfectly with the six functionally relevant groups previously identified by experts [35] (Fig 3B).
These searches identified 38,739 sequences (Table 1) in six groups (DASP2 score threshold ≤1e-14). Proteins identified in each cluster are provided in S2 File. 6,855 of these proteins are annotated in SFLD to the subgroup matching the MISST group [41]. 30,096 proteins were not previously identified by a single DASP search [35] or by SFLD HMM analysis (Table 1); new sequences were identified due in part to their absence from the GenBank database during earlier analyses and to the more robust analysis method used here. To ascertain whether all 38,739 proteins are likely Prx superfamily members, we determined how many contained the canonical Prx active site motif Pxxx(T/S)xxCP [3,56,57]. Across all searches, this fragment is found in 99.3% of all MISST-identified sequences, indicating almost all sequences likely belong to this superfamily.
We next explore how the MISST process agglomerates sequences and subdivides groups. We then quantitatively compare the MISST-identified groups to the previously identified sequences. Because MISST utilizes DASP2 with its focus on functional site features as the search mechanism, we can hypothesize mechanistic determinants important for each group’s function and compare the functional site features of these expanded groups to those described by experts [35,58].
MISST iterations initiated with seven Prx5 proteins in TuLIP group Sct3 ultimately identify 5434 proteins. This coherent group was not further split by PSSM analysis (Fig 2B, purple dendrogram branch), likely because of the strong intragroup active site similarity. 1039 of the MISST-identified proteins are identified as Prx5 sequences in SFLD, representing 97.8% coverage (recall). The Prx5 proteins deposited in SFLD were identified through one DASP iteration [35]; a few more were added through the SFLD curation processes. This group contains no proteins from any other Prx subgroup (Fig 3B); consequently, Sct3 is mapped to Prx5 for subsequent analysis and herein called Sct3_Prx5.
Sct3_Prx5 includes 252 proteins identified in SFLD as belonging to the Prx superfamily, but uncharacterized with respect to subgroup; thus, the functional subgroup of these proteins can now be defined more precisely. 4143 Sct3_Prx5 proteins were not previously identified as Prx5 (Table 1) demonstrating that, if the new identifications are correct, search iterations of MISST add significantly to our knowledge of functionally related proteins. Consequently, the probability that these proteins are actual Prx5 proteins was evaluated by determining the presence or absence of the Prx5-specific active site motif P(G/A)A(F/Y)(T/S)(P/G)xCP [9] (Fig 5A, part of red brace). 97.4% of all Sct3_Prx5 sequences contain this motif. The percentages do not differ between previously known and newly identified proteins: 98.2% of previously identified Prx5 proteins, 96.4% of Prx sequences in SFLD that are uncharacterized relative to subgroup, and 97.2% of new (non-SFLD) proteins contain the motif. Given that the percent of both knowns and new proteins containing this motif is similar, there is high probability that the MISST iterations are consistently identifying proteins that belong to the Prx5 functional family.
To quantitatively evaluate sequence identification, F-measure, the harmonic mean of precision and recall [59], was calculated for Sct3_Prx5 sequences. For this analysis (and similar analyses of other groups), Prx proteins in SFLD are the known sequences; “positive” sequences are the proteins in the subgroup under consideration, while “negative” sequences are Prx sequences in all other subgroups. Thus, if a known Prx5 was identified by MISST, a true positive was counted. If a sequence from another Prx subgroup was identified as part of Sct_Prx5, a false positive was counted. False negatives were Prx5 sequences identified in the previous work [35], but not identified in this search. A true negative is counted if MISST did not identify Prx sequences known to be members of other Prx subgroups. Sequences identified by MISST, but not by previous methods, were not included in this analysis, as their assignment as true or false positives or negatives could not be evaluated. This is a difference between MISST and other methods: instead of subdividing a superfamily in which all proteins are thought to be known at the start [49–51], MISST agglomeratively adds proteins from the database and subdivides the groups.
F-measure analysis demonstrates the high quality of assignments to Sct3_Prx5 (Fig 6A): the F-measure is 0.99 at the DASP2 search score threshold (≤1e-14, dashed line Fig 6A). As the DASP2 score threshold becomes more significant, recall gradually decreases (as some proteins are missed); however, precision never drops below 1 for Sct3_Prx5. Neither precision nor recall decrease in this group as the DASP2 score threshold becomes less significant (yellow, orange, and red bars, Fig 6A), indicating no false positives are identified at ≤1e-8, even prior to cross hit analysis (Fig 6B).
Detailed analysis of the Sct3_Prx5 functional site pseudo-signatures identify mechanistic determinants distinctive to this subgroup (Fig 5A; structures in S2 Fig). These determinants were not identified a priori as input. The Prx active site motif includes elements distinctive to the Sct3_Prx5 subgroup: P(G/A)A(F/Y)(T/S)(P/G)xCP (bold indicates residues almost invariant across the superfamily; [9,35]). Outside of this motif, two defining features are observed: His is almost invariant at signature position 15 (Fig 5A red brace) and a pair of Arg residues (RSxR(Y/F)) at positions 33–37 (yellow brace). The second of these conserved Arg residues is the one recognized to play a major role in activating the peroxide substrate for–O-O–bond scission at the Prx active site [9]. In the structure 1TP9, the side chain of the His residue conserved in Prx5 proteins is hydrogen bonded to the side chain of this invariant Arg (signature position 36; S2A Fig). The location of these side chains in the active site near the CP suggests a role in mechanism, perhaps with the His playing a role in proton transfer. Reasonably well conserved motifs in the pseudo-signatures of this subgroup also include VMxxW at signature positions 20–24 and (C/V)(V/L/I)(S/A)VN at signature positions 39–43 (Fig 5A, purple and fuchsia braces, respectively). The Cys in this second fragment is found in 76% of sequences; 19% of sequences have Val at this position. Further, phylogenetic evidence suggests conservation of this Cys, which sometimes serves as the CR, may be based on phylogeny (S3 File; S3 Fig, red brace).
Starting with just nine structures, MISST agglomerates sequences into a coherent Prx5 cluster. Even though PSSM analysis was performed at each iteration, the Sct3_Prx5 group did not split, suggesting that the PSSM approach does not split functionally relevant clusters.
Sct2 was originally comprised of four PrxQ sequences and nine Tpx sequences (Fig 3A). Known Tpx structures contain the resolving cysteine, CR, in the α3 helix. The CR is not found in a consistent location in the four TuLIP-identified PrxQ proteins. Using the Sct2 TuLIP group as MISST input illustrates sequence agglomeration and increasing coherence within a cluster, despite the group’s initial heterogeneity. At Search0, known PrxQ proteins are identified at less significant DASP2 search scores than the Tpx proteins (Fig 7A). By the second iteration (Search1) known PrxQ proteins are not identified (Fig 7B). Iterative DASP2 searches produce more robust profiles and each successive search produces a more coherent set of sequences that exhibit common active site features.
With each iteration, additional Tpx proteins accumulate, with a plateau reached in Search2 and Search3 (Fig 7C). At Search3, ≥ 95% of sequences used as input to Search3 and ≤ 15% new sequences were identified at significant DASP2 scores (≤1e-14) in the GenBank Search3; thus, self-identification criteria were satisfied following Search3 (Fig 2B, green dendrogram branch). At this point, the group was homogeneous for Tpx proteins and is thus called Sct2_Tpx.
The final Sct2_Tpx cluster contains 4930 sequences—860 are in SFLD and annotated to the Tpx subgroup, 244 are marked as Prx-uncharacterized in SFLD, and 3826 are not in SFLD (Table 1). F-measure shows high precision and recall values for Tpx proteins (Fig 6A). After this group satisfied the self-identification criteria, no false positives were identified even at less significant scores of ≤1e-8. 860 SFLD Tpx sequences represent 90.1% coverage (recall) of known subgroup members; F-measure is 0.95 at the DASP2 score threshold of ≤1e-14 (Fig 6A).
Of the final sequences in this Sct2_Tpx group, 98.5% contained the Prx active site motif distinctive for this subgroup: PS(I/L/V)DTx(V/T/I)CP (Fig 5B, red brace), which refines the motif determined from the previously identified smaller dataset [9,35]. The sequences are 99.94% bacterial (S4A Fig), consistent with what was previously reported on the smaller dataset [58].
Additional mechanistic determinants can be hypothesized for the Sct2_Tpx subgroup. Signature positions 15 and 16 are distinctive in this group: a branched residue (Val or Thr) followed by Arg or Lys (Fig 5B, red brace). A distinctive AxxR(F/W)C motif is observed at signature positions 20–25 (Fig 5B, purple brace). This conserved Cys is the CR in helix α-3. As in Sct3_Prx5 and Sct4_Prx1, the nearly invariant (99.3%) Arg at signature position 36 is the active site residue required for efficient catalysis [9,60]. In the structure 3P7X, the side chain of this Arg is hydrogen bonded to CP. It is preceded by a very well-conserved Leu at signature position 33, the only subgroup with a well conserved hydrophobic residue at this position (Fig 5B, yellow brace). Both Arg (gray) and Leu (black) extend towards CP (S2B Fig). Finally, the Sct2_Tpx subgroup contains a Ser that is almost invariant at signature position 42 (Fig 5B, fuchsia brace). These residues are proximal to the active site, suggesting a functional role (S2B Fig).
This example illustrates how the iterative MISST process creates more coherent groups, even when the original TuLIP group is composed of two subgroups. While the PrxQ structures were not present in the final Sct2_Tpx MISST group, this subgroup was not lost in the MISST process. As discussed subsequently, the PrxQs were identified as a subdivision of the Rlx6 group.
The clustering process described herein starts with proteins of known structure; however, the structure database is a very limited representation of the sequence space universe. Because of this limitation, TuLIP sometimes combines multiple subgroups into one cluster [52], as is the case with Sct4, which contains both Prx1 and Prx6 proteins. Consequently, any agglomerative process aimed at identifying functionally relevant groups must recognize the need for cluster subdivision. PSSM Analysis was developed as an automatable process to do just this.
PSSM Analysis is performed at each MISST iteration after the first (Fig 4A) using the outlined process (Fig 4C; details in Methods). Essentially, the active site pseudo-signatures identified in the GenBank search are used to quantitatively determine if and how the group should be subdivided. If subdivision is required, two new ASPs are created from the appropriate pseudo-signatures. These ASPs are input to a DASP2 search of GenBank. Search outputs are compared to verify the groups are, in fact, unique. Notably, PSSM Analysis was performed at each search iteration for both the Sct2 and Sct3 MISST groups, but distinct, functionally relevant groups were not identified within either group.
The TuLIP-identified Sct4 group includes all known structures from the Prx1, Prx6, and AhpE subgroups (Figs 3A and 4A). At Search1, PSSM Analysis identifies two groups; these groups evolve distinctly through subsequent search iterations (Fig 2B, red and blue dendrogram branches). Notably, though the AhpE subgroup is not identified in Sct4 after Search1, the AhpE subgroup is not lost. It is ultimately identified in Rlx6 using this same PSSM Analysis procedure (discussed subsequently).
Analysis of each search iteration provides insight into the PSSM Analysis of Sct4 (Fig 8). The Search1 DASP2 score distribution is bimodal—Prx1 sequences at more significant and Prx6 sequences at less significant DASP2 search scores (Fig 8A, blue and red bars). PSSM Analysis correctly identifies these two groups (Fig 8A, yellow and green boxes). One ASP is created each for sequences in the yellow and green boxes; each ASP is used in Search2 of GenBank. Prx1 and Prx6 sequences are identified distinctly in Search2 (Fig 8B, Search2 distributions). After just one more GenBank search iteration (Search3), each group passes self-identification criteria.
Ultimately, 9660 and 5212 sequences are identified at significant DASP2 scores in Sct4_Prx1 and Sct4_Prx6, respectively (Table 1). Of the proteins annotated in SFLD, 96.6% of Prx6 proteins and 95.7% of Prx1 proteins are identified (Table 1). Both searches identify Prx sequences annotated as Prx-uncharacterized in SFLD: 127 and 289 are identified as part of Sct4_Prx6 and Sct4_Prx1, respectively. Finally, 4143 and 7241 GenBank sequences not annotated in SFLD were identified as Sct4_Prx6 and Sct4_Prx1 members, respectively (Table 1).
The Prx active site motifs for Sct4_Prx1 and Sct4_Prx6 are distinct: PxDF(T/S)FVCP and Px(D/N)(F/Y)TPVCP, respectively (Fig 5C and 5D, red braces). 93.8% and 96.7% of all sequences in Prx1 and Prx6, respectively, exhibit these motifs, demonstrating that MISST iterations and the PSSM Analysis distinguish these small active site differences. F-measure at the score threshold of ≤1e-14 is high for both groups: 0.98 at a DASP2 score threshold of ≤1e-14 for each (Fig 6A). Thus, PSSM Analysis can effectively subdivide one group into two functionally relevant clusters.
As with the other groups, we can identify mechanistic determinants that distinguish Sct4_Prx6 and Sct4_Prx1. A key distinguishing feature is the TFVC versus TPVC for Prx1 and Prx6, respectively: this one residue in the canonical Prx active site motif distinguishes these two subgroups (Fig 5C and 5D, red brace; S2 Fig, cyan side chains). Another distinguishing feature is a Phe-Tyr (Prx1) compared to Ser-His (Prx6) at signature positions 2 and 3 (Fig 5C and 5D, blue brace). In 2V2G, this His is in the active site, near the CP (S2D Fig, yellow side chains). Again, Arg at position 36 in Prx1 is the active site residue required for efficient catalysis; the fragment containing this Arg is not part of the Prx6 signature. In both subgroups, the almost invariant Ser (at signature position 42) and the almost invariant His (at signature position 21) form a potential path for proton transfer in these subgroups (S2C and S2D Fig, light pink side chains). CR is not observed within the Prx1 group profile because it is contributed from a different chain (the partner subunit of the dimer). There is no CR in most Prx6 members [35]. Interesting phylogenetic observations at specific positions, including the well-known GG(L/I/V)G motif [31], are discussed in S3 File.
Previous sequence analysis methods identified Prx1 and Prx6 as only one group, which the authors named Prx4 [33]. Subsequent expert analysis clearly showed that Prx6 was a distinct functionally relevant group [35]. MISST, a method that focuses on differentiating active site features, has accomplished that which was previously accomplished only by expert curation—to divide these two closely related isofunctional clusters without human curation. This opens the exciting possibility of functionally relevant clustering of superfamilies for which functional groups are not known.
PrxQ and AhpE were members of original TuLIP groups, but were lost from Sct2 and Sct4 searches, respectively, during MISST iterations. TuLIP group Rlx6 contained two of the six PrxQ structures known at the time this research was completed. The task is even more difficult because AhpE is a very small subgroup containing only 25 proteins in 2011 [35] and 112 in the current SFLD; previously, these proteins were found in only one class of bacteria (actinobacteria) [58]. Only one structure is available in the PDB database. Are these groups that are less well represented by structures identified through the iterative MISST process applied to TuLIP group Rlx6? The answer to this important question is yes.
Analysis of the Rlx6 MISST search iterations illuminates how AhpE and PrxQ sequences are identified and subdivided (Fig 2B, pink and yellow dendrogram branches). The Search0 ASP input contained only two PrxQ proteins (Fig 2A); Search0 output contained mostly PrxQ proteins, with a few AhpE proteins (not shown). Per the MISST process (Fig 4), an ASP was created for sequences identified at a DASP2 score threshold of ≤1e-12. This ASP was input to Search1.
Search1 output includes a small number of AhpE and PrxQ proteins at more significant scores; most PrxQ proteins are identified at less significant DASP2 scores (Fig 9A). PSSM Analysis divides Search1 sequences into two groups: AhpE and PrxQ (Fig 9A, blue and green boxes). An ASP is created for each group, and each ASP is input to DASP2 Search2. Results are distinct: one search is populated with mostly AhpE and a few PrxQ proteins, the other populated almost solely with PrxQ proteins (Fig 9B). The Rlx6_AhpE and Rlx6_PrxQ groups subsequently remain distinct (as determined by the agreement criterion; Fig 4B) and pass self-identification criteria at Search3 and Search4, respectively (Fig 2B, yellow and pink dendrogram branches).
The two groups map easily to subgroups identified by experts. One, Rlx6_PrxQ, contains 12,014 sequences; 1786 of these sequences are found in SFLD, which represents 92.1% of known PrxQ proteins (Table 1). 739 sequences are annotated in SFLD to the Prx superfamily but not a specific subgroup. MISST identifies 9489 sequences in this cluster that were not previously assigned to the Prx superfamily (Table 1). Consistent with the other MISST-identified groups thus far discussed, F-measure (and, thus, precision and recall) is quite high, 0.96, for Rlx6_PrxQ (Fig 6A).
Rlx6_AhpE is by far the smallest subgroup identified by MISST: only 1489 sequences are identified in this cluster. 98 of those proteins are currently annotated as AhpE in SFLD, which represents 87.5% of the 112 known AhpE proteins. 1254 Rlx6_AhpE proteins were not previously identified as Prxs (Table 1). Notably, F-measure for this cluster is not as strong as the other MISST-identified groups—only 0.74 at the DASP2 score threshold of ≤1e-14. In addition, the F-measure is never above 0.78, even at less significant score thresholds (Fig 6A). Detailed analysis explains this result. There are 112 nonredundant AhpE sequences in SFLD. At thresholds of ≤1e-14, ≤1e-12, and ≤1e-10, we identify 98, 107, and 108 of them, respectively; thus, recall is high, at 87.5% at ≤1e-14 and increases to 96.4% at ≤1e-10. However, 54 Rlx6_AhpE proteins identified at the DASP2 search score threshold of ≤1e-14 were previously identified as PrxQ subgroup members [35]. These proteins decrease the precision of the result. The question of functional assignment of these 54 sequences is an important one; these sequences are listed in S2 File and discussed subsequently.
The two clusters derived from Rlx6 exhibit common active site features, such as the Phe at signature position 2 (Fig 5E and 5F, blue brace), which is highly conserved in both groups. However, the Prx active site motif is distinct between Rlx6_BCP and Rlx6_AhpE, including the canonical Prx active site motif: P(K/A/R)(D/A)xTxGC and PxAF(T/S)xxC for Rlx6_PrxQ and Rlx6_AhpE, respectively (Fig 5E and 5F, red braces). 90.2% of proteins identified in Rlx6_PrxQ contained its motif, while 94.2% of the Rlx6_AhpE sequences, including 92.9% of those previously identified as AhpE and 88.9% of those previously annotated as PrxQs, contain its motif. Notably, Rlx6_PrxQ is the only subgroup with a Gly strongly conserved immediately preceding CP, which might suggest unique conformational or dynamical motion associated with PrxQ function.
Other positions also distinguish these two Rlx6-derived groups. Glu is invariant at signature position 14 in Rlx6_AhpE, while the residue can be either Glu or Gln in Rlx6_PrxQ. The final two active site fragments are also distinct (Fig 5E and 5F, fuchsia and purple braces, respectively). A G(V/I)SxD motif at positions 40–44 and a Leu at position 49 are strongly conserved in Rlx6_PrxQ. Notably, the invariant Gly, Ser, and Asp of the G(V/I)SxD motif are all in the 5ENU active site, along with the conserved Leu. These distinctive features suggest that, indeed, these two subgroups are functionally distinct.
The question remains: what is the correct functional classification of the 54 sequences previously classified as BCP [35] and classified by MISST as AhpE? A closer analysis of the active site signatures may explain the unexpected clustering. Signature logos were created for the 1786 Rlx6_PrxQ sequences that were previously annotated as PrxQ, the 98 Rlx6_AhpE proteins previously annotated as AhpE, and the 54 Rlx6_AhpE sequences previously annotated as PrxQ (Fig 10). Multiple positions in the active site signature illustrate why the 54 sequences previously annotated PrxQ are now identified in the AhpE MISST group, including a strongly conserved Ala-Phe dyad in the canonical PrxQ active site motif (PxAF(T/S)xxC), a conserved Val or Ile immediately preceding CP, and four other positions (Fig 10, orange highlights). These results demonstrate the DASP2 method identified these 54 proteins in the Rlx6_AhpE subgroup because of common features at the active site. Further, specific residues can be identified that distinguish bacterial (81%) and archaeal (19%) proteins in the Rlx6_AhpE subgroup (full discussion in S3 File). The biological relevance of these observations remains to be determined.
In conclusion, the original Rlx6 TuLIP group contained just two of six PrxQ structures, and the lone AhpE structure was in Sct4, not Rlx6. Despite only one known AhpE structure, PSSM Analysis and iterative DASp2 searches extracted the AhpE functional group from the Rlx6 search results. These results demonstrate the MISST process can identify functional groups for which structural representation is limited. This is an important result, as many protein superfamilies do not contain comprehensive structural representation in all functional families. Over all six Prx functional groups, the iterative MISST process meets the challenges presented by the TuLIP results: all six Prx subgroups were identified in a robust and comprehensive fashion, even though not all groups were well-represented in the structure database.
Results presented thus far demonstrate MISST can both add sequences to functionally relevant groups and subdivide groups into clusters exhibiting distinctive functional features. F-measure (precision and recall, Fig 6A) was described for each group individually. Further quantitative comparison between groups, including cross-hit counts and measures of performance, are essential to determine if these groups are distinct and functionally relevant.
Cross-hits are defined as the same sequence identified in more than one MISST group at a given DASP2 search score threshold. This analysis demonstrates discreteness of MISST groups. In creating the final groups, a cross-hit analysis similar to that previously described [35] is performed (see Methods). Only 20 proteins are removed in this final cross-hit analysis; the identities of the proteins which cross-hit are listed in S1 Table.
To understand the discreteness of the MISST-identified groups, the correlation of cross-hits (counted prior to this final cross-hit analysis) with the DASP2 search score threshold was evaluated (Fig 6C). At DASP2 search score thresholds of ≤1e-16 and more significant, all groups are distinct—the number of cross-hits is zero. At the significance threshold of ≤1e-14, the threshold identified as a “trusted” threshold in the work described here (see Supplemental Methods in S3 File), 20 cross-hits are identified corresponding to a cross hit rate of 0.052%, an extremely low false positive rate (Fig 6C, table and red data point).
Cross-hits increase drastically as the DASP2 search score threshold decreases in significance (Fig 6C). We can observe the evolution of the cross-hits and, thus, better understand the relationship between group active sites by analyzing “fireworks plots,” a form of network analysis (S5 Fig). At a DASP2 search score threshold of ≤1e-8, Prx5 and Tpx subgroups are most distinct and only exhibit a few cross-hits to other groups, which are mostly gone at a threshold of ≤1e-10 (S5A and S5B Fig). At a score threshold of ≤1e-12, the other four groups become more distinct (S5C Fig). At a score threshold of ≤1e-14, only twenty cross hits remain. Ten of these twenty cross-hits at ≤1e-14 are between AhpE and PrxQ (S5D Fig), indicating the functional sites of these groups are more closely related to each other than they are to the other groups, as discussed above. The other ten cross-hits are distributed between Prx1, Prx6, and PrxQ (S5D Fig).
The remaining analysis, F-measure and Performance, assumes that the expert annotations deposited in SFLD [35] are correct. These sequences were identified by a single DASP search of GenBank using expertly-created ASPs. Subsequently, sequences were added using the SFLD HMM approach. The resulting sequences were curated by hand and deposited in SFLD; these annotations are the best known molecular functional annotations for the Prx superfamily. Only 412 sequences previously identified as Prx are not identified as part of the proper MISST group (out of 7267 Prxs in SFLD with subgroup annotations). These sequences are evenly spread over the six groups and are counted as false negatives in the recall calculation of F-measure. 54 of these are the sequences previously annotated as PrxQ, but identified in this analysis as AhpE. 194 sequences were identified above the DASP2 score threshold of 1e-14. About 25 of the sequences are no longer in GenBank.
Since the 2011 analysis of the Prx superfamily, GenBank has grown from 11.9M proteins to over 54.8M proteins at the end of 2015. With this growth comes many new sequences identified in our MISST searches that are not annotated in SFLD. To quantify the performance of MISST, all sequences not annotated in SFLD were not used for the F-measure and Performance analyses as the correct annotation is unknown. (In the previous sections, we demonstrated the likelihood that these newly identified sequences were Prx by evaluating the presence and absence of the canonical Prx active site motif, Pxxx(T/S)xxCP, as well as the active site motif associated with each subgroup.)
To analyze the overall accuracy of the MISST process, a performance score was calculated [49,50] taking into account purity, edit distance, and VI distance [62] (Fig 6D). These measures were calculated by defining the proteins in each group as TP, TN, FP, or FN; these definitions were based on the previous Prx annotation [35] (see Supplemental Methods in S3 File). Purity provides a measure of the proportion of groups which contain only one subgroup. As Rlx6_AhpE is the only group containing false positives, purity remains at 83.3% (5 out of 6 groups are pure) until highly significant DASP2 search score thresholds (Fig 6D, blue). Edit and VI distances measure how many changes are required to transform one grouping method (MISST) to another (SFLD). The high correlation between the six SFLD subgroups and the six MISST groups leads to low edit and VI distances, particularly at less significant score thresholds (Fig 6D, black and red). The increase in edit and VI distance values at more significant scores is due to the presence of “singlets,” which in this case are Prx proteins in the SFLD not identified as a member of any MISST group. Typically, edit and VI distances are used to compare two clustering methods which both start with the same set of proteins. However, MISST is an agglomerative method and does not begin with the full set of proteins; therefore, some proteins in the SFLD are not identified by MISST. Thus, as the DASP search score threshold becomes more significant, more proteins are classified as “singlets” because they are not identified in any MISST groups at the given threshold.
Purity, edit, and VI distance were combined into an overall performance measure (Fig 6D, purple). A maximal performance score of 90.3 is found at a score threshold of ≤1e-8; the performance at the threshold of ≤1e-14 is 88.9 (Fig 6D, colored arrows). Performance does not reach 100 at any point because not all known Prx proteins are identified by the six MISST groups and some PrxQ-annotated proteins are identified in the Rlx6_AhpE group. Performance increases slightly at the less significant score thresholds, simply due to the behavior of edit and VI distance with “singlets.” The value of 88.9 at a score threshold of ≤1e-14 compares well with the performance values reported for clustering of other gold-standard SFLD superfamilies (amidohydrolase, crotonase, enolase, HAD, terpene cyclase, VOC) by SCI-PHY (performance ranged from 54.99 to 91.70, with an average of 75.36) and GeMMA with a generalized cutoff (performance ranged from 53.64 to 90.70, with an average of 80.42) [49]. It is important to note that performance scores vary widely for both SCI-PHY and GeMMA, indicating more superfamilies must be tested using MISST to complete a full-scale comparison between methods. However, this initial test using the Prx superfamily demonstrates the feasibility of the current approach.
Previous work has illustrated how different comparison measures (sequence, structure, functional site) can produce different clusters within a protein superfamily [54]. Here we explore that further, by evaluating full sequence similarity between the functionally relevant MISST clusters.
A representative network (RepNet) was built from the 38,739 sequences identified in the six MISST groups. Each of the 1,369 nodes represents proteins sharing 55% sequence identity; each edge represents the pairwise BLAST score (sequence comparison) between the representatives of the two nodes. Nodes are colored based on the MISST group to which the sequences belong (see Methods for more details). The network is filtered at a variety of BLAST score thresholds to visualize the full length sequence similarity among the MISST groups (Fig 11).
As the threshold for edge becomes more stringent, groups begin to separate. Notably, and as expected, the sequences within MISST groups are more similar to each other than they are to proteins from other groups. This observation is the reason that full sequence comparison methods (like BLAST) do reasonably well at identifying the superfamily level of function, like peroxiredoxin.
Notably, no single threshold can be identified to distinctly identify the six known subgroups, an illustration of why full sequence comparison methods are less successful at identifying detailed levels of molecular function, such as distinguishing between Prx1 and Prx6. At less stringent edge (BLAST score) thresholds (Fig 11B), some subgroups are indecipherable from one another (such as Prx1 to Prx6 and AhpE to PrxQ), and at more stringent thresholds (Fig 11D), some subgroups begin to split unnecessarily (such as PrxQ, AhpE, and Prx5). Unsurprisingly, the Prx1 and Prx6 subgroups are difficult to distinguish from one another until the most stringent threshold. Previous work has demonstrated that the similarity between these 2 subgroups makes it difficult to separate them based on sequence comparison alone [33]. MISST focuses on active site features to define isofunctional groups, thus eliminating reliance on full sequence comparison for detailed molecular function analysis.
In this work, active site features are utilized to define functionally relevant clusters. A method, MISST, which uses self-identification of clusters to define functional relevance is introduced. The method is both agglomerative and divisive. As ASPs become more robust, DASP2 searches agglomerate more functionally related sequences. Likewise, at each stage, clusters are evaluated for the presence of groups that exhibit distinct functional site features. Functionally relevant clustering of the Prx superfamily presents several challenges for the method: How are sequences agglomerated (Tpx and Prx5)? How are clusters subdivided when they contain two distinct isofunctional groups (Prx1 and Prx6)? And, how are functionally relevant groups identified when structural representation is extremely limited (AhpE)?
A defining feature of functional annotation is the hierarchy under which groups of proteins are classified [40,41,54], and it is important to understand how the MISST results fit into a functional hierarchy. Members of the Prx functional superfamily all perform a similar redox chemistry at CP; differences lie in substrate recognition and details of how CP is regenerated. The six expert-annotated groups of Prxs are classified as subgroups in the SFLD, which indicates that group members share more features among themselves than with members of other Prx subgroups.
MISST distinguishes these six subgroups, and members thereof, identifying the differences between the mechanisms from which hypothesis-driven experiments can be developed. As expected, many new sequences were identified—the Prx data in SFLD is from 2010 and 2016 GenBank is significantly larger. Over 99% of newly identified sequences contain the canonical Prx active site motif. Additionally, with the exception of the AhpE subgroup, the phylogenetic distribution for each subgroup is reasonably consistent with the original Prx data, as recently reported by Poole and Nelson [58].
The current work demonstrates the feasibility of this novel, agglomerative approach of using self-identification to identify isofunctional clusters. Notably, the MISST process does not require human- or expert-based analysis and is automatable, with the exception of identification of the key functional residues which are input to TuLIP. Two MISST parameters may require further adjustment to demonstrate generalizability: score thresholds and self-identification criteria. However, our work on the enolase and Prx superfamilies suggests the score thresholds are generalizable (S3 File). The feasibility of MISST is demonstrated here on the Prx superfamily. More extensive parameterization, validation, and generalizability will be demonstrated once the method is automated. Ultimately, we envision that MISST could be applied to cluster any protein superfamily automatically, thus laying the foundation for functionally relevant clustering of the universe of protein sequences.
The peroxiredoxin (Prx) superfamily contains six subgroups previously identified by expert analysis: Prx1 (formerly AhpC/Prx1), AhpE, PrxQ (formerly BCP/PrxQ), Prx5, Prx6, and Tpx [28,32,34,35]. These expertly-identified subgroups are available in PREX [39]. Curators at the Structure-Function Linkage Database (SFLD) constructed hidden Markov models (HMMs) for each subgroup and have updated the proteins in each subgroup in SFLD (Prx superfamily, EC 1.11.1.15) [40,41]. As of March 7, 2016, there were 7,267 annotated Prx proteins (unique EFDIDs) in SFLD, distributed among the subgroups as follows: 2,225 Prx1, 112 AhpE, 1,939 PrxQ, 1,062 Prx5, 975 Prx6, and 954 Tpx. Additionally, there were 4,695 proteins assigned to the Prx superfamily but not assigned to a subgroup (uncharacterized) in the SFLD.
Active site profiling is a method used to identify the residues in the structural vicinity of a protein’s functional site (Fig 1) [36]. Briefly, key residues important for catalytic activity (Fig 1, black residues) are identified using a combination of the Catalytic Site Atlas (CSA) [64] or literature research and structure alignment. All residues within 10 Å of each key residue (Fig 1A, gray spheres) are identified and extracted from the full protein sequence and aligned N- to C- terminus to create an active site signature (Fig 1B). Fragments containing three residues or fewer are removed from the active site signature as they lack sufficient length for statistical significance. Multiple signatures are aligned to create an active site profile (ASP), characterizing the active site features of all proteins in the group (Fig 1B). An ASP score is calculated indicating the residue variation among the signatures in the profile [36]. ASP scores range from -0.5 to 1.0, where 1.0 indicates perfect alignment and conservation across all signatures.
The Deacon Active Site Profiler (DASP) is a tool that uses ASPs to search sequence databases for proteins with fragments similar to the active site motifs [37,38,55]. The ASP is separated into aligned motifs which contain contiguous fragments within the signatures (Fig 1, colored fragments). For each aligned motif, a position specific scoring matrix (PSSM) [65] is calculated, detailing the propensity for specific residues to appear in each position of the motif, normalized to the background frequency of each residue in the database [35,37,55]. Starting with the longest motif, a sliding window search is performed along each sequence in the database. A p-value defining the similarity between the ASP motif and the sequence fragment is calculated for every position; the most significant p-value indicates the best match between a fragment and the motif for a given protein. All motifs are searched in this manner to identify the best matching fragment with the caveat that fragment matches cannot overlap. For each protein sequence, the p-values for each “best matched” fragment are combined using QFAST [66] to calculate a DASP search score. This score represents the probability a given sequence contains the fragments matching the ASP motifs by chance. This process is completed across all protein sequences in the database, such that each protein is associated with a DASP search score indicating the statistical significance of the match between the protein fragments and the ASP fragments.
To efficiently perform iterative DASP searches, a new version of DASP named DASP2 was developed to support variable input formats and decrease GenBank search times. DASP2 testing demonstrated DASP and DASP2 return essentially identical data, but DASP2 searches are significantly more efficient [53]. Additionally, expanding the supported input formats allows the identified fragments of one search to be used as the input of the next search, opening the door for iterative database searches used in the MISST process. While these changes do not alter the search results, the latest version supports efficient, iterative GenBank searches which are critical to this work.
Previously, Leuthaeuser and coworkers demonstrated that clusters identified using pairwise active site similarity networks often share more functional details than those identified using full sequence or full structure similarity networks [54]. Building on this, the Two Level Iterative clustering Process (TuLIP) was developed to identify functionally relevant groups of protein structures based on active site similarity. Validation was previously performed on the enolase and GST superfamilies. Results demonstrated significant correspondence to known functional groups [52].
Initially, an all-by-all network was created using the 47 non-redundant Prx structures in which each node represents one protein structure and each node pair is connected by an edge representing a pairwise ASP score. The edge threshold was incremented and the MCL clustering algorithm [67] was applied until distinct subnetworks form, such that no edges connect subnetworks to each other. At this point, an ASP is created for each subnetwork and used to search the PDB with DASP2. If the PDB search using the subnetwork’s ASP identifies only itself (the proteins within the subnetwork) at significant DASP search scores, it is defined as “functionally relevant” and removed for further analysis. For all subnetworks which are not identified as functionally relevant groups, the edge score is incremented and the process repeated. This iterative clustering process is continued until each protein is either part of a functionally relevant group or separated out as a singlet, which signifies the end of the strict clustering stage.
The full iterative approach is then repeated for the relaxed stage: a fully connected network is formed from all singlets and the edge threshold is incremented to form subnetworks which are used to search the PDB for identifying functionally relevant groups. The relaxed stage uses more relaxed parameters for evaluation of the functional relevance of each subnetwork. Again, any subnetwork that meets the relaxed parameters is defined functionally relevant and is removed. The edge threshold is then incrementally increased. Once all proteins are either members of a functionally relevant group or singlets, TuLIP is complete. Utilizing two stages of iteration allows identification of functionally relevant groups whose relationship might be obscured by the more coherent groups identified with the strict clustering parameters.
TuLIP is performed only on proteins of known structure. A single DASP2 search can expand the group into sequence space; however, the identified sequences are limited by the diversity of the search ASP, which, in turn, is limited to those sequences represented in the structure database. To expand functionally relevant clustering, so that the diversity of sequences and functionally relevant groups are fully comprehended, the Multi-level Iterative Sequence Searching Technique (MISST) was developed (Fig 4). This process utilizes iterative DASP2 GenBank searches to populate each TuLIP group with sequences sharing active site similarity, thus increasing robustness of the search ASP. Additionally, a novel PSSM Analysis method identifies when and how a MISST group should be subdivided into distinct functionally relevant groups.
To initiate MISST, an ASP is created for each TuLIP group; each ASP is used in an initial DASP2 search, Search0, of GenBank (Fig 4A). Given the limited representation in the structure database, the active site diversity of these initial ASPs is limited; thus, the goal of Search0 is to create a more robust ASP better representing each group’s functional site diversity. A DASP2 score of ≤1e-12 was chosen as the threshold for inclusion of sequences in the more robust profile. Previous work had identified ≤1e-8 or ≤1e-10 as “generous” and “trusted” DASP score thresholds in a single search of Prx subgroups [35]. Subsequent work on the enolase superfamily demonstrated that cross-hits (sequences identified as members of more than one functional group) decreased to zero at ≤1e-13 in the 26 subgroups and families of the enolase superfamily [52]. Balancing performance, precision and recall on the enolase superfamily, a “trusted” score threshold of ≤1e-12 was identified and is therefore used here. A detailed analysis and discussion of these score thresholds is provided in Supplemental Methods in S3 File.
An ASP is created from the pseudo-signatures of sequences identified with DASP2 search scores more significant than the score threshold. To create each pseudo-signature, fragments identified in each sequence as matching each ASP motif are concatenated (in length order, longest to shortest). The pseudo-signatures are aligned to create a new ASP for each group; each ASP is then used as input into a second DASP2 search of GenBank, termed Search1 (Fig 4A). At this point, an iterative process of sequence acquisition and data analysis begins for each TuLIP group. The DASP2 score threshold for Search1 and beyond is ≤1e-14, rather than ≤1e-12 used at Search0. 1e-14 was determined to be a more appropriate threshold because the ASPs become more robust and the DASP2 scores of known true positives shift to more significant scores with the addition of new sequences at each search (S3 File, S6 Fig). Notably, there is no score shift after Search 1 as the average DASP search score for true positives does not improve between Search 1 and Search 2 or beyond (S3 File, S6 Fig).
Beginning with Search1, each group is analyzed against two self-identification criteria to determine if the group is self-contained and stable (Fig 4A). This approach to identifying functionally relevant groups is novel as groups are not identified based on a specific threshold, but instead all groups are required to pass a self-identification test to be considered functionally relevant. In this way, groups which are functionally distinct and easier to identify can be fully identified in few iterations, while groups sharing similar active site features with other groups may take more iterations to be distinctly identified. This approach prevents the simultaneous subdivision of some groups and combination of other groups that is prevalent in most clustering.
A group is complete and removed from the iterative process when a GenBank search demonstrates self-identification; that is, all inputs are identified with significant DASP2 search scores and nothing else is identified with significant DASP2 search scores, within a small range of error. Quantitatively, two metrics define the self-identification criteria: percent new hits and percent inputs hit. The first metric tracks whether the search identified sequences not identified in the previous search: if ≤15% of the sequences identified at a score threshold of ≤1e-14 are “new” (not identified ≤1e-14 in the previous search), the group meets this metric. The second metric evaluates whether the proteins used as input were identified in this search. To pass, ≥95% of input proteins must be identified at a DASP2 score threshold of ≤1e-14 (see Supplemental Methods in S3 File for more detail). A MISST group is removed from the iterative process if it meets both metrics (Fig 4A).
The values of these two parameters were chosen based on data from the Prx superfamily, but will be evaluated on other superfamilies in the future. For completed groups in the current data set, percent new hits averaged 5.4% with a range from 2.2% to 11.3% and incomplete groups averaged 63.4% with a range from 29.9% to 98.2%. Similarly, percent inputs hit averaged 99.7% with a range from 99.5% to 100% for complete groups and averaged 66.8% with a range from 50.9% to 99.3% for incomplete groups. Preliminary analysis with other SFLD superfamilies (enolase, crotonase, and radical SAM) suggests these parameters are relatively generalizable, but comprehensive testing is required on more data sets.
Once all groups meet the self-identification criteria, a final ASP is constructed from each MISST-identified group and used to search GenBank one additional time to obtain the final MISST search results for that superfamily. The ASPs of completed searches can additionally be used at any future time to identify new sequences recently added to GenBank.
At each iteration, all groups that do not pass self-identification criteria are evaluated using the following protocol (Fig 4A, gray box):
These three steps are completed for each MISST group at each search iteration (Fig 4A). After completion of these three steps, PSSM Analysis (Fig 4C; see subsequent section) is performed to determine potential group subdivision.
Position Specific Scoring Matrix (PSSM) Analysis is a novel approach using PSSMs to determine whether a group of protein sequences contains more than one identifiable functionally distinct group based on residue similarity within the active site signatures. In this way, MISST groups that contain multiple functionally-distinguishable families can be appropriately subdivided.
PSSM Analysis begins by placing every protein identified by one group’s search into order of magnitude “bins” based on the DASP2 search score at which they were identified. Each order of magnitude is considered a bin, such that proteins with DASP search scores >1e-9 and ≤1e-8 are placed into the bin labeled “8” (Fig 4Ci). All proteins with DASP search scores ≤1e-25 are placed into the bin labeled “25.” Bin-specific ASPs are created from the proteins in each bin (using the pseudo-signatures described previously) and a PSSM [65] is calculated for the each ASP, resulting in 18 bin-specific PSSMs (Fig 4Ci). The PSSM values are based on the count of each residue in each position of the profile, normalized to the overall count of that residue in the database.
To identify the similarity between proteins in each pair of bins, a modified Pearson correlation coefficient is calculated pairwise between bin-specific PSSMs. A PSSM is a two-dimensional array, the first dimension representing each of the 20 amino acids; the second dimension representing a position in an ASP (positions in an ASP are indicated by arrows in Fig 1B). The standard Pearson correlation coefficient is calculated between analogous columns of a pair of PSSMs. To get the overall comparison between two PSSMs, column correlations must be summarized, but averaging correlation coefficients can lead to bias [70]. Therefore, a Fisher transformation is executed prior to computing the average. Due to the nature of the transformation, all coefficients >0.9999 are set equal to 0.9999, and the Fisher transform is performed to produce a z-score for each column. The z-scores are then averaged across all columns and back transformed to r, producing the modified Pearson correlation coefficient, which correlates the active site similarity between the proteins in two bins.
To define when a group should be subdivided, a fully connected network is created, with each node representing proteins in a scoring bin (from 8 to 25) and each edge representing the pairwise correlation coefficient between bin-specific PSSMs (Fig 4Cii). Beginning at the lowest correlation value (rounded to two decimal places), a filter threshold is applied to the network, removing all edges below the threshold. The filtered network is clustered using MCL clustering [67] to produce subnetworks (Fig 4Ciii). If distinct subnetworks are formed, in which no edges connect the two (or more) subnetworks to each other, the subnetworks are evaluated based on the following criteria to determine if they might represent functionally distinct groups: 1) each subnetwork must contain at least three nodes; and 2) the nodes (bins) must represent contiguous DASP2 scores (e.g. 8, 9, and 10 rather than 8, 10, and 12). If the subnetworks meet both criteria, the subnetwork containing the nodes with the least significant DASP2 scores is removed as a potential functionally relevant group, while the remaining subnetwork is subdivided further. If a subnetwork does not meet both criteria, it is not identified as a potential functionally distinct cluster. The filter threshold is increased by 0.02 each iteration and the clustering process is repeated.
At the edge threshold of 0.98, PSSM Analysis is completed. If a group has subdivided, ASPs are built from the pseudo-signatures of proteins in each subnetwork and used in the subsequent MISST iteration and search of GenBank (Fig 4A). If the network reaches the 0.98 edge threshold and no subnetworks have been identified, an ASP is created from the pseudo-signatures of the sequences with DASP2 search scores ≤1e-14.
MISST iterations continue, as outlined in Fig 4. Once all groups pass self-identification criteria, a final DASP2 search of GenBank is completed for each MISST-identified group. In this work, these final searches were completed in March 2016. Cross hit analysis then identifies the number of shared sequences between the six groups identified at the significance threshold of ≤1e-14. Cross-hits are identified and removed using the same procedure utilized during the MISST process (Fig 4A). The final list of all proteins identified in each MISST group along with their DASP2 search score, SFLD annotation, and pseudo-signature can be found in S2 File.
The results of these searches were compared to the expert-identified subgroups using quantitative methods previously used to evaluate other similar processes [49,50,59,62]. To calculate these measures, the MISST groups were compared to the sequences in the SFLD as of March 6th, 2016 (http://sfld.rbvi.ucsf.edu/django/). Each of the 6 MISST groups contained the majority of one subgroup; consequently, the analysis was completed using a 1-to-1 correspondence of MISST group to known functional groups (defined in Table 1).
To evaluate how well our clusters compared with known functional clusters, measurements of purity, edit distance, and VI distance were performed, as previously described [50]. Additionally, the combined performance metric suggested by Orengo and colleagues [49] was calculated as well as the F-measure, which is the harmonic mean of precision and recall [59]. Details of these metrics are provided in S3 File.
The consensus Prx motifs for each group were determined based on the conservation of residues in each position of the motif according to the following rules: 1) if the three most conserved residues make up ≤97% of that position, an x is used in the consensus sequence for that position, and 2) for all other positions, all residues identified in ≥3% of the MISST group sequences are annotated in the consensus sequence. Conservation graphs were built using Weblogo [61].
A representative network (RepNet) was created for all 38,739 sequences identified by the six MISST groups in the final searches using Cytoscape [63]. Using CD-Hit [68,69], 1,369 clusters were identified where all members share 55% sequence identity with the representative protein. Each representative is a node in the RepNet and the edges connecting the nodes are pairwise BLAST scores between each pair of representatives. The nodes are colored by the MISST group the proteins were identified by.
|
10.1371/journal.pntd.0000726 | Pulmonary Abnormalities in Mice with Paracoccidioidomycosis: A Sequential Study Comparing High Resolution Computed Tomography and Pathologic Findings | Human paracoccidioidomycosis (PCM) is an endemic fungal disease of pulmonary origin. Follow-up of pulmonary lesions by image studies in an experimental model of PCM has not been previously attempted. This study focuses on defining patterns, topography and intensity of lung lesions in experimentally infected PCM mice by means of a comparative analysis between High Resolution Computed Tomography (HRCT) and histopathologic parameters.
Male BALB/c mice were intranasally inoculated with 3×106 Paracoccidioides brasiliensis (Pb) conidia (n = 50) or PBS (n = 50). HRCT was done every four weeks to determine pulmonary lesions, quantify lung density, reconstruct and quantify lung air structure. Lungs were also analyzed by histopathology and histomorphometry.
Three different patterns of lesions were evidenced by HRCT and histopathology, as follows: nodular-diffuse, confluent and pseudo-tumoral. The lesions were mainly located around the hilus and affected more frequently the left lung. At the 4th week post-challenge HRCT showed that 80% of the Pb-infected mice had peri-bronchial consolidations associated with a significant increase in upper lung density when compared with controls, (−263±25 vs. −422±10 HU, p<0.001). After the 8th and 12th weeks, consolidation had progressed involving also the middle regions. Histopathology revealed that consolidation as assessed by HRCT was equivalent histologically to a confluent granulomatous reaction, while nodules corresponded to individual compact granulomas. At the 16th week of infection, confluent granulomas formed pseudotumoral masses that obstructed large bronchi. Discrete focal fibrosis was visible gradually around granulomas, but this finding was only evident by histopathology.
This study demonstrated that conventional HRCT is a useful tool for evaluation and quantification of pulmonary damage occurring in experimental mouse PCM. The experimental design used decreases the need to sacrifice a large number of animals, and serves to monitor treatment efficacy by means of a more rational approach to the study of human lung disease.
| Paracoccidioidomycosis (PCM) is a fungal infection caused by the dimorphic fungus Paracoccidioides brasiliensis. It occurs preferentially in rural workers in whom the disease is severe and may cause incapacitating pulmonary sequelae. Assessment of disease progression and treatment outcome normally includes chest x-rays or CT studies. Existing experimental PCM models have focused on several aspects, but none has done a radiologic or image follow-up evaluation of pulmonary lesions considered as the fungus primary target. In this study, the lungs of mice infected with fungal conidia were studied sequentially during the chronic stage of their experimental mycosis by noninvasive high resolution medical computed tomography, and at time of sacrifice, also by histopathology to characterize pulmonary abnormalities. Three basic lung lesion patterns were revealed by both techniques: nodular-diffuse, confluent and pseudo-tumoral which were located mainly around the hilus thus accurately reflecting the situation in human patients. The experimental design of this study decreases the need to sacrifice a large number of animals, and serves to monitor treatment efficacy by means of a more rational approach to the study of human pulmonary diseases. The findings we are reporting open new avenues for experimental research, increase our understanding of the mycosis pathogenesis and consequently have repercussions in patients' care.
| Human paracoccidioidomycosis (PCM) is an endemic fungal infection of pulmonary origin that disseminates to different sites, notably oral mucous membranes, skin, adrenal glands and reticuloendothelial system. The disease tends to run a chronic progressive course while acute cases are more unusual. This mycosis is caused by Paracoccidioides brasiliensis, a thermally dimorphic fungus [1].
Primary infection in humans occurs in the lungs, where it causes chronic granulomatous inflammation of the parenchyma leading to fibrosis and severe restriction of respiratory function [2].
We have developed a model of pulmonary PCM in male BALB/c mice induced by the intranasal inoculation of P. brasiliensis conidia [3]. This model allowed to evaluate histopathologically and immunologically the pulmonary tissue responses occurring during the active and residual stages of the processes [4], [5], [6]; however, radiological follow-up evaluation of pulmonary lesions in the experimental model of PCM have not been previously described.
Non-invasive radiological imaging has recently gained considerable interest in basic and preclinical research for monitoring disease progression and assessing therapeutic efficacy [7]. One of the most noteworthy attributes of non-invasive imaging is the ability to obtain data from individual animals at multiple time points. Therefore, the number of animals required for a study can be minimized [8]. Additionally, each pixel in the image has a value that can be mapped to the density of the tissue being imaged [9]. Furthermore, neither conventional histological analysis nor pulmonary function test provide information, in alive animals, on the three-dimensional (3-D) distribution of lesions over the entire lung volume [10].
Recent improvements in spatial resolution capacity have made possible to manufacture scanners specifically designed for imaging small animals (microscopic computed tomography - “micro-CT”), which produces images with spatial resolution of 50–100 µm [11]. However, the radiation dose delivered to the animal during micro-CT imaging may approach 5% of the median lethal dose in mice (LD50), potentially limiting the number of repeated studies that could be performed over time [12]. Furthermore, radiation exposure from repeated micro-CT scans may have an effect on skeletal growth in normal animals [13].
Some studies reported clinical scanners designed primarily for human application but employed, nonetheless, to follow-up lung fibrosis and tumor progression in mice [14], [15], [16]. However, only few studies have described and followed-up experimental mycoses using computed tomography (CT) [17] and no studies have been published using this radiological tool in experimental pulmonary PCM.
Recognition and monitoring of CT patterns associated with this model of disease could improve our understanding of anatomo-spatial distribution of lesions and their time course in the same animal. Also, it appears possible to identify differences between manifestations in human and experimental PCM models and, finally, imaging could also be used to evaluate in vivo therapeutic responses.
This study focus on a comparative analysis between high-resolution computed tomography (HRCT) and histopathological parameters, determining usefulness of performing noninvasive conventional medical X-ray tomography in the follow up of sequential lung lesions in the experimental PCM model induced in mice by conidial inoculation.
All animals were handled according to the national and international guidelines for animal research and experimental protocols were approved by Corporación para Investigaciones Biológicas (CIB) research ethics committee.
BALB/c mice were originally obtained from Taconic Farms, Inc. Quality Laboratory Animals and Services for Research, New York, USA with the breeding colony being then expanded at the Corporación para Investigaciones Biológicas (CIB), Medellin, Colombia. Male mice, 7 weeks old and approximately 20 g in weight were used in this study. Mice were divided into 2 groups: non-infected control mice (n = 50) and P. brasiliensis infected mice (n = 50).
A P. brasiliensis strain registered at the American Type Culture Collection (Rockville, MD), ATCC-60855, was used in all experiments. This strain was originally isolated from a Colombian patient and it is known to produce abundant conidia (natural infectious propagules) [18], and cause a progressive chronic disease [19]. The strain was previously passed though mice to restore virulence and then used for production of conidia. The fungus was maintained at 18°C in its mycelial form by successive transfers on the modified synthetic McVeigh and Morton (SMVM) medium [20]. The growth was then transferred to an Erlenmeyer flask with liquid SMVM and incubated for 10–15 days (18°C) with constant shaking at 150 rpm (Model G-2 gyratory shaker, New Brunswick Scientific, Co. New Brunswick, N.J.). After this time, growth was collected, homogenized in a blender (Eberbach container assembly semi-micro press with fit cover) for 15–20 seconds in four intervals of four seconds each, and plated in Petri dishes containing a media that stimulates conidia production, namely, water agar medium and dextrose salts agar [21]. Culture dishes were washed with 0.85% saline solution plus 0.01% Tween-20; this suspension was then shaken at 250 rpm for 45 min at 18°C in an Erlenmeyer flask containing glass beads. The homogenized suspension was sonicated twice for 15 seconds at 7 Hz at 4°C with one minute intervals (Sonicator model 200, Branson Ultrasonic Co, Danbury,CT). The fungal slurry was then poured through a sterile syringe packed with glass wool. The conidia suspension that passed through the glass wool was concentrated by centrifugation. The number and viability of the conidia were determined by the fluorescein diacetate-ethidium bromide fluorescence method [22]. The viability of the conidia was consistently higher than 90% of the total number of conidia counted. The inoculum was then adjusted so that 0.06 ml contained approximately 3×106 viable conidia [4].
Mice were anesthetized by the intramuscular injection of a solution containing Ketamin hydrochloride (Park, Davis & Company, Berlin, Germany; 100 mg/kg) and Xylazine (Bayer, Brazil 10 mg/kg) [23]. When deep anesthesia was obtained, 3×106 conidia (in 0.06 ml of the inoculum) were instilled intranasally. Control mice received an intranasal inoculum of 0.06 ml of saline [4].
Fifty animals/group were scanned at 0, 4, 8, 12 or 16th weeks post-inoculation (10 mice/time after inoculation). Mice were anesthetized with ketamin hydrochloride (Park, Davis & Company, Berlin, Germany; 100 mg/kg) and Xylazine (Bayer, Brazil; 10 mg/kg) and placed in prone position inside polypropylene tubes (50 ml), which were arranged together in a wood box with parallel holes (Figure 1 A, B). All the animals were placed with their noses in the same vertical plane. Each animal had a code for future identifications. The box containing the mice was then placed in the CT gantry for thorax scanning (Figure 1 C, D).
CT images were taken in a multislice CT-scanner Lightspeed, (General Electric, EU of 16 canals) applying 140 kV with 165 mAs/second (Kernel U90). Thin-section slices, each of 0.625 mm in thickness, and spaced 1 mm apart covered the complete mouse lung from the apex to the hemidiaphragm. Images were acquired in axial plane and the bone algorithm was applied to better visualize the lung. Field of view was 18 cm to include simultaneously all the animals, with matrix of 512×512 and acquisition time of one second per section. Around 20 to 22 slices covered the entire animal lungs.
Following the scan, the mice were placed in their cages, recovering from the anesthesia within 35 minutes. Animals were supplied with standard laboratory diet and water ad libitum.
Image analysis was performed independently and blindly by two radiologists from the Radiology Department of the University Hospital San Vicente de Paul (Medellín, Colombia). Images were visualized using Advantage workstation version 4.3 General Electric, applying lung and mediastinal windows.
The pulmonary densities were evaluated as described by Plathow (2004) [14], with some modifications. Representative tomographic slides were used to figure out Hounsfield units (HU). Briefly, eight regions of interest (ROI) were selected in the following areas of the right and left lungs: upper or hilar region (about 5 slides below the apex, where the main bronchi enter the lungs), anterior and posterior middle or central region (about 12 slides bellow the apex where the heart presents its larger diameter) and finally, lower lung region (about 18 slides bellow the apex, corresponding to bases of lung). These circles were of 2 mm2 for upper and middle regions and of 4 mm2 for the lower regions. Main bronchi and vessels were omitted to measure parenchyma density (Figure 2, A–D). However, to evaluate the extension of the inflammatory infiltrate that presented a predominantly axial location, the ROIs were enlarged to include the interstitium around pulmonary arteries and bronchi.
After scanning, ten animals per group were euthanized by Thiopental overdose (Sandoz GmbH. Kundl Austria; 1 ml at 2.5%, i.p) at 0, 4, 8 12 and 16th weeks post-inoculation in accordance to animal ethical practice. Lungs of five mice were intracardially perfused with 10% formalin neutralized with phosphate-buffered saline, removed and fixed in the same solution by at least 48 h. Lungs of remainder mice were used for other studies.
Formalin-fixed lungs were embedded in paraffin and coronal sections (5 µm) were stained with hematoxylin-eosin (HE), and picrosirius with fast green (PIFG) [24] to evaluate the inflammatory reaction and determine collagen deposition, respectively. Slides stained by PFIG were automatically scanned by ScanScope® CS (Aperio, USA). The extent of tissue involvement was estimated by histomorphometry, as described below.
Morphometric analysis was done using one panoramic image of both lungs per mouse captured with digital camera (Axiocam MRc5, Carl Zeiss, Germany) adapted to a stereomicroscope (Stemi Sv11, Carl Zeiss, Germany). The images were analyzed by the free Image J software (http://rsbweb.nih.gov/ij/, NIH, USA). Areas of interest (AOIs), correspondent to the inflammatory regions, were manually drawn and measured. The percentage of pulmonary area with inflammatory reaction was calculated dividing the sum of total AOIs by the total area occupied with lung tissue (excluding the air space).
The statistical analyses between groups were performed with Prism 5.0 software (Graph Pad, USA) applying one-way or two-way ANOVA. Values were expressed as mean ± standard error of the mean. p values less than 0.05 were considered statistically significant; p values less than 0.01 were considered statistically highly significant. The frequency of infected mice with increased lung density was determined considering the outlier values established by boxplot graph, using 1.5 times of the interquartile range.
Lung density in healthy BALB/c mice showed local differences according to the region evaluated. The upper or hilar lung density was higher (−432.7±10.82) than the central and lower regions, −492.4±6.5 and −492.7±4.8, respectively, during all observation times, indicating that these latter regions were more aerated than the former; p≤0.001, (Figure 2E).
At week 4 post-infection, Pb-infected mice showed peri-bronchial consolidations that persisted in every one of the evaluation periods. Pulmonary consolidations was associated mainly with a significant increase in upper lung density as compared with controls, −263±29 vs. −426±8 HU at week 4 (p<0.001), −191±25 vs. −403±17 HU, at week 8 (p<0.001), −269±43 vs. −445±12, at week 12 (p<0.001). At week 16, upper consolidations tended to decrease as well as the corresponding density, −356±33 vs. −466±9 at week 16 (p>0.01) (Figure 3C). At weeks 8 and 12, consolidation had progressed to involve also the middle regions with a statistically significant increase in density (Figure 3G, H). The lesions, during every one of the infection times, were mainly located in the hilar region (Figure 3D, H). Lesions were not detected in the lung bases, which presented normal density (Figure 3 I-L).
Left lung was more frequently affected by lesions than the other pulmonary regions (>80% the mice at week 4, 8 and 12), followed, in frequency, by upper right and central or middle lung regions, saving the bases (Figure 4).
The main patterns of lesions were nodular-diffuse, confluent and pseudo-tumoral with some occasional and additional aspects such as atelectasis. These aspects were revealed by both HRCT and histopathology, showing a strong correspondence between the two approaches (Figure 5).
The three-dimensional reconstruction of air-structure showed that the more consolidated zones, predominantly in the left upper lung, provoked a deprivation of the air volume in the correspondent region (Figure 6B arrow). Nonetheless, the quantification of the total lung volume did not show statistically significant difference between control and infected animals at any time of evaluation (Figure 6C).
The cardio-thoracic ratio, vascular structures and main bronchi did not show significant changes at any time during evaluation and no mouse showed a radiologic pattern compatible with lung fibrosis, a common sequela observed in human PCM. Pleural changes were observed in only one mouse.
The granulomatous reaction was initially seen around bronchi, bronchioli and blood vessels, where various compact nodule-shaped granulomas were detected (Figure 5B). Some of them became confluent and contained numerous fungal cells. Additionally, large areas of the lung became consolidated involving hilar regions and lung parenchyma (Figure 5D). After 12th week of infection, confluent granulomas formed pseudotumoral masses (Figure 5F) that obstructed large bronchi. Complete atelectasis of the left lung was observed in one animal (Figure 5H).
The percentage of lung area occupied by inflammatory reaction was 8±2% at 4th week, and gradually increased to 20±6% at 12th week. Finally, at 16th week, the percentage of affected lung area decreased to 13±3% (p<0.01 at all evaluation times) (Figure 7). The extension of the lesions presented a direct correlation with lung density in the upper (Figure 8A) and central region (Figure 8B), while an inverse correlation was observed in the lower lung region (Figure 8C).
By histopathology, the presence of fibrosis was ascertained; it was observed predominantly in the inflammatory reaction around arteries and in the periphery of granulomas, with less intensity in the peribronchial connective tissue and without evidence in the interstitium.
Through a comparative analysis between HRCT and histopathological data, this work revealed that noninvasive conventional medical X-ray tomography is adequate to follow the sequential lung lesions in experimental PCM in mice. This procedure allowed detection of the main pathological patterns, the differential topographic distribution of the pulmonary lesions in both lungs, and their intensity in our experimental model of PCM. Three basic lesion patterns were evidenced by the study: nodular-diffuse, confluent and pseudo-tumoral (Figure 5). Histopathologically, the lesions were predominantly of the granulomatous type and were mainly located around branches of the arterial vasculature and close to the bronchial tree, preserving large areas of the parenchyma. Concerning the topographic distribution, the hilar region of the upper left lung was more frequently involved than other regions (Figures 3 and 4), and while the mechanism of this preference is unknown, it could be influenced by lymphatic drainage. It was surprising to notice the predominance of left lung involvement, considering that the bronchial mouse anatomy reveals a thinner main left bronchus in comparison with the right one and with approximately the same angular deviation from the carina (Figure 6). We do not know the effect of the ventral decubitus position on the differential inclination of the main bronchi.
The intensity of the inflammatory reaction, evaluated by histomorphometry, was crescent until the 12th week of infection with subsequent decrease due to the tendency to form predominantly compact and more isolated pseudotumoral masses (Figure 7). This histopahological behavior was also detected by HRCT, as expressed by lung density measures (Figure 3) that showed a significant correlation mainly in the upper or hilar lung region (Figure 8A).
The absence of difference in the total lung air space volume between the control and infected groups (Figure 6) suggested that the healthy or less affected lung compensated, by hyper-insufflation, the focal volume lost. This effect was also supported by the indirect correlation observed between the percentage of lung area with inflammatory reaction and the lung density observed in the lower region (Figure 8C).
For future studies, it would be of interest to assess pulmonary function in our experimental model of PCM by non-invasive methods such as unrestrained whole-body plethysmography for small animals, or other more accurate procedures for determining physiological parameters although invasive techniques, such as forced pulmonary maneuvers system or forced oscillation techniques [25]. Pulmonary function analysis would respond to the following question: Do the large masses or the consolidated lesions observed in our study by HRCT or the fibrosis recorded by histopathology, decrease the normal function of lungs of mice with experimental PCM?
Although the human and mouse lungs exhibit basic anatomic similarities, they present significant interspecies differences, such as: the absence of respiratory bronchioles in the mouse; number of the subdivisions of the conducting airways; characteristics of pleura structure, interlobular septa, pulmonary and bronchial vasculature, bronchial associated lymphoid-tissue and others that could contribute to explain the different behaviours of fibrogenesis [26], [27], [28].
Radiologic patterns showed by HRCT in the experimental infection, differed slightly from their counterpart in human patients. The most frequent HRCT findings in patients with pulmonary PCM are: ground-glass attenuation areas, small centrilobular, cavitated and large nodules, parenchymal bands, airspace consolidations, interlobular septal thickening, architectural distortion, traction bronchiectasis, paracicatritial emphysema and fibrosis. Most of those HRCT findings predominate in the periphery and posterior regions involving all lung zones, with light predominance in the middle zones [29], [30], [31], [32]. The radiologic patterns described above in patients with pulmonary PCM were dependent on the stages of the disease and the exclusion or not of patients who had received previous treatment.
Values of lung density of healthy mice differed from those of the human counterpart. In BALB/c mice the apices were denser than the bases suggesting that there was less air in the former regions (Figure 2). On the contrary, human lung apices are more ventilated and less perfused than bases which suppose a decrescent button to up gradient of density probably due to human upright position, which considerably reduced the flow of low pressure pulmonary artery blood in the upper lung for long periods. Otherwise, ventilation to the various parts of the lung is much less affected by body position. As a consequence, oxygen uptake and carbon dioxide excretion are impaired in the upper as compared with the lower parts of the lungs in the erect position. This situation produces a more oxygenated environment in the upper lungs, which favor proliferation of some pathogenic agents like Mycobacterium tuberculosis [33]. This kind of mechanism appears not to interfere on paracoccidioidomycosis infection in human and experimental animals.
In conclusion, this study demonstrated for the first time that conventional-HRCT is a useful, precise and non-invasive technique to evaluate and quantify the pulmonary damage occurring in the mouse experimental paracoccidioidomycosis. This procedure will contribute significantly to decreases the need of killing large number of animals, and to monitor treatment efficacy in animal models with an approach that reflecting the way human pulmonary diseases are studied.
|
10.1371/journal.ppat.1002047 | Phospholipids Trigger Cryptococcus neoformans
Capsular Enlargement during Interactions with Amoebae and
Macrophages | A remarkable aspect of the interaction of Cryptococcus
neoformans with mammalian hosts is a consistent increase in capsule
volume. Given that many aspects of the interaction of C.
neoformans with macrophages are also observed with amoebae, we
hypothesized that the capsule enlargement phenomenon also had a protozoan
parallel. Incubation of C. neoformans with Acanthamoeba
castellanii resulted in C. neoformans capsular
enlargement. The phenomenon required contact between fungal and protozoan cells
but did not require amoeba viability. Analysis of amoebae extracts showed that
the likely stimuli for capsule enlargement were protozoan polar lipids. Extracts
from macrophages and mammalian serum also triggered cryptococcal capsular
enlargement. C. neoformans capsule enlargement required
expression of fungal phospholipase B, but not phospholipase C. Purified
phospholipids, in particular, phosphatidylcholine, and derived molecules
triggered capsular enlargement with the subsequent formation of giant cells.
These results implicate phospholipids as a trigger for both C.
neoformans capsule enlargement in vivo and
exopolysaccharide production. The observation that the incubation of C.
neoformans with phospholipids led to the formation of giant cells
provides the means to generate these enigmatic cells in vitro.
Protozoan- or mammalian-derived polar lipids could represent a danger signal for
C. neoformans that triggers capsular enlargement as a
non-specific defense mechanism against potential predatory cells. Hence,
phospholipids are the first host-derived molecules identified to trigger
capsular enlargement. The parallels apparent in the capsular response of
C. neoformans to both amoebae and macrophages provide
additional support for the notion that certain aspects of cryptococcal virulence
emerged as a consequence of environmental interactions with other microorganisms
such as protists.
| A key event in C. neoformans pathogenesis is capsule enlargement
in mammalian hosts. Historically, this phenomenon was attributed to high
CO2 and iron deprivation but the magnitude of capsular
enlargement observed in vivo cannot be consistently replicated
in vitro. This paper reports that C.
neoformans responds to polar lipid extracts with massive capsule
enlargement, with some cells having dimensions comparable to the giant cells
observed in vivo. Phospholipids are identified in this paper as
the inducers of capsule enlargement. Our work is important because this is the
first host-derived molecule that has been identified as a stimulus of massive
capsule enlargement thus providing a potential mechanism for the capsular
enlargement observed in vivo. Furthermore, the fact that the
signal is common to both macrophages and amoebae suggests that the capsule
enlargement response to phospholipids is a mechanism for fungal sensing of
phagocytic cell predators. This provides another example of a correspondence
between a possible environmental signal and a mechanism of virulence.
| Certain environmental microbes exist that have no obvious need for animal virulence
with regards to their survival or propagation yet these organisms have the ability
to cause infection and disease in a human host. One such organism is the soil fungus
Cryptococcus neoformans, a major pathogen for immunocompromised
individuals, such as those with advanced HIV infection. C.
neoformans has several well-characterized virulence factors [1], and the most
extensively studied virulence factor is its polysaccharide capsule [2], [3]. The capsule
is believed to contribute to virulence through multiple mechanisms as it is both
anti-phagocytic and capable of causing detrimental effects on host immune system
functions [3].
The polysaccharide capsule is also a powerful free radical sink that protects the
fungal cell from oxidants, such as those produced in the oxidative burst of
phagocytic cells [4]. A remarkable property of the capsule is its ability to
undergo enlargement during infection and this phenomenon is associated with
cryptococcal virulence in the mammalian host [5]. This enlargement can result
in gigantic cells that exceed the size of macrophages [6], [7]. Several factors have been
shown to induce this capsular enlargement, including high CO2, low iron,
basic pH, and mammalian serum [8]. Additionally, capsular enlargement intensifies protection
against both phagocytosis and oxidative damage [4], [9].
C. neoformans is a facultative intracellular pathogen with a unique
replication strategy in macrophages [10], [11]. The sophisticated virulence
strategies utilized by C. neoformans in the human host and the
ability of cryptococcal polysaccharide to interfere with the immune response might
suggest that such virulence factors as the polysaccharide capsule have evolved for
evading mammalian defenses. However, given that C. neoformans does
not require a mammalian host for replication and survival, the evolutionary origin
of such sophisticated virulence strategies has been a perplexing problem in the
field. Consequently, there has been considerable interest in characterizing the
interactions of C. neoformans with other soil organisms.
Acanthamoeba polyphaga was shown by Bunting et
al. to interact with and ingest cryptococcal cells in classic studies
carried out in the 1970s [12]. In 2001, Steenbergen et al.
demonstrated that the interaction of A. castellanii with C.
neoformans was similar to that observed with macrophages [13]. Recently,
this concept has been extended to the emergence of fungal virulence for insects
[14]. Our
group also described the interaction of C. neoformans with three
Paramecium spp., which were shown to rapidly ingest and kill
the fungus [15].
These results led to the proposal that the virulence strategies used by C.
neoformans for survival in mammalian hosts had emerged and developed
through environmental interactions, due to the constant selection by predation [16], [17]. In this
scenario, cryptococcal virulence factors are the result of environmental selection
and serve this microbe in mammalian infection by the accidental adaptation to the
host [18].
Additional evidence for this theory comes from the finding that non-lytic exocytosis
from macrophages [19], [20] is also observed with amoeba [21].
Consequently, we hypothesized that the increase in capsule size may also occur in
interactions with amoebae, perhaps as a mechanism to avoid phagocytosis by those
predators. Upon co-incubation of the amoeba and the fungus, we observed that
C. neoformans responded to the protist by increasing its
capsule size. We have characterized the properties of capsule–inducing amoebic
extracts, including their composition, stability, and effects on the C.
neoformans cells. Additionally, we have observed that the same response
could be elicited by both live and dead macrophages. These observations led to a
search for the signal sensed by C. neoformans and we report that
phospholipids can trigger both capsule enlargement and giant cell formation.
Co-incubation of C. neoformans with amoeba elicited an
approximately four-fold increase in the cryptococcal capsular volume as compared
with yeast cells in the absence of amoeba (Figure 1A and 1C, *p<0.0001). Increase in
cryptococcal capsular volume was similarly observed when C.
neoformans cells were incubated with macrophages as previously
observed [22] (Figure
1A, *p<0.05). Increasing the incubation time to 48 h produced
similar results (data not shown). Co-incubation of C.
neoformans with either dead protozoa or macrophages also resulted
in capsular enlargement, indicating that phagocytic cell viability was not
required for this effect (Figure
1A), although both viable amoeba and macrophages induced a more
pronounced effect, twice as much as the capsule increase observed with dead
organisms.
We investigated whether the capsule inducing molecule was diffusible by using
24-well flat-bottom plates where C. neoformans cells were
separated from A. castellanii by 0.4 µm cell culture
inserts which prevented the passage of either organism, but allowed the passage
of small soluble molecules. The C. neoformans cells to be
assessed were placed in PBS below the inserts to allow diffusion to carry any
molecules to these cells. The conditions above the inserts were varied, and
included PBS alone as a control, A. castellanii to determine if
A. castellanii alone produced a stimulant molecule, and a
combination of A. castellanii and C.
neoformans to determine if the combination was necessary to
stimulate production of a small molecule or its chemical modification. The
C. neoformans capsules were measured at 24 and 48 hrs.
Measurements included conditions in which only PBS was placed below the insert
for 24 hrs and then C. neoformans was added for an additional
24 hrs in the event that the reaction required more time because of the physical
separation, but no consistent effects on capsule enlargement were observed (data
not shown). Additionally, when C. neoformans was placed in
A. castellanii cell-free supernatants, no capsular
enlargement was observed, suggesting that A. castellanii does
not secrete a capsule-inducing molecule in solution (data not shown). We then
set up a large volume co-incubation of A. castellanii and
C. neoformans and tested the concentrated supernatant from
this interaction in the activity assay. No activity was found, suggesting that
the active moiety was either not released into the supernatant, that it remained
on the surface of the amoeba, or that it rapidly lost activity in solution (data
not shown).
We considered whether capsular enlargement was a result of mechanical stimulation
by a foreign object such as the amoebae. This was tested by incubation of fungal
cells with 9.2 µm polystyrene beads. There was no statistical difference
between the capsule volume after incubation with beads and the volume when
C. neoformans was incubated alone in PBS (Figure 1A, p>0.05), despite
documenting that yeast and bead cells could be found in close proximity to one
another (data not shown).
Comparisons of the C. neoformans volume alone in PBS and after
incubation with A. castellanii revealed that the capsule volume
was enlarged at room temperature and above, specifically at 28°C and
37°C (Figure 1B,
*p<0.05). The highest ratio of capsule induction was obtained with
incubations at 28°C, which is the optimum temperature for the growth of the
amoeba.
In considering common triggers found in both macrophages and amoebae, we focused
on membrane lipids, given that cell membranes are highly conserved in eukaryotic
evolution. A. castellanii lipid extraction was performed using
the Folch Method [23]. This procedure gave rise to three fractions (aqueous
upper phase, interface, and organic lower phase) which were separated and
incubated individually with C. neoformans serotype D strain
24067. After co-incubation with the extracted fractions for 24 and 48 h, we
observed that only the upper polar phase, normally considered the
“non-lipid” fraction, induced a significant increase in capsular
volume that was comparable to that observed with intact A.
castellanii (Figure
2A, p>0.05), while no effect was observed with the lower organic
phase or the interface fractions. This effect of the upper polar phase on
capsule enlargement was also observed using a C. neoformans
serotype A strain, H99 (Figure
2B), with a more pronounced increase compared to controls, despite
the lower absolute numbers compared to the serotype D 24067 strain.
Additionally, when J774.14 macrophage-like cells were subjected to the same
lipid extraction protocol and tested in the activity assay, we observed parallel
results. Upon co-incubation with C. neoformans, the macrophage
upper polar phase was again found to cause capsular enlargement equivalent to
the intact cell (data not shown).
Given that capsule enlargement has been linked to stationary cell growth [8], we
investigated the effect of the A. castellanii polar extract on
fungal growth. Using C. neoformans strain 24067, we compared
the growth in PBS and in SDB for 24067 incubated either alone or with various
concentrations of polar extract. In SDB, the growth curves were found to be
identical. In PBS, growth rates were much slower, however, after 40 h, cells
grown with lipid extract manifested increases in growth rate relative to PBS,
presumably as a result of the fact that the extract could provide nutrients
(data not shown).
We noted that amoeba polar extracts often lost their ability to induce capsule
growth upon storage or additional purification. Analysis of the stability of the
capsule-inducing component over time revealed a rapid decrease of activity
(Figure 2C) such that
the half life of activity decay was calculated to be 1.385 days.
We evaluated the production of both capsular and exopolysaccharide following
overnight incubation with amoeba extract in PBS or in minimal medium. Incubation
of C. neoformans with the amoeba upper phase polar extract in
PBS resulted in a 2-fold increase in capsular polysaccharide and a 6-fold
increase in exopolysaccharide relative to the amount produced in PBS alone
(Figure 3). When
C. neoformans was incubated in minimum medium containing
high glucose concentrations, no difference was observed for either capsular and
exopolysaccharide (p>0.05) production, whether incubated with amoeba extract
or alone in the medium.
Since the capsule enlargement phenomenon required contact between C.
neoformans and amoeba cells, we considered whether the release of
polar lipids from amoeba membranes could be a step in amoeba-mediated capsular
enlargement and thus evaluated the requirement for fungal phospholipase in this
process. Phospholipase B (PLB) can release both sn-1 and
sn-2 fatty acids from phospholipids [24]. C.
neoformans produces extracellular PLB and both PLB-deficient
(plb1) and reconstituted
(plb1REC1) strains have been generated on an H99
background [24].
Both the parental and the reconstituted strains exhibited the same increase in
capsule volume when incubated with either intact A. castellanii
cells or with the upper phase fraction, but not when incubated with PBS (Figure 4A, *p<0.05).
However, the plb1 mutant strain was unable to increase the
capsular volume under any of these conditions, indicating a necessity for PLB
production.
The necessity for PLB activity led us to investigate whether other phospholipases
may also play a role in the interaction between the two organisms. The
C. neoformans strain 24067, as well as the parental (H99),
the phospholipase C mutant (Äisc1), and the reconstituted
(Δisc1REC) strains were tested in the
activity assays with A. castellanii. The C.
neoformans phospholipase C mutant strain did not display any
defects in capsule enlargement upon co-incubation with A.
castellanii (data not shown).
The C. neoformans capsule can be enlarged by incubation in
10% fetal calf serum (FCS), as described [22]. Consequently, we
investigated whether the capsule-inducing properties of serum were also due to
polar lipids and whether the effect was also PLB-dependent. Using the same
extraction protocol that was used with A. castellanii, FCS was
separated into upper phase, interface, and lower phase fractions. C.
neoformans strains H99, plb1, and
plb1REC1 were tested in activity assays where
they were each incubated with the FCS fractions (Figure 4B). Similar to the results with
amoebae, the FCS capsule-inducing activity was also found in the upper polar
fraction (p<0.05). However, unlike the A. castellanii polar
extract, the FCS polar lipid fraction induced enlargement of the
plb1 capsule, suggesting that for serum-derived polar
lipids, there is no PLB requirement for capsular enlargement.
Both heat and glucanase treatments were found not to affect the ability of the
extract to trigger capsule enlargement (Figure 5A, p>0.05). However, when the
fractions were treated with Proteinase K and heat, we observed a dramatic
39% increase in the capsule induction activity when compared to the
untreated extract, suggesting that the active compound may be complexed with a
protein in solution and that its cleavage helps to release the active compound.
The fraction treated with Proteinase K and heat was subsequently run on a silica
TLC plate with a mobile phase of a 65∶25∶4 ratio of
methanol∶chloroform∶water (Figure 5B). A new band appeared, with a
higher Rf, most likely consisting of free lipids released after proteinase
digestion. We performed a new fractionation using the Folch method following the
Proteinase K digestion. New upper and lower phases incubated overnight or for 48
h with C. neoformans showed different results. Activity was
fractionated to the lower phase, which induced capsule enlargement compared to
PBS, with similar levels to treatment with Proteinase K without fractionation,
after both overnight (Figure
5C, *p<0.05) and 48 h (Figure 5D, *p<0.05) incubations.
Interestingly, at both time-points evaluated, the new upper phase material
obtained after fractionation of the Proteinase K digestion resulted in an
increase in cell body volume when compared to PBS.
The requirement of PLB for capsular enlargement combined with the new band
observed in the polar extract TLC after Proteinase K treatment suggested that
the enlargement activity was due to a type of phospholipid or
phospholipid-derived molecule from the A. castellanii extracts.
Thus, we tested the ability of a few purified commercially available lipids and
lipid-derived molecules to trigger capsule enlargement.
One of those molecules was phosphatidylcholine (PC), a major component of amoeba
cell membranes. We observed that PC was able to induce a dose-dependent capsule
enlargement in two different strains of C. neoformans
comparable to the one obtained in the co-incubation experiments (Figure 6A and 6B; Figure
S1B). The average increase varied from 2- to 8-fold (p<0.001)
depending on the conditions of the experiment, with larger increases when the
cells were incubated in MM instead of PBS and for at least 48 hours. In addition
to PC we also tested phosphatidic acid (PA), phosphatidylethanolamine (PE),
phosphatidylglycerol (PG), phosphatidylinositol (PI), phosphatidylserine (PS)
and lysophosphatidylcholine (LC) (Figure S1B). All the compounds with the
exception of PG and PS produced significant enlargement of the C.
neoformans capsule, however the effects were highest in the
presence of either PC or LC.
Additionally, mass spectrometry of samples from the amoebae polar extract treated
with or without Proteinase K, demonstrated the presence of phospholipids.
However, phospholipid concentration was very low when compared to the amount
detected in the untreated non-polar samples by mass spectrometry (data not
shown), therefore precluding molecular identification.
C. neoformans PLB activity has been linked to the generation of
arachidonic acid from fungal phospholipids and to the subsequent production of
eicosanoids, including prostaglandins [25]. Thus, we hypothesized that
arachidonic acid or one of its products could be responsible for the capsular
enlargement, but none of the compounds tested promoted capsule growth (Table
S1).
In addition, we considered that the effect on the C. neoformans
capsule could be caused by the polar head group of the phospholipids, which also
can be derived from PLB activity. It has previously been shown that GPC is the
only degradation product of PC upon treatment with C.
neoformans supernatants containing PLB activity [26]. Thus, two
commercially available polar head groups, GPE and GPC, were tested with
C. neoformans strains 24067 and H99 in the activity assay
(Figure 6A and 6B, respectively). We observed
that both molecules were able to induce capsule enlargement with differences in
their effect dependent on the C. neoformans strain used (Figure 6A and 6B). After 24 hours of
treatment, 10 µM of GPC was able to induce an average 2-fold increase in
the capsule volume of 24067 and H99 cells (data not shown) reaching a 5-fold
increase in the first strain after 48 hours of treatment. GPE produced an
average 2-fold increase with both strains, however, was slightly but
significantly more active with H99 cells. Additionally, GPE and GPC were able to
induce capsule enlargement in the PLB-deficient strain comparable to the
enlargement previously observed in the presence of amoeba and polar extracts
(Figure 6C). In this
case, GPE was also shown to be more active against the plb1
mutant than GPC and was also able to induce the presence of giant-like cells in
the mutant cultures. As observed with the polar extracts, both GPC and GPE were
also shown to lose their activity very rapidly when in solution. We tested
concentrations of GPC and GPE up to 1 mM, however, concentrations higher than 10
uM did not result in further increases in the capsule enlargement (Figure S2).
Conversely, incubation of plb1 mutant cells with intact
phospholipids, such as PC, was not able to induce capsule enlargement, thus
supporting the necessity of the phospholipase B activity in this process (data
not shown).
C. neoformans PLB contains three enzyme activities in one
protein, phospholipase B (PLB), lysophospholipase (LPL) and lysophospholipase
transacylase (LPTA). These activities have been found to be either secreted or
cell associated (either membrane bound or in the cytosol) [27]. PLB removes both acyl
chains from phospholipids; LPL removes the single acyl chain from
lysophospholipids; and LPTA adds an acyl chain to lysophospholipids to produce
phospholipids. To further evaluate the role of PLB in the capsule enlargement,
we tested the effects of three phospholipase inhibitors (as described [27]) on the
capsule enlargement induced by PC. The first inhibitor was alexidine
dihydrochloride (compound AX) which primarily inhibits secreted PLB activity at
the tested concentration. Another inhibitor was dioctadecyldimethylammonium
bromide (compound O) which acts mainly on secretory and cytosolic LPL and LPTA
and on cell-associated PLB. The third inhibitor was palmitoyl carnitine
(compound PAC), which has been found to be a potent inhibitor of PLB activity at
0.5 mM while affecting LPL and LPTA activities by only by 35% [28]. We found
that compound AX did not affect the capsule enlargement induced by PC, however
both compound O and compound PAC, which target cell-associated PLB and secreted
and cell associated LPL and LPTA activities, abolished the capsule enlargement
(Figure
S3). These results further support the role of phospholipase B in the
capsule enlargement and suggest that the PLB activity involved in the process is
possibly cell-associated.
Incubations of C. neoformans yeast with A.
castellanii, macrophages, and their respective extracts were
evaluated in longer incubation periods for the induction of cell gigantism. At
days 2, 4, and 6, co-incubation with amoeba, macrophages, and the extracts all
induced larger capsule volumes when compared to incubation in PBS or minimum
medium alone (Figure 7A,
p<0.05). After 8 days, we observed an increase in the cell body volume of
these cells and a concurrent reduction of relative capsular volume, but the
overall volume of the C. neoformans yeast did not display a
statistically significant difference. Cells incubated with amoeba extract and
with the intact amoeba cell have a distinct pattern, with a double-layered
capsule and two regions of different density. These cells had diameters ranging
from 15 to 20 µm, approximately the size of giant cells previously
described as forming in vivo
[7], [22] (Figure 7B). Analysis by
indirect immunoflourescence of capsule-induced C. neoformans
cells after 6 days revealed stronger binding than the capsules of C.
neoformans cells in PBS, consistent with the capsule enlargement
phenomenon previously described (Figure 7C). Additionally, India ink staining and immunofluorescence
of C. neoformans cells exposed to PC manifested very large
C. neoformans cells as early as 48 hours that were even
larger than those observed after extended incubation with A.
castellanii, macrophages, and their respective extracts (Figure 8A, 8B, and 8C). Those
cells did not constitute the majority of the cells, but were notably larger than
untreated cells (Insets in Figure
8A and in 8B and Figure
8D). The whole cells averaged from 20 to 30 µm and both cell
body and capsule were significantly enlarged and their relative abundance
appeared to increase after longer incubation periods (Figure 8C and 8D).
C. neoformans capsular enlargement is a phenomenon that has
frequently been associated with its virulence in mammals [5]. Numerous signals are known to
trigger capsule enlargement including CO2
[29], serum [8], iron
deprivation [30], pH, and certain growth conditions [22]. The fact that chemically
diverse signals trigger capsule enlargement suggests that this phenomenon may be a
non-specific defense against fungal-perceived stress, threats, and danger. When
murine lungs are inoculated with C. neoformans, capsular
enlargement proceeds rapidly. The phenomenon, in its extreme and combined with
cellular growth, can also result in the formation of gigantic cells [31]. However,
since the primary ecologic niches of C. neoformans are soils and
trees, and animal infection may be a relatively rare event involving only a minute
fraction of the cryptococcal fungal mass on the Earth, it is unlikely that capsular
enlargement evolved for the specific purpose of defense in animal hosts. Given that
soil amoebae have been reported to be major predators of C.
neoformans in that niche [32], we investigated whether interactions with protozoa also
induced capsular enlargement.
Incubation of C. neoformans with A. castellanii
resulted in capsular enlargement. The effect required contact between the fungal and
protozoan cells but did not require amoebae viability. Since the capsule protects
C. neoformans against amoebae ingestion, and since the diameter
of the capsule correlates inversely with the efficiency of phagocytosis [9], [21], capsular
enlargement is a likely defense mechanism against phagocytic predators. The absence
of a capsular enlargement response when cryptococci are incubated with beads implies
that the stimulus is more than merely mechanical and that fungal cells discriminate
between inert spheres and cells.
Since both live and dead amoebae triggered capsular enlargement and since protozoan
cells represented a very different trigger than previously reported stimuli, we
investigated the nature of the responsible component by fractionating amoebae cells
and testing these fractions for their ability to elicit capsule growth. One of the
approaches was to submit A. castellanii cells to lipid extraction.
The upper polar phase, normally called the non-lipid fraction, had comparable
efficacy to intact amoeba cells in promoting capsular enlargement. In contrast,
neither interface nor the lower phase lipid extract fractions demonstrated any
capsule enlargement activity. Concurrently, due to the requirement for cellular
contact, we investigated the requirement for fungal phospholipase in
amoebae-promoted capsular enlargement. Phospholipase B, but not phospholipase C, was
required for C. neoformans to respond to amoebae with capsular
enlargement.
Combined, these two results were conflicting since the partitioning to the upper
phase during lipid extraction suggested that the enlargement activity was not due to
a lipid molecule, given that most of the lipids are normally found in the lower
organic phase. Additionally, the requirement for phospholipase B suggested that the
molecule was a phospholipid or at least a phospholipid degradation product. The
treatment with Proteinase K and the subsequent TLC analysis gave us a possible
explanation for this potential inconsistency. Instead of abolishing the activity, as
would be expected if the activity was due to a polypeptide chain, the treatment
actually increased the extract activity and produced a new band in the TLC that was
compatible with the release of a phospholipid. In the case of lipids that are
covalently associated to proteins or carbohydrates, they could be carried to the
non-lipid extract during phase partitioning [33]. Thus, our hypothesis is that
the phospholipids responsible for the enlargement activity are strongly associated
with a polypeptide and this association results in their partitioning to the upper
polar fraction during the extraction. As expected, the activity was transferred to
the lower phase upon Proteinase K digestion and new fractionation, supporting the
hypothesis that free lipids are released. The treatment with Proteinase K disrupts
this interaction, further building upon the activity of phospholipase B. Additional
support for this observation comes from the fact that, even in the absence of strong
interactions with other molecules, some phospholipids and other highly polar lipids,
such as gangliosides, partition into the upper phase [34]. Furthermore, mass
spectrometry of both intact crude amoeba polar extracts and those treated with
Proteinase K indeed revealed the presence of different classes of phospholipids in
our samples but their identity could not be established due to small quantities.
Given that phospholipases are known to damage membranes, we interpreted this result
as indicating that fungal phospholipase B catalyzed the release from the membrane of
lipids and/or lipid fractions that are subsequently sensed by the fungal cells.
Phospholipase B is known to be a virulence factor for C.
neoformans, but the dependence of capsule enlargement on this activity
implies a potential new role in cryptococcal biology.
Incubation with amoeba fractions also altered the production of both capsular and
exopolysaccharides. We measured an approximately two-fold increase in the capsular
polysaccharide and a six-fold increase in the exopolysaccharide production in the
presence of amoeba lipid extracts. Given the structural differences in the capsular
and exopolysaccharide fractions, the quantitative differences in production are
consistent with the notion that these compounds have independent pathways of
production and/or secretion [35].
The observations that protozoan phospholipids triggered capsular enlargement prompted
us to investigate whether mammalian lipids had the same effect. The polar fractions
extracted from macrophages and serum were also shown to trigger capsular
enlargement. An interesting difference between the effects observed with amoebae and
with serum was the absence of a phospholipase B requirement in the capsular
enlargement response to serum-derived polar lipids. This observation suggests that
serum lipids are responsible, at least in part, for the ability of serum to trigger
capsule growth. However, serum also contains iron binding proteins that could
conceivably indirectly trigger capsule growth through iron limitation [30].
We then attempted to identify the specific compound responsible for the capsular
enlargement. Our first approach was to further fractionate the upper phase from the
lipid extraction using a variety of techniques. However, activity was inevitably
lost with progressive fractionation. Size exclusion and reverse phase chromatography
purification of the polar fractions revealed activity in at least two fractions but
mass spectrometry analysis of the most active fractions was not revealing
(unpublished data). This suggested that the compound was not stable and/or that the
effect required more than one molecule. Indirect evidence for the instability of the
compound comes from the observation that we were never able to demonstrate capsular
growth induction in experiments where fungal and amoeba cells were separated by
diffusible membranes. Direct evidence for the instability of the capsule-inducing
compound comes from the observation that extracts rapidly lost their activity when
stored at room temperature. The instability of the activity suggests an explanation
for our difficulties in the attempts to further purify and identify the active
compound(s) responsible for capsule enlargement. In retrospect, the putative
identification of the active compound as a phospholipid suggests an explanation for
its instability since these compounds are rapidly degraded by molecular oxygen and
our protocols did not involve working in oxygen-free conditions. However, it is also
possible that the inability to demonstrate capsular enlargement in assays with
diffusible membranes indicates strong concentration dependence such that the effect
is lost with dilution.
Our second approach to molecular identification was to consider compounds that may be
present in the polar extract, to obtain them in pure form, and to test them
individually, and sometimes in combination, for their effects on capsule growth.
Using this approach, we found that phospholipids, in particular, phosphatidylcholine
(PC) and lysophosphatidylcholine (LC) and two glycerophosphodiesters, GPC and GPE,
that are components of the polar head of phospholipids, were able to reproduce the
C. neoformans capsule enlargement. Additionally, GPC and GPE
were able to overcome the inability of the phospholipase B mutant to enlarge its
capsule in response to the amoeba extract. This was in contrast to intact PC,
supporting the necessity of PLB activity to generate these small compounds that
trigger phospholipid-mediated capsule enlargement. The fact that GPC and GPE are
regularly found in brain and other host tissues [36], also suggests that they
could act as possible triggers for the capsule increase observed with C.
neoformans in the host environment [37].
Although the mechanism by which phospholipids trigger capsule enlargement was not
elucidated as part of this study, our working hypothesis is that certain
phospholipids can trigger signaling cascades in C. neoformans that
in term promote capsule synthesis. In this regard, we note that members of the human
oxysterol binding protein (OSBP) family can bind phospholipids [38] and it is conceivable that in
C. neoformans this highly conserved family, or another
signaling set of proteins, has been specialized to bind phospholipids.
The finding that phospholipids triggered capsular enlargement led to a conundrum;
neither the lower phase extract from amoebae nor from macrophages mediated this
effect, however phospholipids would have been abundant in the organic layer. Our
hypothesis is that the complex lipid solution in the lower phase includes both
stimulators and inhibitors of capsule enlargement. In this regard, we note that this
fraction would also include all the sterols and this class of compounds can trigger
signal transduction by the OSBP-related protein system [39]. In yeast, stimulation of these
proteins inhibits golgi vesicular production [40]. An analogous effect in
C. neoformans could shut down capsule production since the
polysaccharide is synthesized in golgi-derived vesicles [41]. In this regard we note that in
the dose response data with GPC and GPE, the amount of enlargement peaked at 10
µM and declined at higher concentrations consistent with an inhibitory effect.
Alternatively, there could be a possible nutritional explanation. Capsule size is
known to be negatively regulated by nutrient-rich media and high glucose
concentrations (reviewed in [3]). Consequently, it is possible that the lipid-rich
environment of the lower phase is perceived by the fungus as nutrient-rich, leading
to inhibition of capsular enlargement. Given that capsular enlargement has been
associated with poor nutrient preparations and stationary phase growth conditions,
combined with the recent observation that giant cell formation in C.
neoformans follows cell cycle progression without fission [6], [7], we decided to
evaluate the effects of our lipid preparations on cell growth but found no effect.
Similarly, lipid-induced capsule growth was observed at all tested temperatures with
the exception of extremely low, non-physiological temperatures. The occurrence of
capsule enlargement at temperatures ranging from ambient to mammalian, suggests that
this phenomenon can occur both in environmental niches and during mammalian
infection.
We also observed that co-incubations with extracts or cells of A.
castellanii and macrophages for extended periods induced the formation
of very large cells. Although these cells did not achieve the full dimensions of
giant cells described in vivo, they approximated that size and
represented a tremendous increase in both cell and capsule sizes. Given that
detailed studies of giant cells are likely to require the ability to induce them
in vitro, the finding that these extracts promoted their
formation is an important development for future progress in understanding their
cell biology. This type of C. neoformans cells displayed a
double-layered capsule, consisting of a denser region close to the cell body, and an
outer layer, which permitted a higher penetration of India ink particles.
Immunofluorescence studies revealed that binding of mAbs to the capsule of
C. neoformans cells following incubation with amoeba or
extracts was more intense, indicating the presence of more reactive polysaccharides
surrounding the yeast. An increase in the cell body was observed after eight days of
incubation with amoeba or macrophage cells and extracts when compared to PBS alone,
along with a concurrent decrease in capsular volume, but no alteration in the whole
cell volume. This suggests that in the conditions of starvation that accompany the
late stationary phase, C. neoformans might be reusing the capsular
polysaccharides as an energy source.
In summary, we describe a new trigger for cryptococcal capsule enlargement that is
present in the polar lipid fractions derived from amoebae, macrophages, and
mammalian serum. We propose a model for C. neoformans capsule
enlargement resulting from interactions with amoebae or macrophages, whereby
C. neoformans induces release of phospholipids from the
phagocytic cell membrane after action of PLB possibly facilitated by C.
neoformans proteases [42]. Those phospholipids are then cleaved by PLB releasing
their polar heads that are in turn sensed by C. neoformans cells,
triggering capsule enlargement and the formation of giant cells (Figure 9). Since capsule
enlargement reduces the phagocytic efficacy of both amoeba and macrophages [4], [21], we propose
that this is a general cryptococcal defensive response to the sensation of potential
danger. The observations with phospholipase B-sufficient and -deficient cells,
suggest that fungal enzymes are used to damage amoeba cell membranes and release
lipid components that subsequently trigger capsule growth. According to this view,
the fungus would sense the lipid components and/or cleavage products (GPC or GPE) as
signals of potential danger in the form of predatory phagocytic cells in their
immediate vicinity. In this hypothesis, phospholipids join other known triggers of
capsule growth such as Fe deprivation, CO2, and pH as stress signals to
which the fungus responds by capsular enlargement. To our knowledge, these are the
first host-derived compounds identified to promote capsular and cellular
enlargement. Our observations provide yet another striking parallel between the
response of C. neoformans to amoebae and macrophages. Such
similarities, combined with the observations that virulence can be enhanced by
passage in amoeboid cells [1], have been used to argue that the capacity for virulence
in C. neoformans and other soil-dwelling organisms with no
requirement for animal hosts is a consequence of selective pressures in soils from
the presence of protozoa. Recently, the same argument has been put forward to
explain the virulence of certain fungi for insects [14]. The parallel responses of
C. neoformans to macrophages and amoebae provide additional
support for the view that cryptococcal virulence is a result of selection of certain
traits by environmental pressures that also enhance survival in animal hosts.
A. castellanii strain 30234 and C. neoformans
strain 24067 were obtained from the American Type Culture Collection (ATCC,
Manassas, VA). The amoebae were cultured in peptone-yeast extract-glucose broth,
PYG (ATCC medium 712, containing 10 mM glucose), in tissue culture flasks at
28°C. A. castellanii cells were used when confluent on the
bottom of the flask and were passaged every 5-7days, as described [43]. The
C. neoformans strains 24067 (serotype D) and H99 (serotype
A) were grown from frozen stocks and maintained in Sabouraud dextrose broth
(SDB, Difco, Lawrence, KS) or minimal medium (MM, 15 mM glucose, 10 mM
MgSO4, 29.4 mM KH2PO4, 13 mM glycine, and
3.0 ìM thiamine). C. neoformans strains H99,
plb1, and plb1REC1
[24] were
obtained from Drs. Gary Cox and John Perfect (Durham, NC). C.
neoformans strains H99, Δisc1, and
Δisc1REC strains [44] were obtained from Dr.
Maurizio Del Poeta (Charleston, SC).
After growing C. neoformans as described above, the cells were
washed 3 times in phosphate buffered saline (PBS) and 1×106
yeast cells were suspended in either PBS or MM and placed in 96-, 24-, or 6-well
plates. Cryptococcal cells were incubated with 1×106 of either:
live or dead A. castellanii or J774.14 macrophage-like cells.
Alternatively, C. neoformans was incubated with 9.2 µm
polystyrene beads. Incubations were done overnight, for 24 h or 48 h at
28°C. Dead amoebae were obtained by boiling the organism for 5 minutes.
Lysing of amoeba cells was accomplished by forcefully pulling and pushing the
cell suspension through 26.5 gauge syringe needles 15-20 times. Additionally,
C. neoformans cells were incubated in medium conditioned by
prior growth of A. castellanii or cell-free supernatant from a
previous overnight co-incubation experiment. To evaluate the effects of
temperature in the interaction, plates with C. neoformans
strain H99 and A. castellanii were also incubated overnight at
4°C, room temperature, 28°C, and 37°C.
The volume of the capsule both before and after exposure to the various
conditions was measured using India ink suspensions, as previously described
[8].
After overnight incubation for each condition tested, C.
neoformans cells were washed from the wells, spun down, and in some
cases stained with Uvitex 2B (Polysciences, Inc.), and then all aliquots were
spotted on microscope slides, mixed with a drop of India ink, and examined using
an Olympus AX70 microscope at a magnification of 40X (Center Valley, PA). Cells
suspended in India ink were photographed with a QImaging Retiga 1300 digital
camera using the QCapture Suite V2.46 software (QImaging, Burnaby, British
Columbia, Canada). Alternatively, C. neoformans cells were
observed in an Axiovert 200 M inverted microscope using a 40X objective (Carl
Zeiss Micro Imaging, Thornwood, NY) and photographed using a Hamamatsu ORCA ERJ
camera (Hamamatsu Photonics, Hamamatsu City, Japan). The volume of the
C. neoformans capsule was measured using Adobe Photoshop
7.0 for Windows (San Jose, CA.), or AxioVision software (Carl Zeiss Micro
Imaging, Thornwood, NY). The diameter of the whole cell
(Dwc) and the cell body
(Dcb) were each measured and the capsule width
was defined as the difference between Dwc and Dcb.
diameters. The volume of the capsule was calculated using the equation for the
volume of a sphere, 4/3 Π(D/2)3, such that the
capsule volume (Vc) was the difference between the
whole cell volume (Vwc) and the volume of the cell body
(Vcb). For each condition, we averaged the capsule volume for a
minimum of 50 C. neoformans cells.
To determine whether cell contact was needed for C. neoformans
to respond with capsular enlargement, experiments were carried out where fungal
and amoeba cells were separated by means of filter inserts. C.
neoformans cells were placed in 24 well plates and separated from
A. castellanii by the presence of a cell culture insert
with a 0.4 ìm pore size (BD Falcon, Franklin Lakes, NJ). Prior to
co-incubation, C. neoformans cells were suspended in PBS and
placed below the inserts. Above the inserts, either A.
castellanii with C. neoformans, A.
castellanii alone, or PBS was then placed. The plates were
incubated for either 24 or 48 h at 28°C. A third group was incubated for 24
h with PBS below the filter prior to the addition of the C.
neoformans and then C. neoformans was added for an
additional 24 h of incubation. All organisms were suspended at a density of
1×106/mL and at an initial 1∶1 ratio of fungal to
amoeba cells.
Cells from confluent A. castellanii cultures were collected by
centrifugation at 320 x g for 10 min. J774.14 macrophage-like cells were also
collected by centrifugation after growth to confluence; however, they were spun
at 320 x g for 7 min. Cell pellets were washed three times with PBS. Resuspended
pellets (108 cells in 10 mL) of A. castellanii,
macrophages, or aliquots of Fetal Calf Serum (FCS) were each incubated with a
mixture of chloroform and methanol (2∶1 v∶v) for 2-3 h on a bench
top rocker at room temperature. The samples were then centrifuged for 10 min at
1100 x g for phase partitioning and the three phases obtained (upper, interface,
and lower) were collected and dried overnight in a vacuum centrifuge (Eppendorf,
Hauppauge, NY, USA). Interface and upper phase lipid fractions were resuspended
in PBS. Lower phase lipid fractions were resuspended in Dimethyl sulfoxide
(DMSO).
After resuspension, the amoeba and macrophage extracts were tested in capsular
enlargement activity assays with C. neoformans strains 24067
and H99 and serum extracts were added as well for the tests with C.
neoformans strains H99, plb1, and
plb1REC1. For each activity assay, C.
neoformans cells were washed, counted, and resuspended at
1×106 cells/mL in PBS. A 1 mL volume of the cell suspension
was added to each well of a 6-well plate. An additional 1 mL of PBS, and
1×106 of A. castellanii cells, were always
added to the first and second wells, respectively. In general, 1 mL of a
solution of the fraction to be tested was added to each of the subsequent wells,
with the concentration determined by the particular experiment. The plates were
incubated at 28°C overnight, 24 or 48 h and capsule volume was measured
using India ink staining (described above).
In addition to strains 24067 and H99, experiments were performed with C.
neoformans strains H99, plb1, and
plb1REC1 to determine the effect of
phospholipase B deficiency on the ability to respond to co-incubation with
amoeba extracts [24]. Strains H99, Δisc1, and
Δisc1REC were tested to determine the
effect of phospholipase C deficiency on the ability to show activity [44].
The upper phase of the amoeba extract was suspended in PBS and tested for its
ability to induce capsule enlargement as described above. A series of aliquots
were left at room temperature and one was tested each day for the ability to
elicit capsule enlargement. This experiment was conducted until no effect on
capsule enlargement was observed.
In order to evaluate the chemical characteristics of the active molecule(s) in
the polar extract, the extracts were submitted to various treatments. Treatments
included: (1) heat for 1 h at 65°C, (2) Proteinase K treatment [100
µg/mL in 50 mM Tris-HCl (pH 8.0) and 1.0 mM CaCl2] for 1 h
at 37°C followed by enzyme inactivation at 70°C for 30 min, or (3)
treatment with 1 U of Aspergillus niger β-glucanase (Sigma
Aldrich). The capsule volumes of the C. neoformans cells after
overnight incubation with the various treated extracts were then compared to
cells incubated in PBS alone or incubated with untreated extracts.
To investigate what was released after the enzymatic treatments listed above,
thin layer chromatography (TLC) was performed. A similar volume of all the
fractions to be tested was dried, resuspended in chloroform, and 25 µL
were spotted onto the membrane. A general separation of phospholipids based on
head group polarity was done using a mobile phase composed of
chloroform∶methanol∶water (65∶25∶4). Four of the main
phospholipids known to be present in A. castellanii cells,
L-α-Phosphatidylethanolamine (unsaturated, from Glycine max),
L-α-Phosphatidylcholine (unsaturated, from Glycine max),
L-α-Phosphatidylinositol (unsaturated, from Glycine max), and
1,2-Diacyl-sn-glycero-3-phospho-L-serine (unsaturated, from bovine brain) were
purchased from Sigma-Aldrich (St. Louis, MO) [45] and used as standards. TLC
plates were dried and stained in an iodine vapor chamber until the spots
formed.
Upon treatment with Proteinase K as described above, samples were submitted to a
second round of fractionation with a mixture of chloroform and methanol
(2∶1 v∶v) for 2–3 h as described above. Upper and lower phases
were then tested for capsular enlargement activity as described above.
Phospholipids known to be present in A. castellanii were
purchased from Sigma-Aldrich (St. Louis, MO) [45]. 3-sn-Phosphatidic acid
sodium salt from egg yolk lecithin (PA), L-α-Phosphatidylcholine from egg
yolk (PC), L-α-Phosphatidylethanolamine from egg yolk (PE),
L-α-Phosphatidyl-DL-glycerol sodium salt from egg yolk lecithin (PG),
L-α-Phosphatidylinositol from Glycine max (PG),
1,2-Diacyl-sn-glycero-3-phospho-L-serine from bovine brain (PS), and
L-α-Lysophosphatidylcholine from bovine brain (LC), were each tested with
C. neoformans for their ability to induce capsular
enlargement. Arachidonic Acid, Epoxyeicosatrienoic Acid, Thromboxane B2,
Prostaglandin E2, Prostaglandin I2, Leukotriene B4, and Leukotriene C4 were also
purchased from Sigma-Aldrich (St. Louis, MO). The powders were resuspended in
PBS, MM, or ethanol, serially diluted, and tested in the activity assay with
C. neoformans strains H99 and 24067. Purified
glycerophospholethanolamine (GPE) and glycerophosphocholine (GPC) were purchased
from Avanti Polar Lipids (Alabaster, Alabama). These two substances were tested
in activity assays with C. neoformans strains 24067, H99,
plb1, and plb1REC1. For GPC and
GPE, we tested concentrations ranging from 0.1 µM to 1 mM and found that
10 µM was the lowest concentration where we observed activity. The
activity did not increase at higher concentrations. For PC, we chose a 5 mM
concentration based previous studies [26]. Additionally, we carried out
a dose response study and found that for PC the effect was higher at
concentrations equal to or higher than 1 mM. As the effects of GPC, GPE, and PC
were stronger in MM in comparison to PBS, most of the tests were done in this
condition.
C. neoformans cells were incubated with amoeba extracts in
either PBS or MM and the resulting pool of polysaccharide was evaluated by
ELISA. Exopolysaccharide and capsular polysaccharide were measured by an
inhibition ELISA on reaction plates to which 10 µg/well of
glucuronoxylomannan (GXM) [46] was affixed overnight, followed by blocking with
1% (w/v) bovine serum albumin diluted in PBS (blocking buffer), and then
subjected to mAb binding [47]. A second blank ELISA plate (inhibition plate) was
blocked for 1 h at 37°C and a solution of 2 µg/mL of mAb 18B7 [48] was
incubated with serial dilutions of GXM (concentrations of 10 µg/mL to 0.06
ng/mL) or capsular and exopolysaccharide samples at 37°C for 1 h. Contents
of the wells were transferred to blocked reaction plates with adherent GXM as
antigens. After incubation at 37°C for 1 h, the plates were washed and
anti-mouse IgG conjugated with alkaline phosphatase (1∶1000 in blocking
buffer) was added to the wells for 1 h at 37°C. The plates were again
washed, incubated with a p-nitrophenyl phosphate substrate
solution and read at 405 nm. The concentration of GXM in the samples was
calculated extrapolating from the GXM standard curve.
Aliquots of C. neoformans suspensions following incubation with
amoeba extracts were washed with PBS, centrifuged, and suspended in a 50
µg/mL solution of 18B7 mAb in a 5% Bovine serum albumin solution in
PBS. Tubes were incubated for 1 h at 37°C while shaking and then washed
three times with PBS. The pellets were suspended in 100 µl of a
1∶100 dilution of FITC-conjugated goat anti-mouse IgG1 (Southern
Biotechnology) in blocking solution. Tubes were again incubated for 1 h at
37°C and washed with PBS. Pellets were then suspended in mounting medium
(Biomeda Corp, Foster City, CA) and spotted onto a microscopy slide. Slides were
examined with an Olympus AX70 fluorescence microscope using a 495 nm filter, at
a magnification of 40X. Alternatively, 107
C. neoformans cells were stained for 1 h with 10 µg/mL
DTAF-labeled 18B7 mAb and 0.01% Uvitex 2B. After washing, cells were
suspended in Prolong gold anti-fade mounting medium (Invitrogen, Carlsbad, CA)
and imaged using a 63x 1.4 objective. Z-stacks were collected and deconvolved
using a constrained iterative algorithm from AxioVision software (Carl Zeiss
Micro Imaging, Thornwood, NY).
A Bioscreen-C Automated Growth Curve Analysis System (Growth Curves USA,
Piscataway, NJ) was used to measure the growth of C. neoformans
strain 24067 in the presence of the A. castellanii polar
extract. The fungal cells were incubated either alone in PBS or in PBS and 1
µL or 10 µL of polar extract and compared with the cells incubated
in SDB alone or in SDB and 1 µL or of 10 µL polar extract.
Polar extracts and polar extract samples after treatment with proteinase K and
lipid re-extraction by the FoIch method were submitted to mass spectrometry
analysis by the Kansas Lipidomics Research Center (Manhattan, KS).
Statistical analysis was performed using GraphPad Prism 5 (La Jolla, CA). The
Shapiro-Wilk test was used to verify normal distribution of the measurements. To
determine significance, One-way ANOVA tests were used, followed by either
correction with the Bonferroni test for multiple pairwise comparisons for
normally distributed values or the Kruskal-Wallis analysis for measurements that
were not normally distributed. P values of less than 0.05 were
considered significant.
|
10.1371/journal.pntd.0007441 | Antifungal activity of two oxadiazole compounds for the paracoccidioidomycosis treatment | Paracoccidioidomycosis (PCM) is a neglected disease present in Latin America with difficulty in treatment and occurrence of serious sequelae. Thus, the development of alternative therapies is imperative. In the current work, two oxadiazole compounds (LMM5 and LMM11) presented fungicidal activity against Paracoccidioides spp. The minimum inhibitory and fungicidal concentration values ranged from 1 to 32 μg/mL, and a synergic effect was observed for both compounds when combined with Amphotericin B. LMM5 and LMM11 were able to reduce CFU counts (≥2 log10) on the 5th and 7th days of time-kill curve, respectively. The fungicide effect was confirmed by fluorescence microscopy (FUN-1/FUN-2). The hippocratic screening and biochemical analysis were performed in Balb/c male mice that received a high dose of each compound, and the compounds showed no in vivo toxicity. The treatment of experimental PCM with the new oxadiazoles led to significant reduction in CFU (≥1 log10). Histopathological analysis of the groups treated exhibited control of inflammation, as well as preserved lung areas. These findings suggest that LMM5 and LMM11 are promising hits structures, opening the door for implementing new PCM therapies.
| Paracoccidioidomycosis (PCM) is a granulomatous fungal infection with clinically severe forms and serious pulmonary sequelae. The current limited arsenal and prolonged treatment regimen demonstrate the need for new antifungals. This study reveals two fungicidal oxadiazole compounds for PCM treatment. The in vitro assay showed the antifungal activity and the synergic effect with Amphotericin B. The high doses administration in mice showed absence of toxicity, which allowed to demonstrate the in vivo antifungal activity.
| Paracoccidioidomycosis (PCM) is an endemic fungal disease in Latin American countries, which presents high prevalence in South America. The lung is the most affected organ, mainly during chronic form, presenting pulmonary architectural distortion, which can lead to hypoxemia and hypercapnia in 90% of patients with PCM [1]. In the last 30 years, the presence of pulmonary damage ranged from 63.8 to 100% in the patients [2, 3, 4, 5]. Furthermore, this injury remains even after the treatment and promotes pulmonary fibrosis with loss of respiratory function in 50% of patients [6, 7].
Considering this worrying scenario, the current available antifungal drugs are limited. In Brazil, only three therapeutics options are available for PCM treatment, such as polyenes, sulfanilamide and triazoles. The azoles action on the sterol biosynthetic pathway leads to many side-effects. Amphotericin B (AmB), a polyene, is the antifungal of choice in severe and acute cases. The treatment time should be as short as possible, between two and four weeks, due to its high toxicity [8]. The sulfanilamide is treatment options according to the severity of the disease; however, several disadvantages have been reported such as hypersensitivity reactions, gastrointestinal symptoms, hemolytic anemia, agranulocytopenia and crystalluria [9]. On the other hand, the most commonly used antifungal agent for treating mild and moderate forms of PCM is itraconazole (ITZ), but the time of therapy may reach 18 months and presents some collateral effects [10].
The major therapeutic challenges of this disease are the long period of continuous use of systemic antifungals, the possibility of relapses and the appearance of sequelae in the lung [1]. This, associated with the limited antifungal arsenal, evidences the necessity of the emergence of a new antifungal class.
Thus, the development of a drug that selectively acts on the target pathogenic fungi without producing collateral damage to mammalian cells is a pharmacological challenge. Biotechnological methods have become an important approach in pharmaceutical drug research and development. For example, the in silico methodologies not only reduce the cost associated with drug discovery, but they may also reduce the time it takes for a drug to reach the market [11]. This is a modern strategy to explore the interaction of compounds with a specific target [12].
By comparative genomics, ten potential targets for drugs occurring in eight human pathogenic fungi—Candida albicans, Cryptococcus neoformans, Aspergillus fumigatus, Blastomyces dermatitidis, Coccidioides immitis, Histoplasma capsulatum, Paracoccidioides brasiliensis and Paracoccidioides lutzii—were described [13]. One of these targets is thioredoxin reductase (Trr1), a flavoenzyme that acts primarily on resistance to oxidative stress, and it is essential to cell growth [14]. The trr1 mutation may result in hypersensitivity to hydrogen peroxide and to high temperatures [15]. In addition, this Trr1 isoform is found only prokaryotes and fungi [14]. Therefore, Trr1 a good target for the development of new anti-PCM therapies [16]. By molecular modeling and virtual screening, several compounds were selected as Trr1 ligands. Preliminary results showed that two compounds, which belong to the oxadiazole class, present antifungal activity against important pathogenic fungi such as Candida spp., Cryptococcus neoformans and Paracoccidioides spp. [17]. For this purpose, the antifungal activity of two oxadiazole compounds selected by in silico methods was tested both in vitro and in vivo against Paracoccidioides spp.
All the procedures were performed according to the regulations of the Ethical Committee for Animal Experimentation, State University of Maringá, Brazil (approval no. CEUA 9810191015, 22/04/2016). The animal’s experimentation were conducted according to the Guideline for the Care and Use of Laboratory Animals (CONCEA).
The compounds selected by virtual screening against thioredoxin reductase were commercially purchased from Life Chemicals Inc. (Burlington, ON, Canada). These compounds were named by LMM5 is 4-[benzyl(methyl)sulfamoyl]-N-[5-[(4-methoxyphenyl)methyl]-1,3,4-oxadiazol-2-yl]benzamide, and the chemical name of LMM11 is 4-[cyclohexyl(ethyl)sulfamoyl]-N-[5-(furan-2-yl)-1,3,4-oxadiazol-2-yl]benzamide (Fig 1). The stock solutions were prepared in dimethyl sulfoxide (DMSO) at concentration 100 μg/mL for LMM11 and 50 μg/mL for LMM5.
Nine isolates of Paracoccidioides spp. were used, three P. brasiliensis (Mg0113, Mg0213 and Pb18), three P. lutzii (Pb01, 8334 and Mg0114) and three isolates not identified yet at species level (Mg0116, Mg0216, Mg0115). Candida parapsilosis (ATCC 22019) and Candida krusei (ATCC 6258) were included for quality control. The isolates are part of the collection from Laboratory of Medical Mycology of the State University of Maringá, Brazil. The yeast phase was maintained by weekly passaging at 37°C in Fava Netto's solid medium. For each experiment, the viability of Paracoccidioides spp. was determined by counting viable cells in a Neubauer chamber by the trypan blue method. The assays were performed with ≥80% of viable cells [18].
Balb/c male mice, approximately six weeks old, with an average weight of 20 g, were raised at animal facilities of the State University of Maringá, Brazil. The animals were divided in groups and maintained in ventilated cages, with free access to tap water and food, in a controlled animal facility having a constant temperature of 23°C and a 12 h light/dark cycle.
The minimum inhibitory concentration (MIC) was determined by the broth microdilution method, following the standard methodology by the Clinical Laboratory Standards Institute (CLSI) published in document M-27A3, with modification for Paracoccidioides spp. [19, 20]. The oxadiazoles compounds’ concentrations ranged from 1 to 512 μg/mL. The inoculum was adjusted to 2 × 104 yeast cells/mL and diluted 1:2 into a 96-well plate with RPMI-1640 medium. Negative controls were only medium without inoculum, and positive controls were medium plus inoculum. The incubation time was 5 days at 37°C. Interpretation of the growth cutoff point was performed visually based on the comparison of growth in the positive control wells. The MIC values was defined as the lowest oxadiazoles concentration that resulted in at least an 80% reduction in growth relative to the positive control [21]. For AmB, it was considered to be the concentration causing 100% inhibition compared to the control without the antifungal drug. The drug controls were performed with AmB against C. parapsilosis (ATCC 22019) and C. krusei (ATCC 6258), according to the CLSI (document M27-A3) [19].
The minimum fungicidal concentration (MFC) of each compound was determined by transferring aliquots of 5 μL of each well from MIC microplates to brain-heart infusion (BHI) agar plates and incubating at 37°C for 7 days. The fungicidal activity was considered the lowest drug concentrations at which no colonies were able to grow. The following assays were performed with isolate Pb18.
The P. brasiliensis isolate Pb18 was cultivated in McVeigh Morton Chemically Defined Culture Medium (MMcM) for 7–10 days at 37°C under agitation at 150 rpm, to obtain yeast cells with typical multiple budding [22]. This culture was adjusted to 2.5 × 104 CFU/mL and treated with different concentrations of LMM5 and LMM11 (8 and 16 μg/mL) for 1, 3, 5, 7 and 14 days at 37°C. The untreated yeasts were used as controls. At each time interval, yeasts of each group were diluted in phosphate-buffered saline (PBS), and 100 μL was plated on Brain Heart Infusion (BHI) agar medium supplemented with 5% of Pb18 culture filtrate and 4% of Fetal Bovine Serum and incubated at 37°C for at least 14 days. The CFU were counted. The effect was considered fungicidal only when the CFU reduction was 3 log10 (≥99.9%); otherwise, it was considered fungistatic [23].
The metabolic activity of yeast cells of Pb18 was analyzed after exposure to the MIC concentrations of LMM11 and LMM5 (both 8 and 16 μg/mL, each). The assay was performed using FUN-1 and FUN-2 stains according to the manufacturer's protocol (Molecular Probes). Yeasts were suspended in MOPS buffer containing 2% glucose. The fungal cell activity was estimated with 0.5 μM FUN-1 (100 mM stock solution, dissolved in DMSO) and expressed as a change in the ratio of red fluorescence (k = 575 nm) to green (k = 535 nm). The viability of fungal cells was determined from examination of at least 200 cells in a biological replicate by fluorescence microscopy. A dead control was done using 70% ethanol to kill Pb18. Metabolically active cells fluoresce as red in their structures, while dead cells or cells with little or no metabolic activity exhibit bright diffuse green cytoplasmic fluorescence and lack of intravacuolar fluorescent inclusions [24].
AmB was chosen to test in combinations with LMM5 and LMM11 against Pb18 isolate. The compounds (starting at 4× MIC) were distributed vertically while AmB (4× MIC) was added horizontally as described by Bagatin et al. [25]. A 2 × 104/mL yeast cell suspension was added to 96-well plates and incubated at 35°C for 7 days. Inhibition was read visually and confirmed by XTT viability (492nm). The fractional inhibitory concentration (FIC) was determined by calculating ΣFIC = FICA + FICB = (CombAmB/MICAmB) + (CombLMM/MICLMM). For a strongly synergistic effect, FIC < 0.5; a synergistic effect, FIC < 1; an additive effect, FIC = 1; no effect, 1 < FIC < 2; and an antagonistic effect, FIC > 2 [26]. The Bliss-independent interactions were analyzed by Combenefit software [27].
Male Balb/c mice at 6 weeks old were divided into four groups: Control group treated with vehicle (PBS, DMSO 1%, and Pluronic F-127 0.2%); LMM5 group treated intraperitoneally with LMM5 at 25 mg/kg; and LMM11 group treated intraperitoneally with LMM11 at 50 mg/kg. The animals were monitored by Hippocratic screening at times 0, 15, 30, 60, 120, 240 and 480 minutes. After the 14th day, the mice were anesthetized for blood collection and euthanized. The biochemical examinations were performed, and the liver, heart and kidneys were weighed, on the day of euthanasia. The assay was performed in accordance with Salci et al. [28].
After inoculation with 106 Pb18 yeast (intratracheally), animals were randomly divided into experimental groups: LMM5, LMM11, ITZ (group treated with itraconazole) and control. The treatment started after 24 hours of infection. The compounds and ITZ were administered at 5 mg/kg, once per day for 14 days, intraperitoneally. The animals were euthanized by isoflurane vaporizer, and the number of CFU/g of the lung tissue was determined [20].
Mice were euthanized 15 days post-infection, and lungs were collected. The organs were fixed in 10% formalin and embedded in paraffin. Five-micrometer sections were stained with Grocott's methenamine silver (GMS) and counterstained with hematoxylin–eosin (H&E). From the histological sections of the lung, the area was determined, and CFU/mm2 were counted. The calculation consists of the total number of fungal cells divided by the lung area [29]. The lung sections were analyzed about the cellular changes, and presence of fungi and inflammatory cells, using a Motic model BA310LED microscope, Moticam 5.0 MP digital camera (100, 400 and 600x magnification) and Motic software. Thus, 20 fields of at least two histological sections were classified according to the presence of inflammatory infiltrates categorized as severe (3+ or more), moderate (2+), mild (1+) and non-inflammatory (0+) [30].
Statistical analysis of the different experimental groups was performed by GraphPad Prism software (GraphPad Software, San Diego, CA, USA). Reduction of fungal burden from in vivo treatment was reported as log10 of mean ± standard deviation using unpaired Student’s T-test. The significance of differences in histopathological score was determined by Student’s T-test. The level of significance was set as p< 0.05.
LMM5 was able to inhibit the growth of all isolates of Paracoccidioides spp.. 77.8% of isolates presented MIC values ranging between 8 and 32 μg/mL (Table 1). The Mg0114 isolate was the most sensitive (MIC = 1 μg/mL). All isolates presented the minimum fungicidal concentration (MFC) values similar to the MIC values. Otherwise, the MIC values of LMM11 was 8 μg/mL for most of the isolates (88.9%). The MFC were 8 and 16 μg/mL, corresponding to 66.7 and 33.3% of the isolates, respectively (Table 1). The susceptibilities of isolates to AmB are shown by MIC values of 2 μg/mL (66.7% of the isolates) and 1 μg/mL (33.3%).
The change in growth over time was evaluated by time-kill curves during 14 days (Fig 2). The yeast treated with LMM5 or LMM11 exhibited 80% reduction in the Pb18 cell viability from the 5th day post-treatment. The fungicidal profile was determined by CFU reductions of ≥3 log10 as compared with control growth. The LMM5 fungicidal profile can be observed from the 7th day (Fig 2A). For LMM11, the fungicidal effect was detected on the 7th day (16 μg/mL) as shown in Fig 2B. In addition, the largest difference between groups was observed on the 14th day, for both compounds.
The time-kill curve results were corroborated by LIVE/DEAD assay, in which the cellular viability was evaluated by fluorescence microscopy. For this evaluation, Pb18 cells were treated with LMM5 (Fig 2C) or LMM11 (Fig 2D). Both compounds were able to produce a diffuse bright green fluorescence profile indicating cell death or yeast with little metabolic activity. This fluorescence profile is quite different from what was observed in the control group (not treated), in which live cells presented yellow-orange intravacuolar structures.
AmB, when combined with LMM5 or LMM11, showed better antifungal activity than alone, reducing the MIC value from 2 to 0.5 μg/mL. The new compounds’ interactions with AmB reduced the three-fold MIC value of LMM5 (from 32 to 4 μg/mL) and the two-fold MIC value of LMM11 (from 16 to 4 μg/mL). These results indicate a synergistic effect of AmB with LMM5 or LMM11 (Table 2). The synergic effect revealed by FIC values was validated by the result of the Bliss independence surface analysis. In this way, the AmB combination with each of the oxadiazoles showed predominance of blue areas, indicating a positive ΔE and thus confirming the synergic capacity (Fig 3).
In vivo toxicity parameters showed mild behavioral changes, such as abdominal contortion and motor impairment, within 30 minutes after intraperitoneal administration of the compounds in all groups evaluated. After this period, no alterations were observed. For both compounds, it is important to note that there were no differences in the body weight of animals, in the hematological profile and in the macroscopic analysis of the organs after 14 days. Regarding the biochemical parameters, although the serum levels of amino transferase aspartate (AST) from mice receiving LMM5 were significantly higher than that from the control (p <0.05), their values were within those expected for normal mice (Table 3). Similarly, the LMM11 group showed no statistical differences in the AST values compared to that from the control group (Fig 4A). Both amino transferase alanine (ALT) and creatinine levels did not present a statistical difference between the groups evaluated (Fig 4B and 4C). According to Fig 4D, the liver did not exhibit changes in its weight, while in the kidney, although there was a slight reduction in the kidney weight of animals receiving LMM5 and LMM11, no statistical difference was found (p >0.05) (Fig 4E). There were no significant changes in heart weight (Fig 4F).
Since LMM5 and LMM11 presented promising in vitro antifungal activity and no toxicity in vivo, the next step was to evaluate them through an in vivo experimental PCM treatment. Daily therapy with the new compounds for 14 days showed a significant reduction of pulmonary fungal burden in relation to the control (p <0.05) for LMM5 (1.2 Log10 CFU/g) and LMM11 (1.0 Log10 CFU/g), as well as for the group treated with ITZ (1.5 Log10 CFU/g) as shown in Fig 5. There was no statistical difference among the groups treated (LMM5, LMM11 and ITZ); all were equally efficient in reducing fungal burden in mice (p>0.05).
Because pulmonary fibrosis is the main sequelae of PCM, even after treatment it is essential to evaluate the therapy effect of the new compounds on the inflammatory response triggered by P. brasiliensis. A quantitative analysis of the histological sections allows determination of the number of fungal cells present in each histological lung section. Fig 6A demonstrates that conventional antifungal treatment with ITZ was as effective as the new compounds in reducing the number of yeast cells/mm2 when compared to a control (p<0.05), but no statistical difference between compounds and ITZ was found (p>0.05). These results corroborate the significant reduction of fungal burden presented previously (Fig 5). The inflammation level, indicated by the presence of inflammatory infiltrates in lung tissue, was significantly reduced in the three treatments tested, in relation to the control (p<0.05) (Fig 6B).
A qualitative analysis of the histological sections of groups treated was performed. The lungs of the infected mice that received only vehicle (DMSO 1% and Pluronic 0.2%) showed a predominance of necrotic areas, indicated by black arrows (Fig 7A). In contrast, animals treated with ITZ, LMM5 or LMM11 revealed large areas of preserved lung tissue, indicated by the blue arrows in Fig 7D, 7G and 7J, respectively. In the necrotic areas, it was possible to observe a total loss of pulmonary architecture, leading to no alveolar wall visualization (black arrows).
Severe lesions, characterized by the presence of diffuse inflammatory exudate, were also observed. An intense recruitment of mononuclear cells was detected in the control group, as indicated by red arrows (Fig 7B). In the treated groups, the presence of inflammatory infiltrates was lower (Fig 7B, 7C and 7I). The histopathological evidence showed rounded and multi-budding fungal cells presenting viable protoplasm (white arrows) and nonviable protoplasm (green arrows) in all groups analyzed (Fig 7C, 7I and 7L). Therefore, these results demonstrated that treatment with both compounds can be associated with infection control and maintenance of pulmonary architecture.
The key to a good prognosis is immediate treatment [32]. However, the limited number of antifungal drug classes, the need for long-term treatment, and the high toxicity and adverse effects of the drugs reduce adherence to the treatment by patients [33]. All these concerns indicate a major gap that must be filled in antifungal therapy, especially for severe mycoses treatment. Undoubtedly, the pace of discovery of new antifungals is far from reaching the current needs. However, some groups have sought new therapeutic options through in silico approaches.
Virtual screening based on the ligand or structure target is being performed in the identification of new compounds for PCM treatment [20, 29, 34]. The thioredoxin reductase (Trr1) has been shown as an important target for the development of molecules with antifungal activity [16, 35]. This protein is a flavoenzyme that catalyzes the reduction of NADPH-dependent thioredoxin, protecting cells against oxidative stress [36]. Thus, Trr1 comprises the three main parameters to be a potent candidate for antifungal development: it is an essential gene for fungus survival, it is absent in humans, and it is conserved in several pathogenic fungi. Therefore, it will possibly allow a broad spectrum of action [16].
LMM5 and LMM11 were selected by virtual screening as potential inhibitors of Trr1. Initial tests showed in vitro activity against various invasive fungal infections (IFIs) and absence of toxicity [17]. This work reports the antifungal effects of these two oxadiazoles against Paracoccidioides spp., presenting an inhibition profile with MIC values between 1 and 32 μg/mL. Several studies have used the in silico approach to identify compounds for PCM treatment [16, 20, 25, 30, 35, 37]. The MIC values for these compounds was always similar to oxadiazoles. Three inhibitors of thioredoxin reductase showed MIC values ranging from 8 to 32 μg/mL [16]. An inhibitor of chorismate synthase presented antifungal activity with MIC values between 2 and 32 μg/mL [20].
Furthermore, two homoserine dehydrogenase inhibitors demonstrated antifungal activity (MIC values 32–64 μg/mL) [25]. This group also synthetized 4-methoxy-naphthalene derivatives with antifungal activity against Paracoccidioides spp. (MIC values 8–32 μg/mL) [38]. A thiosemicarbazone derivative tested against 14 isolates of Paracoccidioides spp. showed MIC values between 3.90 and 62.50 μg/mL [39]. In addition, new chalcone derivatives presented antifungal activity with MIC values between 2.9 and 42.2 μM [34].
Drugs with fungicidal profiles are more promising than fungistatic [40]. Thus, our findings indicate that the two oxadiazoles compounds analyzed in this study are promising, because both presented fungicidal profile, especially after 7th day. Fluorescence microscopy results confirmed this fungicidal profile, resulting in cell death and not only growth inhibition. A thiosemicarbazone derivative of lapachol was tested against isolate Pb18 by de Sa et al. [39], and it showed a reduction of 90% in fungal growth after the 5th day; no synergistic effect with conventional drugs was detected.
The antifungal effect of this commercially available drug combined with the novel compounds was evaluated. The synergistic interaction between candidate compounds and conventional antifungal agents may reduce the need for high doses, minimizing adverse effects and providing beneficial attributes for new therapeutic strategies against PCM [41]. In this way, LMM5 exhibited a strongly synergistic effect with the most potent antifungal for PCM, and LMM11 also interacted synergistically. It suggests a possible interaction pathway with AmB, increasing its fungicidal effect [42]. It is possible to suppose that the pores opened by AmB could facilitate the access of the new compounds to the intracellular target, the thioredoxin system, leading to the synergistic interaction observed. Although chalcone derivatives have low MIC values for several isolates of Paracoccidioides spp., no synergistic effect was observed with AmB or another antifungal that was tested [40].
Whereas AmB is the choice drug in the most severe PCM cases, nephrotoxicity affects more than 80% of patients [43]. Recent findings reveal that hepatobiliary changes during treatment with ITZ in patients with PCM are irreversible even if they are not as frequent compared to AmB [44]. The biochemical parameters of in vivo toxicity assays for the new oxadiazoles were analyzed based on the references values for male Balb/c mice suggested by Araujo and collaborators [31]. Therefore, the AST, ALT and creatinine values of mice treated with high doses of oxadiazoles are within normality patterns. These findings reveal that these compounds do not present nephrotoxicity or hepatotoxicity in the murine model.
An important validation of anti-Paracoccidioides activity is to extrapolate to in vivo analysis. The experimental PCM model showed the ability of the oxadiazoles to reduce the fungal lung burden of infected mice. It is suggested that the intraperitoneal treatment with LMM5 and LMM11 reached the lungs and controlled the fungal burden as well as for itraconazole. Comparable results were found for chalcone derivatives in treatment of the PCM experimental. In which the fungal reduction was similar to itraconazole treatment [45]. Cyclopalladated treatment also demonstrated fungal burden reduction and decrease of the damages caused by infection [46].
The major challenge for patients undergoing PCM treatment is sequelae triggered by aggressive pulmonary inflammatory response, which may lead to loss of function [1]. This work evaluated how much of the lung was preserved with the different treatments. Representative images of the lung histopathology (Fig 7) demonstrated untreated animals with large areas of necrosis, filled with Pb18 yeast cells throughout the tissue. Thus, the ability of LMM5 and LMM11 to reduce fungal burden and inflammatory response in the lungs of mice infected with P. brasiliensis seems to be very promising for controlling pulmonary sequelae in PCM.
In conclusion, we have successfully demonstrated that two new oxadiazoles selected by virtual screening presented promising antifungal activity against Paracoccidioides spp., opening perspectives for implementing alternative PCM therapy strategies. Both in vitro and in vivo results indicated that LMM5 and LMM11 could be used as lead structures to new antifungal compounds, with fungicidal profiles and leading to reduced tissue damage caused by fungal infection.
|
10.1371/journal.pcbi.1005948 | The E2.65A mutation disrupts dynamic binding poses of SB269652 at the dopamine D2 and D3 receptors | The dopamine D2 and D3 receptors (D2R and D3R) are important targets for antipsychotics and for the treatment of drug abuse. SB269652, a bitopic ligand that simultaneously binds both the orthosteric binding site (OBS) and a secondary binding pocket (SBP) in both D2R and D3R, was found to be a negative allosteric modulator. Previous studies identified Glu2.65 in the SBP to be a key determinant of both the affinity of SB269652 and the magnitude of its cooperativity with orthosteric ligands, as the E2.65A mutation decreased both of these parameters. However, the proposed hydrogen bond (H-bond) between Glu2.65 and the indole moiety of SB269652 is not a strong interaction, and a structure activity relationship study of SB269652 indicates that this H-bond may not be the only element that determines its allosteric properties. To understand the structural basis of the observed phenotype of E2.65A, we carried out molecular dynamics simulations with a cumulative length of ~77 μs of D2R and D3R wild-type and their E2.65A mutants bound to SB269652. In combination with Markov state model analysis and by characterizing the equilibria of ligand binding modes in different conditions, we found that in both D2R and D3R, whereas the tetrahydroisoquinoline moiety of SB269652 is stably bound in the OBS, the indole-2-carboxamide moiety is dynamic and only intermittently forms H-bonds with Glu2.65. Our results also indicate that the E2.65A mutation significantly affects the overall shape and size of the SBP, as well as the conformation of the N terminus. Thus, our findings suggest that the key role of Glu2.65 in mediating the allosteric properties of SB269652 extends beyond a direct interaction with SB269652, and provide structural insights for rational design of SB269652 derivatives that may retain its allosteric properties.
| G protein-coupled receptors (GPCRs) are targets of more than 25% of prescription drugs on the market. Due to their critical roles in human physiology, competitive modulation of these receptors has been found to be associated with many undesired side effects. Allosteric modulation holds the promise of retaining normal receptor function and improving selectivity. However, the underlying molecular mechanisms of the allosteric modulation of GPCRs have remained largely uncharted. The dopamine D2-like receptors have been implicated in voluntary movement, reward, sleep, learning, and memory. Based on previous experimental findings, we computationally characterized the binding of a negative allosteric modulator of dopamine D2 and D3 receptors, and revealed the dynamic binding mode of this modulator in a secondary binding pocket (SBP) of the receptors. Our results highlight the key role of a Glu in mediating the allosteric properties of the modulator by shaping the dynamically formed SBP, and shed light on rational design and optimization of allosteric modulators of GPCRs.
| G protein-coupled receptors (GPCRs) represent one of the largest protein families, and regulate a myriad of physiological processes in response to diverse chemical or environmental stimuli [1]. Among this family, members of the dopamine D2-like receptor subgroup (consisting of dopamine D2 receptor (D2R), D3R, and D4R) have been implicated in various physiological functions, including voluntary movement, reward, sleep, learning, and memory [2]. Previous studies have established dopamine D2-like receptors as important therapeutic targets for a variety of neuropsychiatric disorders as well as for the treatment of drug addictions [2, 3]. Over the last two decades, significant efforts have been made towards understanding the structure-function relationships of these receptors [4–6]. Despite this success, the high sequence identity within the subgroup presents a formidable challenge for selective drug development [7].
In recent years, several bitopic ligands that target both the orthosteric binding site (OBS) and a secondary “allosteric” binding site in GPCRs have been developed to achieve subtype specificity, improve binding affinity, and lead to a reduction in the side effects compared to orthosteric ligands [8]. Whereas most bitopic ligands show competitive behavior against other ligands that bind the OBS [8], SB269652, a bitopic ligand for D2R and D3R, has been shown to act as an allosteric modulator at both receptors [9–12]. SB269652 is composed of a tetrahydroisoquinoline (THIQ) and an indole-2-carboxamide moiety, connected by a cyclohexyl linker in trans orientation. Molecular modeling of SB269652 in D2R showed that the THIQ moiety binds in the OBS and forms an ionic interaction with Asp3.32 (superscripts denote Ballesteros-Weinstein numbering [13]), while the indole-2-carboxamide moiety protrudes into a secondary binding pocket (SBP) formed by the extracellular portions of transmembrane segments (TMs) 2 and 7. The pose in the SBP establishes a hydrogen bond (H-bond) between the N atom of the indole-2-carboxamide and Glu2.65 [10]. An N-methyl indole-2-carboxamide derivative of SB269652 that is no longer able to make this interaction displayed competitive behavior [14], consistent with an alteration in the binding of the ligand in the SBP. Derivatives based on the indole-2-carboxamide moiety, N-isopropyl-1H-indole-2-carboxamide and N-butyl-1H-indole-2-carboxamide, were recently found to display allosteric pharmacology in D2R and D3R, respectively [12, 15], which suggest that the SBP near TMs 2 and 7 is indeed an allosteric binding site. In addition, SB269652 was inferred to mediate negative allosteric modulation through a dimer interface of D2R [10]. Mutagenesis experiments implicated Glu2.65 near the proposed TM1 dimer interface of D2R [16] as a key determinant for the activity of SB296652, as replacement of this residue with alanine caused a decrease in both SB269652 affinity and negative cooperativity [10]. Similar disruption by the E2.65A mutation of SB269652 binding affinity was also observed at D3R.
However, as the proposed H-bond between Glu2.65 and the indole moiety of SB269652 is not a strong interaction and the E2.65A mutation did not change the pharmacological profile of SB269652 from allosteric to competitive, the H-bond may not be the only element to determine the allosteric properties [10, 14, 15]. Indeed, our structure activity relationship (SAR) studies also suggested that the size and lipophilicity of the indole-2-carboxamide moiety were also determinants of allosteric pharmacology [14]. Thus, another impact of the E2.65A mutation, such as the potentially altered size and shape of the SBP in response to the mutation, may also contribute to the decrease in affinity and negative cooperativity. In the present study, we carried out extensive molecular dynamics (MD) simulations to characterize differences in the binding modes of SB269652 in D2R or D3R, and the impact of the E2.65A mutation. Our results elucidate important mechanistic details of the role of Glu2.65 in the SBP-mediated change in binding affinity and negative cooperativity.
We carried out comparative MD simulations of four conditions: D2R and D3R wild-type (WT) and their E2.65A mutants bound to SB269652. The D2R models in complex with SB269652 were derived from our previous study [10], whereas the starting poses of SB269652 in our D3R models are similar to those in D2R models (see Methods). The first set of simulations was followed by multiple rounds of additional simulations, in which we collected more trajectories for the under-sampled microstates based on the results of the Markov state model (MSM) analysis [17, 18] (see Methods). In total, we collected 145 MD trajectories with a cumulative length of 76.5 μs (Table 1).
Similar to our previous study [10], in the resulting conformations from our extensive MD simulations, the primary pharmacophore (PP) of SB269652, the THIQ moiety, forms a salt bridge with the carboxyl group of Asp3.32 in both D2R and D3R, a key component of ligand binding to aminergic receptors [7]. The secondary pharmacophore (SP), which consists of an indole-2-carboxamide moiety, attached to the PP through a trans-cyclohexylene linker, shows significant dynamics in all our simulated conditions (Fig 1). To characterize the dynamics of the SP poses, we performed MSM analysis to identify the thermodynamic populations of the SB269652 binding poses and to calculate the kinetics of transitions between these populations. Specifically, we used 12 distances between the nitrogen atoms of SB269652 and the Cβ atoms of selected binding-site residues and 4 intra-ligand measures as the input features (see Methods and S1 Fig). The analysis identified two metastable states (MSs) with similar equilibrium probabilities of 52% and 48% in D2R/WT (shown as orange and green spheres in Fig 1C). In the green MS, SB269652 forms two H-bonds to Glu2.65 with both the indole N4 and amide N3 atoms as we described previously [10] (Fig 2A and 2C). However, in the orange MS, the H-bond between the N4 atom and Glu2.65 is lost as N4 reorients toward the extracellular side (Fig 1C). In addition, whereas N3 is in a similar orientation as in the green MS, it has significantly reduced propensity to form a H-bond with Glu2.65 (Fig 2A and 2C).
In contrast to the observed dynamics of the SP among different MSs, in both green and orange MSs, the PP is stable and the salt-bridge interaction between the charged N1 nitrogen in the PP and the key binding-site residue Asp3.32 remained intact (S2A Fig), suggesting that the strong salt-bridge interaction deters the dynamics of the SP from propagating to the PP, although we have found that the poses of the PP and SP of bivalent ligands can affect each other [10, 20, 21].
For D3R/WT, we found that the two states identified by the MSM analysis are similar to those in D2R/WT, in terms of the orientations of the indole-2-carboxamide moiety of SB269652, relative to Glu2.65. Interestingly, in D3R/WT the orange MS in which the N4 atom of SB269652 faces toward the extracellular side also has a slightly higher equilibrium probability (53%) than the green MS (47%) with the N4 atom interacting with Glu2.65 (Fig 1D). Similar to D2R/WT, the PP is stable in D3R/WT as well, with an intact interaction between the N1 nitrogen and Asp3.32 (S2B Fig).
Although the PP is stable in both D2R and D3R, we observed subtly different poses in the OBS of these two receptors. Comparing the representative poses of SB269652 at D2R and D3R, we noticed different interactions between the THIQ moiety and residues from extracellular loop 2 (EL2). Specifically, the subtle divergence of these two receptors at the interface between EL2 and EL1-TM2 accommodates the cyclohexyl linker of SB269652 slightly differently, and this divergence appears to correlate with drastically different orientations of the conserved Ile at the EL2.52 position (second residue after the conserved disulfide-bonded Cys in EL2): while Ile183EL2.52 in D3R forms a favored hydrophobic interaction with the THIQ moiety in the OBS, Ile184EL2.52 in D2R points upwards and is not in contact with SB269652 (Fig 3 and S1 Table). Such a difference is consistent with the results of our per-residue decompositions of the MM/GBSA binding energy calculations of the representative D2/WT and D3/WT frames, in which IleEL2.52 contributed favorably to binding of SB269652 at D3R but not at D2R.
In addition, we found that SB269652 interacts with Ser1935.42, Ser1945.43, and Ser1975.46 in D2R, while it only interacts with Ser1925.42 in D3R. This is likely due to the divergence in both EL2 and TM5 between D2R and D3R –in addition to the divergent EL2.51 and EL2.53 positions in EL2, TM5 is divergent at position 5.52 (Ile2035.52 in D2R and Gly2025.52 in D3R) near the proline5.50-induced kink (Fig 3).
Previously, it was found that SB269652 had more than 10-fold higher binding affinity at D3R than at D2R, and a chimera mutagenesis study that swapped the D2R and D3R segments identified EL2 and TM5 to be important for the different binding affinities [9]. Thus, our findings of the divergent poses of SB269652 in the OBS of D2R and D3R are highly consistent with these results.
In comparison to D2R/WT, our MSM analysis identified 3 MSs for the D2R/E2.65A condition. One MS of D2R/E2.65A is similar to the green MS of D2R/WT; however, given the absence of the H-bond between the indole-2-carboxamide moiety and Ala2.65, the indole ring of SB269652 in the green MS of D2R/E2.65A tends to be more parallel to the membrane compared to in D2R/WT (Fig 1E). In the dominant new pose of the SP in the D2R/E2.65A condition (magenta MS in Fig 1E, which has an equilibrium probability of 60%), both the amide N3 and indole N4 atoms face toward the extracellular side, but the amide O atom faces the intracellular side, which is rarely observed in D2R/WT (Fig 1C and 1E). Interestingly in D3R/E2.65A, the three MSs we identified (Fig 1F) have significant similarity to those three in the D2R/E2.65A, in terms of the distances of N3 and N4 to Ala2.65 (S3 Fig). Even though the orange MS is the most dominant MS (69%) in D3R/E2.65A instead of the magenta MS in D2R/E2.65A (Fig 1F), in both mutant receptors, N4 of SB269652 has a similar tendency to face away from Ala2.65.
We hypothesized that in addition to the H-bonds between the indole-2-carboxamide moiety of SB269652 and Glu2.65, another key to understanding the significance of the E2.65A mutation on the allosteric action of SB269652 lies in conformational changes resulting from this mutation. Our structural analysis identified marked conformational differences between the D2R/WT and D2R/E2.65A conditions bound with SB269652, in the SBP consisting of TM1e, TM2e, TM3e, and TM7e subsegments (see S2 Table for the division of subsegments [20, 22]). These differences were characterized by a significantly larger TM2e-TM7e distance and a shorter TM1e-TM3e distance in D2R/E2.65A as compared to the D2R/WT condition, demonstrating the altered size and shape of the SBP in the mutant construct (Fig 4). Interestingly, the occupation of the SBP by the SP of SB269652 in D2R/WT increased both TM2e-TM7e and TM1e-TM3e distances (Fig 4C) compared to the D2R/WT condition equilibrated with eticlopride (Fig 4B and 4D), a ligand that predominantly occupies the OBS and does not protrude into the interface between TMs 2 and 7. Thus it appears that the SBP is dynamically formed to accommodate the SP of SB269652, and that the E2.65A mutation ablates the ability of SB269652 to increase the distance of TM1e-TM3e through the interaction of its SP with the SBP.
A similar enlargement of the SBP by SB2696952 was observed for D3R/WT as well (Fig 4C and 4D). Comparing the two D3R conditions bound with SB269652, the E2.65A mutation results in larger TM2e-TM7e and smaller TM1e-TM3e distances, similar to the observations for the D2R (Fig 4C).
In both D2R and D3R, Glu2.65 of TM2e faces Ser7.36 of TM7e, and we found that the disruption of this polar interface by the E2.65A mutation contributes to the larger TM2e-TM7e distances. However, Glu2.65 and Ser7.36 have a significant probability to form a H-bond in D3R/WT (green MS, 54.7±0.6%; orange MS, 51.7±1.8%, for the dataset used in Fig 2) but not in D2R/WT (green MS, 14.3±0.5%; orange MS, 9.0±1.3%). Thus, the TM2e-TM7e distance appears to be larger in both the D2R/WT and D2R/E2.65A conditions than in D3R/WT and D3R/E2.65A (Fig 4C), likely due to the shorter EL1 in D2R, consistent with our previous observations and with differences in the tendencies of this interface to accommodate the SP of the bitopic ligands [23].
The impact of E2.65A on the SBP is associated with altered conformations of N terminus (NT) as well. While the NT always adopts flexible loop conformations in our simulations, our loop clustering analysis (see Methods) indicates clearly distinct equilibria and preferences of the loop conformations in different conditions. For the combined analysis of D2R/WT and D2R/E2.65A, we found that the mutation significantly shifts equilibrium of the NT conformation towards one of the two most populated clusters shown in WT, and has ~70% occupancy for the dominant magenta MS of D2R/E2.65A (Fig 5 and S3 Table). Thus, the NT appears to be more dynamic in D2R/WT and adopts multiple conformations, whereas the E2.65A mutation reduces such dynamics. Similarly, we found the most populated cluster in D3R/E2.65A has a significant higher population and is significantly different from that of D3R/WT (Fig 5, S3 Table). Interestingly, it appears that residues 9–13 in D3R have a significant tendency to form a helical conformation, whereas in D2R residues 20 and 21 are dominantly in a bend conformation (S4 Fig). In all conditions, the NT bends down and forms a lid over the extracellular vestibule bringing some of the residues in direct contact with the SP of the ligand (S1 Table).
Taken together, our data show that the E2.65A mutation alters the shape and size of the SBP in both D2R and D3R, which in turn affects the NT conformation.
In recent years, there have been significant advances in the development of allosteric modulators for GPCRs that have high selectivity and novel modes of action. These modulators may lead to therapeutic agents that have fewer side effects [24–26]. One such an example is SB269652, which acts as a negative allosteric modulator in both D2R and D3R. Whereas our previous studies identified Glu2.65 as a critical residue for allosteric modulation of SB269652 [10], our follow-up SAR study suggests that other elements are also involved in determining the allosteric properties of SB269652 [14]. By carrying out comparative MD simulations in combination with MSM analysis of D2R and D3R WT and their E2.65A mutants, we sought to comprehensively characterize the binding poses and dynamics of SB269652 and the impact of E2.65A mutation on the size and shape of the SBP.
The results of our MD simulations and MSM analysis revealed that in both D2R/WT and D3R/WT, SB269652 has significant probabilities of not forming the H-bonds with Glu2.65, and its SP is in dynamic equilibria between two poses, although they essentially occupy the same space in the SBP in each receptor (Figs 1 and 2). These results suggest that the direct H-bond interactions between the ligand SP and Glu2.65 are not the only factor that governs the allosteric property of SB269652 in either D2R or D3R. Indeed, the mutation E952.65A did not cause a switch from allosteric to competitive pharmacology, but rather caused a decrease in the affinity and negative cooperativity of SB269652 [10]. The dynamic equilibria of the binding poses for the allosteric moiety of SB269652 is likely a common feature shared by the binding of other bitopic ligands in allosteric pockets–in many cases, such pockets of GPCRs are located at peripheral regions, more exposed to the water milieu, and may have more dynamic and flexible properties than the OBS. While unlikely to be revealed by crystallography, such dynamic features can be readily identified and characterized by extensive MD simulations in combination with MSM analysis.
Our results also indicate that the E2.65A mutation significantly alters the dynamic equilibria of the SP of SB269652 in the SBP and results in new poses of the ligand. The new poses (magenta MSs), although not observed in WT for both receptors, are similar to the orange MSs of WT that have the N4 atom facing away from Glu2.65 (Figs 1C and 2A). From the perspective of receptor conformation, we found the substitution of the charged and larger Glu2.65 residue to a smaller Ala residue significantly affects the packing in the SBP, leading to larger TM2e-TM7e and smaller TM1e-TM3e distances in both receptors (Fig 4). Thus, we propose that the combined impact from both the removal of the H-bonds and the altered SBP is responsible for the decrease in affinity and cooperativity of SB269652 observed in the E2.65A mutants. Given the significant role of the size and shape of the SBP in mediating the allosteric properties of SB269652, we can envision that some SB269652 derivatives may have allosteric properties even without the capability of forming H-bonds with Glu2.65, as long as they can induce the necessary conformational changes of the SBP. Such conformational changes may be impaired by the E2.65A mutation irrespective of whether a ligand has the capacity to H-bond with Glu2.65. Of note, our SAR studies reveal that an N-methyl indole-2-carboxamide derivative of SB269652 (MIPS1500) displayed apparently competitive behavior at D2R/WT, but acted as a negative allosteric modulator of dopamine at D2R/E952.65A [10]. While such observations may reflect the inability of this ligand to form a H-bond with Glu2.65 as we originally proposed, the addition of a methyl group also adds bulk to the SP. This may change its orientation within the SBP. Thus, by changing the configuration of the SBP, the E952.65A mutation may change the orientation of the N-methyl indole-2-carboxamide moiety of MIPS1500 within the SBP causing the ‘gain’ of allosteric pharmacology. Indeed, the size of the SP has also been shown to be an important determinant of the allosteric pharmacology of SB269652, as derivatives in which the indole moiety was replaced by a pyrrole or proline moiety display apparently competitive pharmacology at the D2R [14]. Such derivatives retain the ability to form an interaction with Glu2.65 but lack the lipophilicity and bulk of the indole moiety. Interestingly, our recent mutagenesis studies reveal that, in a similar manner to the N-methyl indole moiety, the pyrrole derivative displays allosteric pharmacology at D2R/E952.65A (Draper Joyce et al., manuscript in preparation). Such observations are consistent with the conformation of the SBP, and the influence of the SP on this conformation, being central to the allosteric pharmacology of SB269652.
Our results also show that the altered size and shape of the SBP in E2.65A mutants could bias the NT towards specific conformations, and in the case of D3R, a distinct one from the most populated conformation in the WT. In all conditions, the NT forms a lid over the extracellular vestibule and is in direct interaction with SB269652, suggesting a previously unappreciated role of the NT in ligand binding at D2R and D3R. Indeed, the functional roles of the NT have been documented recently in a few closely related homologs, including the α1D-adrenergic [27], 5-HT2B [28] and μ-opioid receptors [29, 30]. By systematically examining all the available high-resolution crystal structures of class-A GPCRs bound to small compounds, we found 9 structures of 6 receptors showing direct interactions (within 5 Å of the heavy atoms) between NT residues and the small-molecule ligands that at least partially occupy the OBS. Interestingly, many of these small-molecule ligands protrude into the interface between TMs 2 and 7 (S5 Fig).
Taken together, our findings highlight the key role of the size and shape SBP, which is determined by Glu2.65, in mediating the allosteric properties of SB269652, and provide structural insights for the rational design of SB269652 derivatives that may retain these allosteric properties.
The binding mode of SB269652 at D2R was investigated based on our previous study [10]. Briefly, to acquire a reference binding mode of the PP (tetrahydroisoquinoline (THIQ)) of SB269652 in the high-resolution crystal structure of D3R (PDB code 3PBL [31]), THIQ in the protonated form was first docked into the D3R structure with the induced-fit docking (IFD) protocol [32] implemented in Schrödinger suite (release 2016–1, Schrödinger, LLC: New York, NY). The lowest MM/GBSA energy pose from the largest binding mode cluster was selected as a reference pose for the PP of SB269652 at D3R. Assuming that binding modes of THIQ in the near-identical OBSs of D3R and D2R should be similar, we docked the THIQ into the D2R model [20, 21, 23, 31, 33] and selected a pose that is closest to the THIQ reference pose in the D3R structure. The full-length SB269652 was then docked into the D2R and D3R models by restraining the PP core [21] to the respective THIQ reference poses (with RMSD tolerance for the heavy-atom restraints of < 2.0 Å). To investigate the effect of the E2.65A mutation, Glu2.65 was mutated to Ala in representative frames from equilibrated WT trajectories, and the charge of the system was neutralized by removing a Na+ ion from the water milieu.
MD simulations of the receptor–ligand complexes were performed in the explicit water and 1-palmitoyl-2-oleoylphosphatidylcholine (POPC) lipid bilayer environment using Desmond MD System (version 4.5; D. E. Shaw Research, New York, NY) with the CHARMM36 force field [34–37] and TIP3P water model. The ligand parameters were obtained through the GAAMP server [38], with the initial force field based on CGenFF assigned by ParamChem [39]. The system charges were neutralized, and 150 mM NaCl was added. The average size of the simulation systems was ~110000 atoms.
The protein-membrane relaxation was carried out with a protocol modified from that developed by Schrödinger, LLC. Briefly, the initial energy minimization was followed by equilibration with restraints on all protein and ligand heavy atoms in the beginning for 1 ns, then with restraints only on the protein backbone and ligand heavy atoms for 6 ns. For both the equilibrations and the following unrestrained production runs, we used Langevin constant pressure and temperature dynamical system [40] to maintain the pressure at 1 atm and the temperature at 310K, on an anisotropic flexible periodic cell with a constant-ratio constraint applied on the lipid bilayer in the X-Y plane.
For each condition, we collected several rounds of multiple trajectories following the procedure described below.
The MSM analysis was performed using the PyEMMA program (version 2.3.2) [41]. For the input featurizer, we chose the features based on the following considerations to describe the interactions and orientations of SB269652 within the receptor binding sites. The polar and charged interactions between the ligand and protein contribute significantly to ligand binding, while these interactions can be more conveniently defined and characterized by simple geometric measures, compared to hydrophobic and aromatic interactions. For SB269652, the nitrogen atoms are distributed in both the THIQ and indole-2-caboxyamide moieties, so that the dynamics of the entire ligand can be properly characterized using distances between these atoms and protein residues. Therefore, we identified protein residues with their Cβ atoms within 7.0 Å of any of the four nitrogen atoms of the SB269652, and used these Cβ-N distances as input features. To better account for the orientation of the indole-2-caboxyamide moiety of SB269652, two additional intramolecular distances from the N4 atom of indole ring were included in the features—one to the N3 nitrogen and one to the oxygen of the amide bond. Further, we calculated the vectors from the centers of mass of 5- or 6-member rings of the SP to N4, and the projections of these vectors on the axis perpendicular to membrane were included to identify the orientation of the indole ring relative to the plane of the membrane. In total, 16 input features were used (S1 Fig).
The slow linear subspace of the input coordinates was estimated by the time-lagged independent component analysis (TICA) [42, 43] on the combined data set of D2R/WT, D2R/E2.65A, D3R/WT, and D3R/E2.65A conditions, and a dimension reduction was achieved by projecting on the 4 slowest TICA components (which represent 61.7% of cumulative kinetic variance). k-means clustering was then employed to discretize the simulated subspace and 100 microstates (MIs) were obtained. For a range of numbers of MIs (50, 75, 100, 150, 200, 300), we estimated an MSM for each situation and concluded that the 100-MI MSM performs best based on two analysis. We first calculated scores in terms of variational principle [44, 45], using cross-validation [46] as previously described [47]. This analysis showed the variational scores were comparable for small numbers of MI and decreased when the number of MI is larger than 150. In addition, the 100-MI MSM showed better convergence of the implied time scales (ITS) in terms of the lag times (S6–S8 Figs).
The discretized combined data set was then divided into individual simulated conditions to estimate Bayesian MSMs [48]. The Bayesian sampling was used to compute statistical uncertainties of 500 transition matrix samples at each lag time. Convergence of ITS for all MSMs was achieved at 96 ns lag time (S7 Fig), which was used to estimate Bayesian MSM for each condition. The PCCA++ method [49] implemented in PyEMMA was then used to stitch the MIs into metastable states (MSs). In the resulting MSMs, 2 MSs were assigned to D2R/WT and D3R/WT conditions each, whereas 3 MSs were assigned to D2R/E2.65A and D3R/E2.65A conditions each. The identity of common MSs between different conditions was determined based on the number of shared MIs between them. Further structural and kinetic analysis was performed using frames from those MIs that have > 70% probability belonging to their respective MS.
In Fig 1, the transition rate between two MSs is the inverse mean first passage time, which is the expected hitting time of one MS when starting from the other MS. The π value denotes the equilibrium probability of a given MS, which is the probability to be in the MS that remains unchanged in the Markov model as time progresses. The transition rate and equilibrium probability were computed as described previously [50].
The validity of the MSMs was assessed using Chapman-Kolmogorov tests (S9 and S10 Figs) which showed that the MSMs estimated at 96 ns were consistent with the simulation data within the 95% confidence interval computed by 500 bootstrapped samples of trajectories. Generally, the Chapman–Kolmogorov test checks if the MSM models estimated at lag time τ can be used to make predictions for the data at longer times kτ within statistical error, i.e., if Eq (1) can be satisfied:
P(kτ)=Pk(τ)
(1)
where P(τ) is the transition matrix estimated from the data at lag time τ (the Markov model), and P(kτ) is the transition matrix estimated from the same data at longer lag times kτ. In practice, we use P(kτ) and Pk(τ), respectively, to propagate probability starting from one of the metastable states, and measure how much probability ends up in each metastable state [47].
To adequately and efficiently i) explore the conformational space and ii) sample the transitions between MSs, we developed an iterative MD sampling protocol to guide the simulations to i) the less-than-well-sampled regions and ii) the saddle points on the energy surface that are likely in between MSs. Thus, by taking advantage of the MSM analysis after each round of simulations, we correspondingly identified both i) the single-frame microstates (MIs), and ii) the under-sampled MIs (< 10 frames) that are in between MSs (i.e., the MIs having similar equilibrium probabilities to be in two or more MSs), as the starting points for the next round of simulations. For the selected MIs with more than one frames, we select the representative frames from the more advanced stages of MD simulations. The procedure to identify the MIs satisfying the criteria and to select the frame has been automated with an in-house python script. For the MSM-guided simulations, we collected 300 ns for each trajectory. The simulations were considered to have reached convergence until the biggest change in the equilibrium probabilities for the updated MSMs of each condition was < 5% after including data from the new round of the simulations.
For the identifications of ligand contact residues shown in Fig 2, the results are based on 500 Bayesian Markov model samples 3 frames each from those MIs having > 70% probability of belonging to each MS. For each of the MS we identified residues within 5.0 Å of ligand (heavy atom-heavy atom distances) in D2R/WT and D3R/WT conditions, and the means and standard deviations of three sets of such samplings are shown in S1 Table.
For the sub-segment distance calculations shown in Fig 4, the TMs in both D2R and D3R were divided into subsegments (extracellular, middle, and intracellular) as described in [20] (see S2 Table), the results are based on 3000 Bayesian Markov model samples for each condition with the number of samples for each MS proportional to their equilibrium probability, from those MIs having > 70% probability of belonging to each MS.
We performed the clustering analysis of the conformations of the N terminus (NT) using the same dataset extracted for Fig 2 (see above) for each MS in each condition, and combined data sets for one receptor together. The clustering is based on pairwise RMSD of selected NT residues by iteratively excluding the residues with high (> 5.0 Å) root mean squared fluctuation (RMSF). The final clustering results are based on residues 6–20 and 22–30 for D2R and 2–25 for D3R to perform superimpositions and RMSD calculations. The computed RMSD matrix was then subjected to hierarchical clustering using Ward algorithm implemented in SciPy. The number of clusters for each receptor was determined so that the intra-cluster mean pairwise RMSD for each cluster is within 5.0 Å, unless the given cluster has less than 5% of the total frames. The population of each cluster was re-weighted by equilibrium probabilities of the MSs that their members belong to.
The means and standard deviations for the three largest clusters are shown in S3 Table.
|
10.1371/journal.pcbi.1002251 | High Degree of Heterogeneity in Alzheimer's Disease Progression Patterns | There have been several reports on the varying rates of progression among Alzheimer's Disease (AD) patients; however, there has been no quantitative study of the amount of heterogeneity in AD. Obtaining a reliable quantitative measure of AD progression rates and their variances among the patients for each stage of AD is essential for evaluating results of any clinical study. The Global Deterioration Scale (GDS) and Functional Assessment Staging procedure (FAST) characterize seven stages in the course of AD from normal aging to severe dementia. Each GDS/FAST stage has a published mean duration, but the variance is unknown. We use statistical analysis to reconstruct GDS/FAST stage durations in a cohort of 648 AD patients with an average follow-up time of 4.78 years. Calculations for GDS/FAST stages 4–6 reveal that the standard deviations for stage durations are comparable with their mean values, indicating the presence of large variations in the AD progression among patients. Such amount of heterogeneity in the course of progression of AD is consistent with the existence of several sub-groups of AD patients, which differ by their patterns of decline.
| In recent decades, our understanding of Alzheimer's disease (AD) has increased; however, some basic questions still remain unresolved. One of them is: how homogeneous is AD? Is the course of progression more or less the same for most patients, or are there large variations? Our paper studies a large cohort of AD patients which comes from a 23-year-long study, and performs a statistical analysis of progression speed. We quantify the amount of spread in GDS/FAST stage durations (a staging system widely used by clinicians). We arrive at an astonishing conclusion that the mean length of AD stages is comparable with their standard deviation! This means that individual courses of AD progression may differ very much from each other, and from the textbook mean values. This has implications both for clinical trials (how do we assess if a new drug is effective, if the amount of natural spread is so large in untreated patients?), and for our understanding of this disease, which appears to be comprised of sub-diseases with different patterns of decline.
| The temporal progression of Alzheimer's Disease (AD) shows a pattern of high variability, with patients transiting the stages of the disease having time-courses ranging from months to decades [1], [2]. While the biological correlates of this variability have been investigated by many groups [2]–[19], the underlying reasons for such variations remain largely uncertain. One of the challenges posed by a high variability of a temporal disease course is the difficulty in treatment efficiency assessments. For any current and future progression-delaying drug, it is important to be able to establish whether and by how much it delays the deterioration caused by AD. To this end, it is necessary to have a reliable quantification of the heterogeneity of the disease.
Global Deterioration Scale (GDS) was proposed in [20] and allows professionals and caregivers to chart the decline of people with AD. While a number of scales exist, GDS is one of the most widely used instruments to stage the course of AD. It measures cognitive, behavioral and functional impairment of patients. There are a total of 7 GDS stages (from stage 1 corresponding to no impairment to stage 7 corresponding to the most severe AD). In particular, stage 4 (mild AD) is characterized by patients requiring assistance in complex tasks such as handling finances, planning a dinner party etc. In GDS stage 5 (moderate AD) patients require assistance in choosing proper attire. In stage 6 (moderately severe AD) patients require assistance in dressing and bathing, and start experiencing urinary and fecal incontinence. GDS has been shown to correlate with both behavioral measures, and anatomic brain changes [20].
Functional Assessment Staging procedure (FAST) was proposed in Ref. [21], [22]. Based on GDS, this procedure describes a continuum of 16 successive stages and substages from normality to most severe dementia of the AD type. The FAST stages have been enumerated to be concordant with the GDS stages from which they were derived [23], although some differences between the two scales have also been demonstrated [24]. One of the advantages of GDS/FAST staging system is that it allows the assessment and staging of AD in its entire range from normal aging to very severe, end-stage, AD [25].
In the literature, the course of AD as characterized by GDS/FAST staging system has been described in quantitative terms. In particular, the stages are thought to follow in a sequential fashion and are characterized by certain stage durations [26]. For example, stage 4 is thought to last for 2 years, to be followed by stage 5 whose duration is 1.5 years, which in turn is followed by stage 6 (2.5 years).
While this quantification is a useful diagnostic tool, it reflects the average course of the disease and provides no information about possible heterogeneity of AD progression. At the same time, quantifying the variance of GDS/FAST stage durations is essential, as one needs to compare the delay gained by a treatment strategy with the amount of natural variation in stage durations, to be able to judge whether there is significance to any improvements observed. In this paper we investigate the heterogeneity of AD by studying the distribution of GDS/FAST stage durations of AD patients. We ask: how much variability is there in the course of AD, and how well do the average values for GDS/FAST stage durations reflect the disease course of individual patients?
The estimates for the cumulative probability distributions of GDS/FAST stage durations are presented in figure 1. We can see that there is a slight difference between the GDS and FAST scale. This is further illustrated in figure 2 where we show the mean values of the GDS/FAST stage durations together with their standard deviations. In both figures, the values pertaining to the GDS system are plotted in black, and those for FAST staging are represented by gray lines. We can see that for stages 4 and 5, the FAST stage mean durations are slightly shorter than the GDS mean durations, and for stage 6, the FAST stage mean duration is longer than that calculated for the GDS system. We can also see that for stages 4,5 and 6, the estimated mean durations are somewhat longer than those given in [27] (the values from [27] for each GDS/FAST stage are shown by dashed horizontal lines). Despite this fact, we can see that, consistent with the literature, the GDS/FAST stage 5 is the shortest of the three stages, followed by stages 4, and 6.
A striking observation can be made by looking at the calculated values for the standard deviations of the stage durations. In figure 2, the standard deviation values are represented by vertical bars around the mean, and are also shown in brackets next to the calculated means. Both for GDS and FAST staging systems, the standard deviations are relatively large. For example, for the shorter stages 4 and 5, the standard deviations are of the order of the mean values for stage durations, and for the longer stage 6, the standard deviations exceed 50% of the mean stage duration values. Given such large standard deviations of stage length durations, it is remarkable that the calculated mean values of stages 4 and 5 are so close to the previously reported durations; and for stage 6, the calculated means are definitely within a standard deviation from the value in [27]. We further observe that the differences between the GDS and FAST measurements are also well within the standard deviation, so we cannot conclude that the two systems yield different mean values [25].
Analysis of a large longitudinal dataset has revealed a significant degree of variation in the lengths of GDS/FAST stages 4–6 of AD. In particular, the calculated standard deviations for GDS/FAST stage durations turned out to have values similar to their mean durations. This is an indication that the patterns of cognitive and functional decline vary significantly from patient to patient.
The suggestion that AD is a genuinely heterogeneous disease, has been proposed in the literature [28]. One paper [29] studies a 4-year longitudinal dataset, and identifies four different subgroups of AD patients which differ by the rate of their intellectual and functional decline as well as other symptoms. Ref. [30] states that AD shows heterogeneity in its clinical, anatomic, and physiologic characteristics, and identifies several patient subtypes with respect to different characteristics, including the time course of progression. In particular, inhomogeneity is observed with respect to the rates of ventricle enlargement, which are related to rates of cognitive decline. In Ref. [31], the presence of aphasia in AD patients is correlated with a more rapid course of the disease. This is done by performing extensive testing of the patients, as well as interviewing reliable informants, in the course of a 2.5 year-long follow-up. Ref. [2] follows patients for 3 year, and discovers an association between relatively severe frontal lobe involvement and a rapid clinical course of AD, measured by using the dementia rating scale and estimating the symptom duration time. A recent paper, Ref. [32], examines AD data from a 15-year longitudinal study, and provides important insights into the patterns of progression of AD. It identifies three groups of patients based on their initial (pre-progression) rate. This rate is estimated by using the (normalized) Mini Mental Status Exam (MMSE) score at base-line, divided by the symptoms' duration. It is found that the different groups remain separate in the course of the follow-ups, which is consistent with our previous finding [33]. Most relevant to our present study, it is found that the average rates of decline for the three groups are different for three types of measures: a cognitive measure (Alzheimer's disease Assessment Scale-Cognitive Subscale), a functional measure (Physical Self-Maintenance Scale), and a global measure (Clinical Dementia Rating Scale Sum of Boxes). Although no direct estimate of the variation has been presented, these results clearly show that AD progression rates are heterogeneous in many respects.
The patient data used here come from a longitudinal study conveyed between 1983 and 2006. It is theoretically possible that the large variation observed in the cohort of patients is a consequence of a change in lifestyle factors, which affected the course of AD progression. To explore this possibility, we have split the cohort of patients into two subgroups based on their dates of visit, and calculated the statistics of stage durations both for the “earlier” and the “later” parts of the cohort. We found that within the subgroups, the variances of the stage durations were as large as the ones reported here, and further, the mean values of stage durations were not significantly different.
Note, however, that the analysis performed here was not specifically designed to discern slight trends in the disease progression over the decades. We cannot perform such an analysis with the data at hand because of the data scarcity issues (using smaller sub-groups of patients necessarily jeopardizes the reliability of the statistics). More data would be needed to catch the trends related to changes in life-style and other generational effects. Here we could only conclude that in both early and late halves of the cohort, the variances were large, and stage durations were statistically not different.
Given a high variability of progression patterns, an important question is finding variables that correlate with progression rates. We have attempted to relate the rate of progression to demographic factors, and determine if it correlates with age at baseline,sex, education, or the age of onset of AD (which was back-calculated by using the information on the estimated stage durations). No significant correlations with these factors have been found, which is consistent with several previous papers [2], [13]–[19]. In the literature, several factors have been proposed to be predictive of the disease progression rate. The work of [34] highlights the heterogeneity of AD, and shows that clusters of CSF biomarker levels are related to patients' cognitive profiles. In particular, it finds that patients with extremely high CSF levels of tau and tau phosphorylated at threonine 181 demonstrate a distinct cognitive profile with more severe impairment of memory, mental speed, and executive functions; importantly, these differences cannot be explained by disease severity. Paper [35] finds that at the time of diagnosis, a combination of high CSF tau without proportionally elevated p-tau-181 is correlated with a faster rate of cognitive decline in AD patients. In paper [36], the variability of AD is explained in terms of specific types of EEG abnormalities. In paper [37], heterogeneity of AD is related to genetic variation in patients, such as that associated with cerebrospinal fluid phospho-tau levels. It is plausible that a combination of many different factors is responsible for a high variability of AD progression rates.
Our main finding is the large heterogeneity in the duration of GDS/FAST stages in AD, which is consistent with the reports cited above. Our methods however are very different. In this study we use a very extensive (23-year long) longitudinal dataset for AD patients, where there is a representation of patients at GDS/FAST stages 4–7 of AD. We calculate the amount of variance in patients explicitly, and demonstrate a large spread in values of GDS/FAST stage values for stages 4, 5, and 6. There are several applications of our results.
To conclude, we analyzed a longitudinal dataset to extract the mean and the standard deviation for GDS/FAST stage durations for stages 4–6 of AD. Applying similar methodology to larger datasets with more frequent assessments will reveal more accurate results.
In order to calculate the probability distribution of stage durations in AD, we used a longitudinal dataset of AD patients, which is an outcome of a longitudinal study performed between the years 1983 and 2006 [33]. The following information is contained in the dataset: the date of each patient's visit to the Medical Center, current GDS and FAST stage, and some demographic information on each patient (such as gender, age and years of education). The total number of AD patients in the dataset is 1321, of which 648 have repeated records (that is, they were seen more than once). The latter group is the one we considered in this study. The mean number of records per patient is 2.6±0.9; the histogram of the number of records for different patients is presented in figure 3(a). The patients' age at the first visit to the clinic is 73.1±8.7 years (see figure 3(b) for the age-distribution). 66% of the patients are female, and 34% male; the average length of education received by the patients is 13.1±3.4 years.
Extracting accurate estimates for the standard deviations for longitudinal datasets is complicated by the practical realities of how the data is collected. First of all, we only know the current stage at the times of assessments, but we have no information on when each stage actually starts and the next one begins (in other words, the data is left-and right- censored). Further complication comes from the fact that the patients' total observation time (time from first to last visit) was 4.78 ± 2.94 years, see the histogram of figure 3(c). This means that many patients in the cohort were not followed for the entire course of their disease. Table 1 shows a split of all the patients into transition classes, that is, it counts the number of patients first seen in stage i, and last seen in stage j. This quantifies exactly how many patients contribute to the calculations for different stages. It is obvious that the information coming from each individual patient is not nearly sufficient to reconstruct all the FAST/GDS stage durations. A method is required which would allow to combine data from different patients to reconstruct the stage duration distributions for the whole cohort (although the information coming from individual patients is incremental). Finally, another problem is illustrated in figure 3(d), where we present the inter-visit time distribution, which shows how long the patients waited before their next visit to the doctor. We can see that: (1) the distribution has a strong peak around 2 years, and then a weaker mode around 4 years, which tells us that the sampling times are strongly biased (the reason for this shape of the distribution is that the next appointment is usually recommended after two years); and (2) the average inter-visit time, which is 3.03±1.59, is comparable with the approximate average stage duration for FAST stages 4–6, which makes this dataset very “coarse” and not ideally suited for extracting stage time variations.
Analysis of long, multistage disease processes has been addressed in literature in many different context [38]–[40]. Statistical approaches to estimating the mean stage durations from a set of AD patients medical records have centered on a linear regression approach [26], where the mean duration of FAST stages were determined, or the use of statistics such as the Kaplan-Meier estimate [32], [41] to determine the survival times of patients. Unfortunately, the linear regression method does not lend itself to calculating the variances of FAST stage durations (see Text S1). Here we used the methodology developed by [42]–[44] to approximate the probability distribution of stage durations.
We view the beginning and the end of each stage as censored events. For each stage i, for each patient, we identify the latest record when they were diagnosed with a stage prior to i (e.g. stage i-1), and then the earliest record where they were diagnosed with stage i or higher. These two time-points give us the interval of time where stage i began, [XL,XR]. Similarly, the latest record in stage i or lower, together with the earliest record at a stage higher than i, give the time-interval where stage i ended, [ZL,ZR]. Some of the right bounds are set to infinity for the lack of appropriate records. We further make an assumption on the patients' first visit, see Text S1 and also [33]: for patients who come to the doctor's office for the first time, we assume that the date of the visit effectively coincides with the onset of the current stage.
We used the iterative approach developed in [43] to approximate the probability distribution function of stage durations for stages 4, 5 and 6. We did not perform the analysis for stage 3 because the number of records for GDS/FAST stages 3 and lower was very small in the database. For stage 7, we were not able to extract meaningful information on the stage duration because of the absence of data on patients' death. The obtained solutions were further checked against a non-parametric numerical estimate of the cumulative distribution function obtained by a straightforward counting method. The two methods are mathematically different, but they revealed very similar results. Further details of the methodology are given in Text S1.
|
10.1371/journal.ppat.1000707 | Analysis of FOXP3+ Regulatory T Cells That Display Apparent Viral Antigen Specificity during Chronic Hepatitis C Virus Infection | We reported previously that a proportion of natural CD25+ cells isolated from the PBMC of HCV patients can further upregulate CD25 expression in response to HCV peptide stimulation in vitro, and proposed that virus-specific regulatory T cells (Treg) were primed and expanded during the disease. Here we describe epigenetic analysis of the FOXP3 locus in HCV-responsive natural CD25+ cells and show that these cells are not activated conventional T cells expressing FOXP3, but hard-wired Treg with a stable FOXP3 phenotype and function. Of ∼46,000 genes analyzed in genome wide transcription profiling, about 1% were differentially expressed between HCV-responsive Treg, HCV-non-responsive natural CD25+ cells and conventional T cells. Expression profiles, including cell death, activation, proliferation and transcriptional regulation, suggest a survival advantage of HCV-responsive Treg over the other cell populations. Since no Treg-specific activation marker is known, we tested 97 NS3-derived peptides for their ability to elicit CD25 response (assuming it is a surrogate marker), accompanied by high resolution HLA typing of the patients. Some reactive peptides overlapped with previously described effector T cell epitopes. Our data offers new insights into HCV immune evasion and tolerance, and highlights the non-self specific nature of Treg during infection.
| Hepatitis C virus persistently infects ∼3% of the world population, leading to life threatening liver diseases and liver failure. It is not well understood why the human immune system often fails to clear the virus, although it is likely multi-factorial. It is accepted that effector T cells are critical for clearing infections, but their function can be suppressed by the somewhat elusive regulatory T cells. Our hypothesis, supported by new data, is that a proportion of the regulatory T cells are specifically stimulated by the virus and that these cells are a stable cell population. We find evidence that these suppressive cells have a distinct set of genes activated and importantly might have a survival advantage over effector T cells, which helps to explain why natural regulatory T cells may influence the outcome of HCV infection. We propose that the new information provides a better explanation of chronic HCV infection and will let us focus on the key experiments to test the hypothesis and to design better treatments.
| Hepatitis C virus is a small positive sense single stranded RNA virus, which causes persistent infection that leads to cirrhosis, cancer and liver failure. In the acute phase of the infection, the host usually mounts strong CD4+ and CD8+ T cell responses, but this wanes in the next few months during the transition to persistence (reviewed in reference [1]). Typically, in persistently-infected patients, the frequency of HCV-specific IFNγ-producing effector T cells is low (usually <0.3% of PBMC by ELISPOT) and that of IL2-producing cells is even lower [2]. T cells, particularly CD4+ T cells, proliferate poorly in response to HCV antigens [3], although CD8+ T cells proliferate slightly better (Li and Gowans, unpublished data). The reason behind the lack of adequate immunity to HCV in human is not well understood, although it is likely to be multi-factorial [1],[4].
IL-10 producing type 1 regulatory T cells (Tr1) may play a role in HCV persistence [5],[6], and more recently, several groups suggested that natural regulatory T cell (Treg, a different type of suppressor cell to Tr1) may be also important [7],[8],[9],[10]. The frequency of circulating CD4+CD25+ cells (the cell population in which Treg are predominantly contained [11]) in the blood of HCV carriers was higher than in healthy donors and individuals who had resolved the infection [7]. In addition, the percentage of CD4+CD25+ cells within the infected liver was much higher than in the peripheral blood [8]. (A review of this topic was published recently [12]). One basic property of Treg is that, once activated via the T cell receptor (TCR), they suppress a wide range of immune responses in vitro and in vivo in a contact-dependent manner [11]. Sugimoto et al. [13] initially showed that depletion of CD25+ cells enhanced the proliferation of the remaining PBMC, while Cabrera et al. [7] and several other groups [8]–[10] further showed that CD4+CD25+ T cells isolated from patients' PBMC could suppress the virus-specific CD8+ T-cell response, suggesting that this population contains HCV - specific Treg. The suppressor function of CD4+CD25+ T cells in response to polyclonal stimuli was further analysed recently in a longitudinal acute phase HCV cohort [10], and it was found that Treg from patients who progressed to persistence were more suppressive than either those from patients who resolved the infection spontaneously or from uninfected healthy donors. In summary, these studies supported the concept that progression from acute to persistent infection is associated with functional changes in the Treg compartment. It is currently unknown, however, to what extend the total Treg pool in HCV-infected individuals is HCV-specific or how Treg react to viral infection as part of the adaptive immune response.
Our group has previously reported [14] that a proportion of natural CD25+ cells isolated from the PBMC of HCV patients substantially upregulated CD25 expression in response to HCV peptide stimulation in vitro, and we proposed that virus-specific Treg were primed and expanded during the disease. Somewhat disturbingly, the frequency of the hypothetical HCV-specific Tregs far exceeded the well-documented low frequency of IFNγ producing anti-viral effector T cells in chronic infection [1], prompting us to seek more insight to these cells in this study.
When the CFSE-CD25+/CD25− co-culture from patients was stimulated for 5 days with the HCV peptide pool (pp), CD25 expression on the CFSE+ fraction was sustained or up-regulated compared to the non-antigen stimulated control (Figure 1A). This observation is reproducible and statistically significant (p<0.05) (Figure 1B). When healthy donor cells were cultured under the same conditions, the CD25 expression profile in the HCVpp culture was similar to that of the non-antigen control (Figure 1A, right panel and Figure 1B). In healthy donors, the baseline level of CD25 expression was sometimes higher (Figure 1B) compared to HCV patients, but there were no major differences between baseline and HCV pp stimulation. Consistent with the manufacturer's technical datasheet, freshly isolated cells expressed more homogenous and intermediate levels of CD25 (Figure S1). These data supported our previous observation with core and NS5 peptides [14], that a proportion of natural CD25+ cells can sustain and/or up-regulate CD25 expression (now termed CD25+/↑ cells) in the presence of HCV peptides and this phenomenon is likely to be disease specific.
The transcription factor FOXP3 plays a critical role in the development and function of natural Treg, but in humans this molecule is also transiently expressed by activated conventional T cells [15],[16]. We have recently shown that epigenetic DNA modification of an evolutionarily conserved element within the FOXP3 locus, named Treg-specific demethylated region (TSDR), correlates with a stable Treg phenotype [17]. In the current study, we applied this principle to determine whether the CD25+/↑ cells, which were previously shown to express FOXP3 [14], are Treg or activated conventional T cells.
HCVpp stimulated CFSE-CD25+/CD25− co-cultures were FACS sorted on day 5 into 3 fractions (Figure 2A): CD25+/↑ cells (P5, >95% of which are CD4+, Figure S2), CD25low (P6) and conventional T cells (P7). Analysis of DNA purified from the above sorted cells by bisulphate sequencing revealed (Figure 2B, left) a highly demethylated TSDR in the HCV-responsive fraction (CD25+/↑ cells, P5), which suggest that these cells are true Treg with stable FOXP3 expression and function. As expected, the TSDR in the conventional T cell fraction (P7) remained highly methylated. The TSDR in the HCV-non-responsive fraction (CD25low cells, P6) showed various degrees of demethylation, which reflects a mixed population of known or unknown cell types. Some P6 cells expressed FOXP3 (Figure 2B, right), but the proportion varied greatly among patients (from ∼5% to ∼40%, data not shown).
To further understand the putative disease-associated CD25+/↑ Treg, genome-wide transcriptional profiles were generated on RNA isolated from the cells, cultured and sorted as described above (Microarray datasets are deposited in Gene Expression Omnibus under series record GSE16576, and can be reviewed via the following link: http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE16576). The Illumina platform was chosen because it requires only 100ng RNA, and given that cell numbers in P5 (CD25+/↑) and P6 (CD25low, or HCV-non-responsive natural CD25+ cells) were limited, this allowed us to analyse each patient individually without pooling samples and thus permit rigorous statistical analysis. Of ∼46,000 genes (or probe sets) analysed, 307 genes were differentially expressed between P5 and P6, followed by 272 genes differentially expressed between P5 and P7 and 155 genes between P6 and P7 (Figure 3A). Some transcript changes were found in more than one comparison (Figure 3B). This constitutes ∼1% of the entire known transcriptome, while the remaining ∼99% of genes were expressed at similar levels by all three T cell fractions. Table S1 provides the full list of genes that were differentially expressed in P5 compared to P6 or P7 (Table S1-A), and in P6 compared to P7 (Table S1-A). Figure 3C shows selected examples of these genes and demonstrates that the data are highly reproducible.
The key Treg signature genes, such as FOXP3, GITR, CD25, IL7R and CTLA4 were differentially expressed as expected among the 3 fractions (Table 1 and Figure 3C) and provide confidence that the experimental system was able to generate quality data. A number of transcription factors (Table 1 and Table S1) were among the differentially expressed genes. This is not particularly surprising because studies in mice suggested that transcription factors are among the genes regulating or regulated by Foxp3 [18].
Ingenuity Pathway Analysis (Ingenuity Systems, www.ingenuity.com), a literature based online annotation tool, was used to identify the relationships and biological significance of the affected genes (Figure S3 and S4). This is the first study in which the putative HCV-specific Treg (CD25+/↑) were analysed against the putative non-HCV-specific Treg (P6), as well as conventional T cells (P7). Most interestingly, a group of genes (Table 1, Figure 3C and Table S1) that were known to be implicated in T cell survival or proliferation (within the top function, immune response, in Figure S3) were differentially expressed by P5 compared to P6 and/or P7. This includes the up-regulation of BCL2 and BCL2L1 (anti-apoptosis), TNFRSF1B and FLT3LG (promote T cell proliferation and activation), IL7 (T cell survival signal) and IL32 (a cytokine released following T cell activation, reviewed in reference [19]), and the down-regulation of the pro-apoptosis gene BMF. This pattern suggests that cells in P5 are likely to be more activated and perhaps have a survival advantage over cells in P7 and/or P6. Figure S3 summarizes the major networks of interactions between these affected genes.
It is known (reviewed in [11]) that Treg must be activated via their TCR to gain suppressor function, and we applied this principle to test the activation status of CD25+/↑ cells (N = 3). We used CD4+ conventional T cells as control because the CD25+ cells isolated from PBMC were almost exclusively CD4+ (Figure S2). The responder cells were a short term autologous CD8+ T cell line driven by HCVpp. The sorted cells (see Figure 4A for a simple illustration and Figure 2A for technical details) were added to responder cells at a ratio of 1∶2 and cultured for 7 days. CD25+/↑ cells strongly suppressed HCV-specific CD8+ T cell proliferation, as measured by Ki67 expression on the responder cells (as the effector frequency is low in HCV patients we found that the Ki67 assay is more sensitive than 3HTdR incorporation in assays with low proliferating cell numbers). Cells from P6 suppressed to a lesser degree, reflecting that this was a mixed population of various cells of unknown nature, while conventional CD4+ cells had no suppressive activity (Figure 4B). These results were confirmed in studies with cells from two additional patients (data not shown). In addition to suppression, P5 also expressed a higher level of IL32 mRNA than P6 (Table 2, in 3 of 4 patients) and P7 (Table 2, in 4 of 4 patients), analysed by qRT-PCR. The role of IL32 in HCV infection is unknown and requires future investigation. Taken together, P5 at the population level correlated with cytokine production and suppressor function, although at present we do not have a reporter molecule that could independently validate the TCR recognition of HCV antigens at the single cell level, a challenging area that is currently being investigated in our laboratories.
A number of genes related to B cell phenotype and function, such as toll like receptors, CD19, CD72, CD86, BLNK, etc. were up-regulated in P6. Interestingly, the same category of genes was also up-regulated in healthy donor natural Treg compared to conventional T cells (Barry, unpublished data). The implication of this is currently unclear. Genes related to CD8+ T effector cell functions (such as CD8, perforin and granzymes) were upregulated in P7 (Table 1 and Table S1), consistent with the fact that this was the only fraction which contained CD8+ T cells, while the original CD25+ fraction (now P5 and P6) contained mainly CD4+ cells (Figure S2).
The HCV NS3 protein has been proposed as a suitable immunogen for vaccine development [20]. The NS3 peptide array (provided by BEI resources, ATCC) consists of 97 overlapping peptides that cover the length of this protein (Table S2 lists the sequence of each peptide). We tested each of the peptides for their ability to induce CD25+/↑ cells following individual peptide stimulation (N = 8). Our working hypothesis is that such a phenomenon directly or indirectly reflects Treg recognition of HCV antigens.
Comprehensive HLA typing of all common loci including class I (HLA-A, B, C) and class II (HLADRB1) was performed for each patient by DNA-based sequencing methods (Figure 5 and Table S3). We found, as expected, that the HLA diversity amongst individuals was high, which may explain why the reactive peptides were not overtly consistent among patients. While the exact location varied among patients, for a given patient, only a few peptides could induce CD25 up-regulation (Figure 5), which is consistent with our earlier findings with the HCV core protein [14]. Some of the reactive peptides are located close to or overlapped with previously described T cell epitopes (Table S4). The implications of this need to be further investigated.
The mechanisms of the positive responses are unknown but our data suggested that it could be related to the HCV-specific nature of Treg. To test this working hypothesis, we designed a HLA (DRB1*1301)-peptide (WKCLVRLKPTLHGPTPLL, the p92) tetramer, which is, to our knowledge, the only HCV HLA class II -peptide tetramer developed based on non-T-helper responses. Compared to a HLA mismatched control, more tetramer+ cells were detected in the patient with DRB1*1301 (7% in SA67 compared to 1.2% in PH 35 in Figure 6A), suggesting the staining signal is likely true. The control tetramers 0701-p92 (mismatched HLA loaded with the same peptide) and 1301-empty (the correct HLA but loaded with no peptide) showed minimal background staining, further suggesting that the staining is genuine. Importantly (Figure 6B), a high proportion (>60%) of the tetramer+ Treg cells were CD25+, while the vast majority (>90%) of tetramer+ T-helper cells were CD25−, supporting our hypothesis and also implying that the tetramer+ T-helper are likely not functional (given that CD25 is an activation marker for conventional T cells).
Conventional protocols to culture human Treg usually involve long term expansion in the presence of high doses of rhIL2. We have previously described a novel co-culture system [14], which we believe to be more physiological. In this system, PBMC-derived CD25+ cells are labelled with CFSE, mixed with CD25− cells from the same donor and finally stimulated with HCV peptides. This approach, used throughout the current study, allowed us to identify a HCV-specific response within the natural CD25+ cell population by observing their response to HCV antigen with conventional T cells as an internal control. We found that the CD25+ population isolated from PBMC of HCV patients, despite a failure to proliferate (which is consistent with the literature that Treg are hypo-proliferative in vitro), responded to HCV peptide stimulation by sustaining and/or up-regulating CD25 surface expression, a phenomenon that does not occur, or at least to a lesser degree, in healthy donors. It is not known if human Treg can down regulate CD25 expression in vitro in the absence of antigen, but we think this can not be excluded.
In naïve inbred pathogen-free mice, CD25+ cells isolated from PBL are almost entirely Foxp3+ natural Treg, but in adult humans, the CD25 expression level is more heterogeneous, as this population is expected to contain activated effector T cells and other known or unknown cell types, particularly during infection. The transient expression of FOXP3 by activated human conventional T cells [15],[16] further complicates the interpretation of human data. We found that natural Treg and Treg converted in vivo under tolerogenic conditions [21] exhibited a completely demethylated TSDR, whereas activated conventional T cells and TGF-β induced Treg contained almost 100% methylated CpG motifs. We therefore proposed the TSDR methylation status as a reliable criterion for the identification of natural and stable subsets of induced Tregs [17]. Using the same criteria, we confirm here that the CD25+/↑ cells in our culture are not activated conventional T cells or TGF-β converted unstable Treg, but are “hard-wired” stable Treg. Since the origin of human Treg is unclear [22],[23], CD25+/↑ cells could either belong to the natural Treg lineage, or be converted from peripheral HCV-specific conventional T cells during the infection, but if it is conversion, the conversion is thorough, as demonstrated by the epigenetic imprint. More Treg were found in HCV-infected liver than periphery blood [24], where a surprisingly high proportion (∼80%) of T cells expressed FOXP3. In vivo expansion of HCV-specific Treg is possible, as Treg from a HCV-experienced chimpanzee had a lower TCR excision circle content compared to naïve animals [25]. The induction and expansion of HCV-specific Treg could have profound effects on the quantity and quality of the anti-viral effector T cell responses.
We next generated gene expression profiles of CD25+/↑ cells (P5), using CD25low (P6) and conventional T cells (P7) as controls, to understand the molecular program that governs the role of these cells. In addition to typical Treg gene patterns, which are either consistent with our FACS data or with the literature, P5 also expressed genes patterns that are less known, such as the survival profile. In an independent study (Barry, et al, unpublished data) we generated transcriptional profiles for ex vivo isolated (FACS sorted CD25high cells) resting, as well as polyclonal stimulated Treg and conventional T cells from healthy donors. Comparing our current dataset to the healthy donor dataset provides hints as to transcriptional changes which could be unique in HCV patients and thus likely to be associated with HCV infection. BCL2, BMF, IL7, IL32, CISH, CCL5, CCR7, IFNαR2, IRF4 and IRF8 (Table 1 and Table S1) are all among this “unique” list, and these genes are known to be critical in regulating cell survival or play important roles in immune responses against pathogens. Development of these data is necessary and is currently ongoing in our laboratories. It was recently reported that the gene profile of ex vivo isolated total Treg from HCV patients was very similar to that of healthy donors [26], as only 5 genes were differentially expressed between the two and the change ranged from 0.4 to 2. Interestingly, none of these 5 genes was identified in our experiments. We think that Treg and non-Treg compartments are both likely to be affected by the disease, a detail which would not be revealed by comparing total Treg of patients and healthy donors.
The continued expression and/or up-regulation of CD25 on a proportion of Treg in response to HCV peptide stimulation in vitro is an event associated with HCV infection, because it does not occur, or is greatly reduced, in healthy donors. This could be a consequence of TCR engagement by the HCV antigen in the context of the peptide/HLA complex, a view supported by the suppression assay data, or alternatively, IL2 (and/or other soluble factors) produced by effector T cells within the co-culture may affect CD25 expression on Treg independently of antigen recognition. In the latter scenario, the apparent antigen specificity of Treg is likely to reflect the antigen specificity of the effector T cells. However, the effector frequency within PBMC was very low, as suggested by the literature (reviewed in reference 2). Supernatant IL10 and IFN-γ levels (measured using Cytokine Bead Array, BD Biosciences) also did not consistently correlate with culture conditions viz. the CFSE-CD25+/CD25− co-cultures and the CD25− PBMC cultures with or without antigen, from patients or from healthy donors (data not shown), and IL2 was generally below the detection limit (data not shown). This is consistent with our microarray data, as none of the key gene signatures for Th1/Th2, Th3 and Th17 (IL2, IL4, IL10, IFNγ, IL12p70, IL17, TGFβ, IL6, etc.) were differentially expressed among the fractions upon HCV antigen stimulation. Thus it is unlikely that the common soluble factors produced by conventional T cells or other antigen non-specific cells in culture could determine the apparent Treg responsiveness.
Ideally we should use a Treg-specific activation marker for epitope mapping, but since there is no such marker we used CD25 as a surrogate marker. In almost every patient, the most reactive NS3 peptide induced higher CD25 expression on Treg compared to anti-CD3 (Figure 5B). Given that anti-CD3 induced more conventional T cells to express CD25 than any of the peptides (Figure 5B and data not shown), these data support the concept that soluble factors alone do not completely correlate with the magnitude of the Treg response, as the level of IL2 in the anti-CD3 culture must be otherwise sufficient to achieve the highest CD25 expression. We attempted to match the reactive peptides against published data on effector T cell epitopes, but found this difficult, as studies using class II tetramers only focus on a few epitope/DR pairs, while in studies which did not use tetramers the HLA typing data were incomplete or missing. Further validation of the putative Treg epitopes and their HLA restriction are required, but nevertheless, our data show that the breadth of the reactivity is rather narrow, while the response itself is robust.
Due to the lack of any Treg specific surface marker and a simple functional readout for these cells, it has not been possible to develop tetramers that are restricted to Treg. Using two class II HLA tetramers previously developed based on T-helper responses, Heeg et al [27] detected FOXP3+ cells during acute infection and reported that the frequency of tetramer+FOXP3+ cells was low and did not correlate with disease progress or outcome. It is unclear at present how this reflects a global picture of Treg/Teff balance, as it is not known to what extent the Treg repertoire overlaps with that of Teff, or if Treg and Teff clones of the same antigen specificity would expand/contact with the same kinetics. Unfortunately, our tetramer data is limited at present and could not answer these questions. Further studies are required, but since it is impossible to develop tetramer for every T cell epitope, we believe that it is important to develop a higher throughput or a more practical Treg antigen specificity readout so that a more global picture can be obtained.
This study opens a window to explore the role of Treg and their target antigens in a chronic viral infection of humans. The target antigens recognised by the FOXP3+ lineage in humans are largely unknown and systems to guide the discovery of these antigens would benefit future studies in HCV vaccines and immunotherapy.
The study was approved by the Alfred Hospital Ethics Committee and the Victorian Department of Human Services Human Research Ethics Committee. Written informed consent was obtained from each subject.
HCV-infected participants (N = 31) were recruited from the Alfred Hospital, Melbourne and from an ongoing study of hepatitis C virus in the social networks of injecting drug users. All participants were HCV mono-infected, with either genotype 1a or genotype 3a viruses, and one participant resolved the infection spontaneously. A few patients were treated previously (unsuccessfully) with interferon/ribavirin and the remainder were untreated. Healthy donors were represented by local volunteers or blood donors from the Australian Red Cross Blood Transfusion Service, Melbourne Branch.
The HCV peptide array, which contains 18-mer peptides overlapping by 11aa covering the entire HCV polyprotein, for genotype 1a and 3a were provided by BEI Resources, ATCC. A peptide pool (pp) working stock (containing 100 µg/ml of each peptide) was prepared in DMSO/RPMI. The final concentration of HCVpp within the culture was 0.2 µg/ml in initial experiments and 0.15 µg/ml for subsequent experiments, or as indicated.
PBMC from patients or healthy donor controls were separated by Ficoll Paque centrifugation and CD25+ cells were isolated from PBMC using CD25 microbeads (MiltenyiBiotec) according to the manufacturer's instructions. The CD25+ cells, typically 1–2% of total PBMC, were labelled with CFSE (Sigma-Aldrich) and mixed back with unlabeled CD25-depleted PBMC at a ratio 1∶10. The CFSE-CD25+/CD25− co-culture was stimulated with or without genotype matched HCVpp in RPMI-1640, 2 mM L-glutamine, 100 IU/mL penicillin-streptomycin (Invitrogen) and 5% human AB serum (MP Biosciences) in 24-well tissue culture plates (Interpath, Australia). Cells were harvested on day 5 for flow cytometry analysis or sorting. In some experiments, culture supernatants were collected prior to cell harvesting for cytokine analysis at later stage.
In general, fluorescent dye-conjugated antibodies and isotype controls were purchased from BD Biosciences. PE-conjugated anti-human FOXP3, isotype control and FOXP3 staining buffer set were purchased from eBiosciences. Intra-nuclear staining of FOXP3, as well as Ki67, was performed according to the manufacturer's instructions.
Flow cytometry was performed using a FACScalibur flow cytometer (BD Biosciences,) and Cellquest software. For data analyses, an initial lymphocyte gate was set based on SSC/FSC and additional gates introduced as required. Results are presented as the percentage, or mean fluorescent intensity (MFI) of positively stained cells within the gated population.
Sorting of HCV peptide-stimulated CFSE-CD25+/CD25− co-cultures from HCV patients was performed using a FACSaria located in a PC3 facility. The cultures were sorted on day 5 into 3 fractions as specified, based on their CFSE labelling and CD25 expression. The primary gate was set on lymphocytes based on SSC/FSC and an additional CD3 gate (for methylation analysis and microarray) or CD4 gate (for in vitro suppression assay) was introduced to the CFSE- population to refine the conventional T cell population.
For this series of experiments, we used cells from male patients, as this overcomes the potential X-chromosomal inactivation of one FOXP3 allele, which usually affects the methylation analysis of Treg in females. Genomic DNA was isolated from sorted cells (Figure 2A) using NucleoSpinTissue XS kit (Macherey & Nagel, Düren, Germany) following the protocol for cultured cells. Bisulfite treatment of genomic DNA was performed as described previously [28] TSDR-primers (5′ to 3′ direction) p-TGTTTGGGGGTAGAGGATTT and o-TATCACCCCACCTAAACCAA, amplifying Amp5 [17] were used for bisulphite-specific PCR and sequencing reactions. The primers “p” and “o” produce amplicons based on the +1 strand.
PCR was performed in a final volume of 25 µl containing 1x PCR Buffer, 1U Taq DNA polymerase (Qiagen), 200 µM dNTPs, 12.5pmol each of forward and reverse primers, and 7ng of bisulphite-treated genomic DNA at 95°C for 15 min and 40 cycles of 95°C for 1 min, 55°C for 45 sec and 72°C for 1 min with a final extension step of 10 min at 72°C. PCR products were purified using ExoSAP-IT (USB Corp.) and sequenced using the PCR primers and the ABI Big Dye Terminator v1.1-chemistry (Applied Biosystems) followed by capillary electrophoresis on an ABI 3100 genetic analyzer. AB1 files were interpreted using ESME.
Total RNA from sorted cells (P5 = CD25+CFSE+, P6 = CD25−CFSE+ and P7 = CD3+CFSE−, as illustrated in Figure 2A) was isolated using RNeasy Kit (QIAGEN Australia) according to the manufacturer's instructions. The RNA quality was ascertained by the Agilent Bioanalyser 2100 using the NanoChip protocol.
The microarray experiments were performed, according to the technical manual from Illumina, by the Australia Genome Research Facility. In brief, 100 ng RNA was amplified using the Illumina Total Prep RNA amplification kit (Ambion Cat. No. IL1791) to generate biotinylated cRNA. An aliquot (1.5 µg/30µl) of the labeled cRNA for each sample, prepared in a probe cocktail that included GEX-HYB Hybridization Buffer, was hybridized to an Illumina Sentrix Human-6 Expression BeadChip-v2.0 at 58°C for 16 hours. After hybridization, the chips were washed, coupled with streptavadin-Cy3 and scanned in the Illumina BeadArray Reader. The scanner operating software, BeadStudio, converts the signal on the array into a TXT file for downstream analysis.
Data analysis and visualization were performed using BeadStudio Gene Expression Module v3.3 software (Illumina Inc., San Diego, CA). With Illumina gene expression array, each probe is measured at least 30 times independently on random distributed beads. This large number of technical replicates allows robust estimation of the hybridization intensity and the measurement error for each probe. The signal for each probe or probe set (gene) was averaged and the background (the average signal from the large number of randomly distributed negative control beads) subtracted, and then normalized using quantile algorithms that account for variations between probes and between chips. A detection P value, calculated by comparing the distribution of the transcript signal to that of the negative control signal, was set at ≤0.001 to identify transcripts that were expressed (with a confidence of ≥99.9%) above background. Genes with detection P value≤0.001 in at least one of the three fractions were selected for further analysis. To detect changes in gene expression between samples, the differential P value (Diff Pval) was calculated using the Illumina custom error model, which allows 5% false discovery rate being automatically adjusted. The cut off for the Diff Pval was set at ≤0.05 (a confidence of ≥95% that the given gene is expressed at different levels between the sample and control).
We used the Ingenuity Pathway Analysis online software (Ingenuity Systems, www.ingenuity.com) to help further group the genes in term of networks and functions.
RNA was isolated from sorted cells as above. Real time RT-PCR assay was performed using Mx3000P QPCR system (Agilent Technologies). The gene expression assays for IL32 and house keep control GAPDH, as well as One-Step Master Mix Reagents, were purchased from Applied Biosystems (Foster City, CA, USA). The cycle conditions are 30 min at 48°C for cDNA synthesis, 10 min at 95°C, followed by 50 cycles of 15 sec 95°C, 60°C 1 min. Data were analysed using MxPro software supplied by the manufacturer.
The co-culture was sorted by FACSAria to CD25+CFSE+ (hypothetical HCV-specific Treg), CD25−CFSE+ (Treg of other specificity and other unavoidable contaminating cells) and CD4+CFSE− (conventional CD4+) in a PC3 facility. The target cells were represented by an autologous HCV-specific CD8+ T cell line, for which an equal number of CD8+ T cells and CD14+ monocytes were mixed and cultured in the presence of 0.15 µg/ml HCVpp for 5 days. The in vitro assay was set up in U-bottom 96-well plates in triplicate. Each well, in a final volume of 200 µl, contained 1×105 sorted cells, 2×105 target cells and 2×104 feeder (autologous immature dendritic cells generated as described previously [29]) and the antigens HCVpp (0.1 ug/ml final of each peptide). At the end of the culture period (day 7), cells were pooled from the triplicate wells, stained for Ki67 expression and analysed by flow cytometry, gating on CD8+ lymphocytes (note that the sorted cells in this experiment were CD4+).
The CFSE-CD25+/CD25− co-cultures were set up essentially as described above, except in a 96 well format, containing 2×105cells in 200 ul medium. Each individual NS3 peptide (Table S1), genotype-matched, was added to each different well at 10 µg/ml final. Anti-CD3 (clone 32-2A2, Mabtech) was used as a positive control at 0.1 µg/ml final. The cultures were harvested on day 5 and analysed for CD25 expression by flow cytometry. The criteria for reactive peptides were described previously [14].
The p92, WKCLVRLKPTLHGPTPLL, is located towards the C terminal of NS3 of HCV genotype 3a (Table S1). PE conjugated HLA class II-peptide tetramer complexes (DRB1*1301-p92, DRB1*0701-p92 and DRB1*1301-empty) were synthesized at the Benaroya Research Institute, USA. For staining, the CFSE-CD25+/CD25− co-culture was harvested at day 5, washed and resuspended in fresh RPMI medium (same as for culture but without HCV peptides) at 1×105 cells in 50 ul per well. To each well 1 ul of a tetramer was added and the cells incubated for 3 h at 37°C, then 30 min at 4°C to stain surface molecules CD25 and CD4.
High-resolution HLA Class I and II typing was performed by direct DNA sequencing methods as previously described [30]. Ambiguities were resolved following sequencing with allele-specific subtyping primers. Sequence electropherograms were analysed using Assign™ (Conexio Genomics). Allele assignment was based upon identity at exons 2 and 3 and consistently allocated for the most common expressed allele in the relevant population.
|
10.1371/journal.pntd.0001535 | Eliminating Rabies in Estonia | The compulsory vaccination of pets, the recommended vaccination of farm animals in grazing areas and the extermination of stray animals did not succeed in eliminating rabies in Estonia because the virus was maintained in two main wildlife reservoirs, foxes and raccoon dogs. These two species became a priority target therefore in order to control rabies. Supported by the European Community, successive oral vaccination (OV) campaigns were conducted twice a year using Rabigen® SAG2 baits, beginning in autumn 2005 in North Estonia. They were then extended to the whole territory from spring 2006. Following the vaccination campaigns, the incidence of rabies cases dramatically decreased, with 266 cases in 2005, 114 in 2006, four in 2007 and three in 2008. Since March 2008, no rabies cases have been detected in Estonia other than three cases reported in summer 2009 and one case in January 2011, all in areas close to the South-Eastern border with Russia. The bait uptake was satisfactory, with tetracycline positivity rates ranging from 85% to 93% in foxes and from 82% to 88% in raccoon dogs. Immunisation rates evaluated by ELISA ranged from 34% to 55% in foxes and from 38% to 55% in raccoon dogs. The rabies situation in Estonia was compared to that of the other two Baltic States, Latvia and Lithuania. Despite regular OV campaigns conducted throughout their territory since 2006, and an improvement in the epidemiological situation, rabies has still not been eradicated in these countries. An analysis of the number of baits distributed and the funding allocated by the European Commission showed that the strategy for rabies control is more cost-effective in Estonia than in Latvia and Lithuania.
| This paper reports the strategy of oral rabies vaccination of wildlife in Estonia, the measures undertaken to check the method's efficacy and the results obtained. Initiated in autumn 2005, oral vaccination programmes resulted in a dramatic decrease in rabies incidence. All the recommended tests were regularly applied, including the systematic testing of vaccine baits prior to release in the field, serological testing and bait uptake assessment in adult and young animals as well as the typing of all rabies virus isolates. The disease was completely controlled by March 2008, with only three cases reported in summer 2009 and one case in January 2011 in areas very close to the South-Eastern border. The costs associated with rabies control have been calculated and compared on a similar basis for the three Baltic countries. The example of rabies control in Estonia shows that rabies can be quickly and successfully eliminated through successive oral vaccination campaigns by strictly following the recommendations of international organisations. These recommendations concern general strategy, vaccination method and choice of vaccine. To our knowledge, this is the first study showing extensive data from a rabies control programme. The underlying strategy, leading to rabies elimination, is advantageous in terms of cost/effectiveness.
| Rabies has been a serious public and animal health issue in Estonia for centuries. Up to 1959, rabies was mainly urban. With the extermination of stray dogs and the compulsory vaccination of pets from 1953, Estonia was rabies free from 1960 to 1967. Sylvatic rabies spread throughout Estonia (including islands) from 1968, there being two main wildlife reservoirs: raccoon dogs (Nyctereutes procyonoides) and red foxes (Vulpes vulpes). By 2002, raccoon dogs had become the major rabies-infected wildlife species in Estonia [1].
From 1947 to 1955, one to eight people died of rabies every year [2]. No human case was registered from 1955 until 1983. The three most recent cases occurred between 1984 and 1986 (one case per year) and were caused by wild infected animals (foxes and roe deer).
The compulsory vaccination of pets and the recommended vaccination of farm animals in grazing areas did not succeed in fully eliminating rabies in Estonia because the virus was maintained in the wild fauna, especially foxes and raccoon dogs.
In May 2004, Estonia joined the European Union (EU) and could then benefit from the financial support of the European Commission for wildlife rabies control.
The natural double mutant SAG2 (avirulent Gif Street Alabama Dufferin strain)—a modified live avirulent rabies virus—was selected for the Estonian wildlife vaccination campaign. The SAG2 strain has been shown to be an effective immunogen when administered orally to red foxes. It contributed to the elimination of rabies in Switzerland [3] and France [4] and has been used in Italy since 2009 with successful results [5]. SAG2 is one of two vaccines recommended by the WHO Expert Consultation on Rabies for oral immunisation of wildlife and dogs [6].
An initial pilot vaccination trial conducted on Vormsi island (92 km2) in spring and autumn 2004 demonstrated the feasibility of orally vaccinating wildlife in Estonia [7]. Three vaccination campaigns with SAG2 baits were conducted in autumn 2005 in North Estonia, and in spring and autumn 2006 throughout the territory. Very encouraging results were achieved in terms of rabies incidence, bait uptake and immunisation in foxes and raccoon dogs [1].
The objective of this study was to assess the rabies situation after oral vaccination (OV) of wildlife from autumn 2005 until the end of 2010 and evaluate the financial aspects. Oral vaccination efficacy was assessed through the incidence of rabies in the country, the proportion of the fox and raccoon dog populations which consumed the bait (revealed through a tetracycline biomarker), and the rabies immunisation rates [6]. The overall cost of the OV campaign in Estonia was reported and compared to that of neighbouring countries.
The SAG2 vaccine (RABIGEN®, Virbac Laboratories, Carros, France) is a modified live attenuated rabies virus vaccine registered in the 27 countries of the EU (European Medicines Agency registration) for oral administration in baits to foxes and raccoon dogs [8]. The SAG2 virus strain was selected from SAD Bern—a sub-clone of a virus isolated from the salivary glands of a rabid dog in 1935—in a two-step process of amino acid mutation [9]. Tetracycline (150 mg per bait) was used as a biological marker to assess bait consumption.
Baits were sent from France to Estonia in refrigerated lorries (−20°C) and stored at −20°C prior to use. Before each vaccination campaign, ten baits from each batch of vaccine to be used (except in 2006 when three baits per batch were titrated) were sent to the European Union Reference Laboratory (EU-RL) in Nancy for titration as previously recommended [10], [11].
All vaccine batches from 2006 on (82 batches) have been titrated. The mean vaccine titre of the different batches used for each vaccination campaign ranged from 107.4 to 108.9 TCID50/dose.
The Republic of Estonia covers 45,227 km2, including ∼25,000 km2 of forests. Estonia is divided into 15 administrative counties, two represented by the islands of Saaremaa (2,673 km2) and Hiiumaa (1,023 km2). The country is bordered by Latvia to the South (339 km), Russia to the East (343 km), the Baltic Sea to the West and the Gulf of Finland to the North.
Wild animals were not vaccinated against rabies in Estonia until 2005, except on Vormsi island (92 km2), where a small-scale OV was carried out in 2004 with a manual distribution of baits. In autumn 2005, the first large-scale oral vaccination campaign of wildlife was conducted in the Northern part of Estonia (25,540 km2) from the Western to the Eastern border, including Estonia's islands. The vaccination area was bordered by a continuous line formed by roads, the shoreline of lake Peipsi and the river Narva. Around 0.5 million baits were spread throughout the vaccination area at a density of 20 baits/km2.
From 2006 to 2010 an OV programme co-financed by the EU and the Estonian state budget was implemented throughout the Estonian territory (42,914 km2 after exclusion of marshlands and urban areas). Both the vaccine (SAG2) and the distribution protocols were similar for the different campaigns. OV was carried out twice a year, in spring (May, early June) and autumn (September, October). Approximately 860,000 baits were dropped during each campaign at a density of 20 baits/km2 using small Cessna 127 fixed-wing aircraft. The baits were distributed manually by trained personnel through special tube systems in the plane. Estonia was divided into vaccination areas covering 240 km2 on average. Vaccines were dropped along parallel flight paths 600 m apart. Flights took place at an altitude of 100–150 m at an average speed of 160–180 km/h. No baits were dropped over urban areas, roads, lakes, rivers, deep swamps and active domestic animal pastures. A GPS system (Garmin 196) was used for navigation and to record flight data. During the campaigns, vaccines were stored at the airport in refrigerated lorries at −20°C.
A public awareness campaign was initiated at the same time as the oral vaccination programme by TV, radio and newspapers.
All rabies cases detected in the vaccinated areas since the beginning of OV in 2005 were sequenced at the EU-RL in Nancy, including the latest case in January 2011.
A cohort of 48 samples from domestic and wild animals found to be FAT positive by the VFL was collected for typing. For the extraction of RNA and hnRT-PCR, 10% (w/v) brain material suspensions were prepared using DMEM medium containing antibiotics and 50% heat inactivated foetal calf serum and centrifuged at 1,500 g for 10 minutes. Viral RNA was extracted from 150 µL eluate using a Qiagen Viral RNA mini kit according to the manufacturer's instructions. First and second round polymerase chain reactions were carried out as previously described [17] giving an amplified product of 589 bp. Host RNA control (18S rRNA, 324 bp) was amplified for each sample using hnRT-PCR as previously described [18].
To verify RNA integrity and validate each negative RT-PCR result, 18S rRNA was amplified for each sample. All the samples were therefore analysed twice: once for host rRNA (18S rRNA) and the second with lyssavirus universal primers [17]. The PCR products (589-bp) were separated by electrophoresis in a 2% agarose gel and purified with a commercial kit (Nucleopsin Extract II columns, Macherey Nagel, France) according to the manufacturer's instructions. Gel purified PCR products were sequenced in both directions by Beckman Coulters Genomics (Takeley, United Kingdom) with the same specific primers used for the nested PCR amplification. The sequences obtained were assembled and edited with Vector NTI1.01 software (Invitrogen, France). After the alignment of sequenced amplified PCR products, 25 identical sequences showing 100% nucleotide identity for N gene (−400 nt) were removed from the phylogenetic analysis.
A phylogenetic tree of classical rabies virus nucleoprotein sequences was constructed using the Neighbour Joining method (p-distance model) with Mega version 5 [19]. A phylogenetic tree was established between 18 N sequences (−400 nt) from Estonia, 20 Eurasian reference sequences [Estonia (n = 1), Latvia (n = 4), Lithuania (n = 4) and Russia (n = 11)], two laboratory strains and two sequences acting as outgroup (Table 1). All the sequences reported in Estonia have been submitted to GenBank. Table 1 summarises sequence findings, detailing the year of isolation, the host species and the geographical origin of all isolates included in the phylogenetic study. The bootstrap probabilities of each node were calculated using 1,000 replicates to assess the robustness of the Neighbour Joining method. Bootstrap values over 70% were regarded as significant for phylogenetic analysis.
The costs of OV in Estonia were evaluated from 2005 to 2010 and detailed according to bait vaccine purchases and their aerial distribution, collection of samples, laboratory tests, awareness campaigns and sundry other costs, such as investigations on suspected animals.
We also wished to compare the cost of OV in Estonia, Lithuania and Latvia. Data have been collected since 2006, when OV was first conducted throughout the territories of each of the Baltic States.
The common basis for comparison was the European Commission's annual financial contribution to the rabies control plans, although the comparison included the number of baits distributed and the vaccine strains used (when published). Data are available at http://ec.europa.eu/food/animal/diseases/eradication/legisl_en.htm. The contaminated and uninfected areas in km2 were evaluated on an annual basis for each country by defining a circle with a 50-km radius around each positive case [10] using mapping software. For each country, the annual data—total uninfected area of year n+1 subtracted from the total uninfected area of year n—were cumulated over the 2006–2010 period. These data assess the area freed from rabies.
Different ratios were calculated i.e. cumulated number of baits distributed over the 2006–2011 period/vaccinated area in the country, cumulated EC funding over the 2006–2010 period/area freed from rabies in the country (which are areas newly uninfected).
The epidemiological situation before the first vaccination campaign in autumn 2005 was reviewed [1].
Briefly, from 1994 to 2005, the number of rabies cases ranged from 74 cases in 1995 to 814 in 2003 (Table 2). The distribution of cases among species has been relatively stable over the years. During the 1968–2005 period, most cases involved wildlife (71–76% of all rabies cases), whereas farm animals accounted for 6%, dogs and cats for 18–23%. From 1968 to 2001, red foxes were the most frequently infected animals, but the number of infected raccoon dogs progressively increased over this same period. Since 2002, the raccoon dog has thus become the main reservoir (47.4% of rabies cases in 2005). Other wildlife, such as badgers, deer, rabbits, hedgehogs, ferrets, squirrels, lynx, minks, weasels, hares, marten and mice, have no epidemiological role in rabies transmission (3% of rabies cases in 2005).
In 2005, rabies cases were evenly spread throughout the country, even on Hiiumaa and Saaremaa islands.
The number of rabies cases has dropped dramatically since 2006. The decrease began following the first countrywide OV campaigns in spring 2006 (Table 2). In 2007, four rabies cases were diagnosed (two cattle, one raccoon dog and one badger), and in 2008, three rabies-positive animals (one dog, one sheep and one fox) were found during the winter-spring period. From March 2008 until the end of November 2011, only four rabies cases were reported: three rabid foxes found in May and July 2009, and one raccoon dog detected in January 2011. All four were within five kilometres of the Estonian–Russian Federation border in the South-East (Figure 1). The case recorded in January was found less than one kilometre from the Russian border, and three kilometres from one of the cases diagnosed in 2009.
Brain samples from target animals collected by hunters for OV efficacy checks (not shown in Figure 1, which presents rabies surveillance data on suspect animals and those found dead) were tested negative for the rabies virus (3,461 samples in 2008; 1,756 in 2009 and 1,750 in 2010).
From 2006 to 2010, some 2,800–3,400 samples in all were tested each year for tetracycline at the VFL.
The proportion of samples showing tetracycline line(s) ranged on average between 84% (2007, 2010) and 90% (2008), indicating that most animals had consumed the baits.
Yearly positivity rates ranged from 85% to 93% in foxes and from 82% to 88% in raccoon dogs. The overall proportion of jaws positive for the biomarker was significantly lower in raccoon dogs than in foxes: 89% in foxes and 84% in raccoon dogs (Chi2 = 67.6, p<0.001) (Table 3).
Annual tetracycline positivity rates ranged from 95% to 98% in adult foxes and from 66% to 88% in fox cubs. In the raccoon dog, the tetracycline positivity rate ranged from 90% to 93% in adults and from 70% to 84% in juveniles. The overall proportion of positive jaw samples was significantly higher in adults than in juveniles for both species, with 96% in adults and 80% in juveniles in the red fox population (Chi2 = 311.2, p<0.001) and 92% in adults and 77% in juveniles in the raccoon dog population (Chi2 = 295.0, p<0.01) (Figure 2).
Approximately 6,400 samples were tested for rabies antibodies at the VFL from 2006 to 2010. Annual immunisation rates ranged from 34% to 55% in foxes, and from 38% to 55% in raccoon dogs, without any significant differences between the two species (Table 3). The overall immunisation rates in the 2006–2010 period were similar in both species, with 46% in foxes and 48% in raccoon dogs (Chi2 = 3.4, non significant).
Annual fox immunisation rates ranged from 40% to 61% in adults and from 17% to 42% in fox cubs. Annual raccoon dog immunisation rates ranged from 51% to 69% in adults and from 13% to 45% in juveniles (Figure 2). Results showed overall immunisation rates significantly higher among adults than juveniles in both species, with 54% in adults compared to 37% in fox cubs (Chi2 = 63.2, p<0.001) and 61% in adults compared to 36% in juvenile raccoon dogs (Chi2 = 220.6, p<0.001).
Of the 48 Estonian samples tested, 43 were positive by hnRT-PCR. The amplified products (589 bp) corresponding to the 43 positive RV strains belong to the lineage formed by the classical rabies virus with a bootstrap value of 85% (value corresponding to the nucleoprotein phylogenetic analysis).
Five samples isolated in 2006 were positive by RT-PCR for the internal control, i.e. the 18S rRNA gene with an amplified product of 324 bp and negative for nucleoprotein gene amplification (589 bp). These samples were therefore investigated in the EU-RL Nancy using rabies reference diagnostic methods. All five samples were found negative by FAT, cell culture and mouse inoculation tests. 17 out of 18 Estonian sequences could be placed in one lineage (bootstrap of 98) belonging to the North East Europe (NEE) group of rabies virus [20]. The latter is mainly composed of the 17 isolate sequences from Estonia, in addition to reference sequences from Latvia (n = 4), Lithuania (n = 4) and from Russia (n = 3), representative of the group E (North-western part of Russia), earlier described by Kuzmin et al. [21] (Figure 3). The group NEE was also linked with 3 sequences of group D, representative of the centre of the European part of Russia [21]. The 17 Estonian sequences exhibited 99% identity among them and 98.4% identity against the 9 published reference European sequences (Estonia, Latvia and Lithuania) belonging to the group NEE. A 98.7% identity was also shown between the 17 Estonian sequences and the 3 Russian sequences forming the group E (RV309, RV245 and RV1596).
The phylogenetic tree also showed that the viral strain (Est-RV2011-DR0359) isolated in 2011 in Estonia is placed in the second group C, representative of Russia already described [21]. This group, constituted by 5 referenced Russian isolates and isolate Est-RV2011-DR0359, formed a solid cluster with a significant bootstrap support value (99%). 98.4% identity was also shown between the 2011's sample and the five Russian isolates (Figure 3).
The analysis of 43 Estonian sequences showed that all the amplified isolates belong to the same clade of classical rabies virus, which clearly differs from that of the rabies virus strain used for orally vaccinating wildlife (SAG2).
A sequence analysis of N (400 nucleotides, positions 71 to 747) and G genes (690 nucleotides, positions 3911 to 4600) comparing isolates TR0608902 (raccoon dog isolated in 2006) and TR0603634 (red fox isolated in 2006) using the Vector NTI and BioEdit software showed a perfect nucleotide similarity with 100% identity (data not shown). The same 100% nucleotide identity was also revealed when the G gene was compared with two other samples isolated in 2006: red fox (TR06-12855) and raccoon dog (TR06-11366) (data not shown).
In autumn 2005, Estonia initiated oral vaccination programmes to control rabies. The organisation and implementation of these rabies vaccination campaigns comply with the recommendations of the European Commission [10].
The efficacy of the OV campaigns in Estonia was assessed by monitoring rabies prevalence in mammals, bait consumption and the immunisation rates in fox and raccoon dog populations. Rabies prevalence was the primary efficacy criterion. Passive surveillance was extended throughout the country by analysing field samples from domestic and wild animals suspected of rabies in addition to mammals found dead [13].
Since the beginning of OV programmes, the incidence of rabies in Estonia has dropped dramatically, with only three to four cases reported annually from 2007 to 2009. Since March 2008, no rabies cases have been detected in Estonia with the exception of three cases reported in summer 2009 and one case in January 2011 in areas close to the South-Eastern border. This last case was identified as a spill-over from Russia and measures have already been adopted to increase rabies awareness in this county. It must be noted that this drop in rabies incidence was observed despite the recent expansion of fox and raccoon dog populations in Estonia (data not shown).
Oral vaccination campaigns were monitored from late July of the OV year to the following March by collecting head and serum samples from foxes and raccoon dogs. Autumn campaigns target both adults and juveniles, while spring campaigns target mainly adults, because fox cubs are usually born from 15 March to 15 April, and raccoon dogs in May [26], [27]. As we expected the animal's age to have an effect on bait uptake and immunisation levels as previously demonstrated [28], we determined the age of all the animals collected from 2007 on.
The proportion of tetracycline-positive samples was high and stable in adults of both target species (≥95% in foxes and ≥90% in raccoon dogs). These results confirmed the efficacy of the aerial distribution strategy and the attractiveness of SAG2 baits for both raccoon dogs and foxes [29], [30]. In juveniles, bait uptake was significantly lower than in adults, ranging from 66% to 88% in fox cubs, and from 70% to 84% in young raccoon dogs. Few studies have evaluated the effectiveness of OV according to the age of the target species [28], [31]. Bruyère et al [28] reported the difficulties in reaching juveniles during baiting campaigns, especially when spring campaigns were conducted in April instead of late May. The tetracycline positivity rates obtained in Estonia are consistent with those reported in foxes during the 1994–2001 period in France with values of 86% after the spring and autumn campaigns in adults versus 63% and 79% after the spring and autumn campaigns in fox cubs [32]. An investigation of diurnal and nocturnal movement patterns of juvenile foxes in the UK suggested that this low bait uptake is explained by their reduced ranging behaviour and the concentration of their activity at secure sites (“rendezvous sites”) [33]. Unless baits are distributed at these secure sites, the probability of vaccinating cubs before the dispersal period is therefore limited. New strategies should be tested to improve the efficiency of OV in cubs, which constitute over half of the target population.
Immunisation rates were similar in both species, ranging in adults from 40% to 61% for foxes and from 51% to 69% for raccoon dogs. In 2007, immunisation rates were significantly lower than those observed in 2006 and from 2008 to 2010. This result may be explained by the earlier dates chosen for the spring 2007 campaign (22 April to 15 May) compared to vaccination periods since 2008 (15 May to early June). This had a clear impact on the immunisation rates observed in young animals after the spring campaign. The rate of immunisation was 45% in adults versus 8% in young animals for samples collected from July to November 2007 after the spring campaign earlier that year. In contrast, 47% of adults versus 43% of young animals had sero-converted after the autumn campaign of 2007 (samples collected from December 2007 to March 2008) (data not shown). From 2008, immunisation rates in adults from both species were close to 60%, except in foxes in 2010. This level corresponds to the vaccination coverage (60%–70%) estimated to be sufficient to break the rabies cycle [34]. Immunisation rates were significantly lower in juveniles, with values ranging from 17% to 42% in fox cubs, and from 13% to 45% in young raccoon dogs. Similar data have also been obtained in France, with lower immunisation rates in fox cubs than in adult foxes [28]. Immunisation rates were significantly lower than tetracycline positivity rates in both adults and young animals, the difference being more marked in juveniles. Several hypotheses may explain these discrepancies. First, the bait casing may be ingested while the capsule containing the vaccine is not. Brochier et al. [35] postulated that cubs may chew the baits without puncturing the vaccine capsule. When they are not hungry, foxes hide their baits to eat later, so the vaccine may be inactivated [36]. Fluorescence in teeth may be seen without any tetracycline absorption or animals may find sources of tetracycline other than vaccine baits: in France, 9.5% of foxes (34/357) were tetracycline-positive in non vaccinated areas (unpublished data). Other hypotheses have been already given [28]. Vixens feed their weaning cubs by regurgitation, and while regurgitated baits still contain tetracycline, the SAG2 strain is likely to be destroyed by gastric acidity. Contact between the vaccine suspension and the oro-pharyngeal mucosa may sometimes be insufficient for immunisation. The production of antibodies may be transient or absent, or reach a low titre. In this last case, the sensitivity of the test used is a critical factor. Recently, Knoop et al. [37] showed that the Bio-Rad immunoassay used for the titration of rabies antibodies has a poor sensitivity using fox field samples from OV areas (32.4% as compared to the Rapid Fluorescent Focus Inhibition Test [RFFIT]). This study showed that antibody titres expressed in EU/mL were 2 to 5 times lower than those obtained with sero-neutralisation assays resulting in a constant underestimation of titres with this immunoassay. These results are consistent with those of a previous investigation conducted at the EU-RL in Nancy [38]. Consequently, multiple-vaccinated adults with high antibody titres may thus be more easily found positive by an ELISA test than primo-vaccinated juveniles.
International organisations also recommend that “all rabies virus isolated should be typed in areas where attenuated rabies virus vaccines are used, in order to distinguish between vaccine and field virus strains” [10]. Phylogenetic analysis demonstrated that all 43 field isolates from Estonia found positive by hnRT-PCR belong to the classical rabies virus (genotype 1) and are all closely related. This shows that no vaccine-induced rabies cases occurred and all positive animals were infected with wild rabies strains present in Estonia. Of the 48 samples detected positive by FAT in Estonia, five samples were shown to be negative by reference techniques and hnRT-PCR at the EU-RL in Nancy. This discrepancy may be due to a degradation of the RNA (storage conditions, transportation to France…).
Our phylogenetic study showed that rabies strains isolated in the 2004–2010 period from the two main wildlife reservoirs in Estonia—the raccoon dog and red fox populations—belong to the same group, NEE. The same phylogenetic results were observed in Lithuania by comparing various isolates (raccoon dogs and red foxes) against published isolates from neighbouring countries [39]. In our study, the comparison of nucleoprotein and glycoprotein gene sequences between two foxes and two raccoon dog samples isolated in 2006 showed 100% identity. The same perfect identity was also shown between one isolate (raccoon dog) from 2006 and three strains (3 red foxes) isolated in 2009 in Estonia. This genetically close relationship between the isolates strongly suggests that the variant circulating in fox and raccoon dog populations have the same origin. As suggested by Bourhy et al. [40], dogs may have served as an early vector for interspecies rabies virus transmission, generating viral lineages that then spread to other taxa. Phylogenetic data suggest the hypothesis of single rabies epidemics in red foxes and raccoon dogs in Northern Europe as previously interpreted [39]. Our phylogenetic study showed that the latest positive case that occurred in 2011 in Estonia belongs to the group C of Russia, suggesting a spill-over from Russia.
The three Baltic States—Estonia, Latvia and Lithuania—share similar features. Rabies incidence had increased at the beginning of the 2000s, in particular in wildlife. The main vectors and virus reservoirs are the red fox and raccoon dog populations [22], [41]. Phylogenetic analyses have shown that rabies viruses from Estonia, Latvia and Lithuania belong to the same clade [22], so it was considered important to compare the evolution of rabies in the three Baltic States.
Since vaccination campaigns were incomplete or absent before 2006 for all three countries, we compared the number of baits distributed and the EC's financial contribution in the three Baltic States since 2006.
Rabies incidence has rapidly decreased since 2006 in the three Baltic States following the implementation of oral vaccination programmes in those countries from 2006. However, while no rabies cases have been recorded in Estonia since summer 2009 (except one case in January 2011 within one kilometre of the Eastern border with Russia), rabies cases were still diagnosed in Latvia and Lithuania in 2010. In France, similar results were obtained when comparing the efficacy of SAD B19 and SAG vaccines in the field: results demonstrated a faster, more durable decrease in rabies incidence when using SAG baits [42]. Despite OV campaigns since 2006 in all infected areas, no area has been successfully freed of rabies in Latvia and Lithuania. It should be noted that rabies incidence differed in the three countries during the period considered for comparison. The period following 2006 was chosen as it corresponded to the start of large scale oral vaccination in all three countries. Lithuania in particular recorded a huge number of cases in 2005 and 2006 then a dramatic drop as early as 2007. Other experiences in Western European countries [4] have shown that the time required to eliminate rabies using oral vaccination campaigns is not correlated with the number of cases in infected areas.
Indeed, the effectiveness of vaccination campaigns relies on three critical elements, i.e. the vaccination strategy, the vaccine and bait used, and the geographical situation of the countries. A tender procedure was used in all three countries for the procurement and aerial distribution of vaccine baits.
All three Baltic States used modified-live vaccines derived from the original SAD (Street Alabama Dufferin) strain isolated in a naturally rabid dog in the U.S.A. The SAD Bern strain is a cell-culture-adapted derivative of SAD. Different vaccines have also been derived from SAD Bern: SAD B19 after attenuation on cloned BHK21 cells, SAG1 and SAG2 vaccines after one or two successive mutations of arginine 333, associated with reduced rabies virus pathogenicity and selected by using monoclonal antibodies [9]. SAG2 was used in Estonia from the very first campaign in autumn 2005 in the Northern part of the country. Both SAD Bern and SAD B19 vaccines have been used in Latvia, SAD B19 being used since 2009. In Lithuania, SAD Bern has been used since 2009 (no information from 2006 to 2008). SAG2 baits have been shown to be stable in the environment, resistant to mechanical forces, water and heat (http://www.ema.europa.eu/docs/en_GB/document_library/EPAR_-_Scientific_Discussion/veterinary/000043/WC500067900.pdf). In contrast, SAD Bern baits have been reported to lack stability in field conditions, with deformation and disintegration of bait casing and loss of vaccine titre depending upon weather conditions such as sunlight and rain [25].
Since 2006, the vaccination campaigns using aerial distribution have been conducted in spring and autumn in the three Baltic States. The bait density per campaign in Estonia and Lithuania was 20 baits/km2, and 23–25 baits/km2 in Latvia [23], [25]. A spatial simulation showed that a higher bait density (over 20 baits/km2) did not reduce the number of under-baited fox groups, had no beneficial effect on the success of OV and wasted resources [43]. The distance between flight paths for the aerial distribution was 600 m in Estonia, as recommended [10], and 1000 m in both Latvia and Lithuania [23], [44]. Although this smaller mesh slightly increased the cost of the flights, the flight line spacing selected in Estonia may increase bait access by the target species. Since autumn 2010, the distance between flight paths was reduced to 500 m in Latvia, and the number of vaccine baits distributed was increased to 26.5 baits/km2 in areas of the country where vaccination was interrupted because of tendering problems [45]. Tetracycline positivity rates are higher in Estonia than in Lithuania or Latvia in both target species. This difference in bait uptake is more marked in raccoon dogs than in foxes. Tetracycline positivity rates in foxes were 93% and 91% in Estonia in 2008 and 2009 respectively, 79% in Lithuania in 2009 [44] and 74% and 75% in Latvia in 2008 and 2009 respectively [23], [46]. Tetracycline positivity rates in raccoon dogs were 88% and 86% in Estonia in 2008 and 2009 respectively, 58.3% in Lithuania in 2009 [44] and 50% and 65% in Latvia in 2008 and 2009 respectively [23], [46]. Bait uptake levels in Estonian fox and raccoon dog populations in 2009 were statistically higher than those obtained in Latvia (no statistical analysis was carried out for Lithuania because the number of animals tested was not available). Apart from the baiting distribution strategy, a lack of bait palatability to raccoon dogs may be critical. Surprisingly, to our knowledge very few data are available regarding both the attractiveness and efficacy of available oral vaccines for the raccoon dog model under experimental conditions. Furthermore, there are no published data on the minimal titres of SAD Bern vaccine required to properly immunise foxes and raccoon dogs. Prior to their use in Estonia, the safety, attractiveness and efficacy of SAG2 baits were demonstrated in caged raccoon dogs according to European guidelines [8]. All animals had seroprotective neutralising antibody titres after ingesting the baits, and all animals vaccinated in this way proved to be protected after a virulent challenge performed six months after OV, whereas all the control animals succumbed to rabies [30]. Since raccoon dogs are one of the two major wildlife vectors and a reservoir in Northern European wildlife, raccoon dogs are a priority target for rabies control strategies in the Baltic States. Efficacy and safety data should be compulsory for all vaccine baits claiming to be suitable for use in that species.
The geographical situation is also an important factor for rabies control, as wild animals ignore borders. The three Baltic States are bordered by rabies-contaminated countries: Estonia with Russia and Latvia; Latvia with Russia, Belarus and Lithuania; and Lithuania with Latvia, Belarus, Poland and Russia (Kaliningrad Oblast). In Poland and the Kaliningrad region, wildlife OV programmes have been conducted throughout the territory since 2002 [47] and 2007 respectively [45]. The disease is endemic in Belarus and Russia, and the EU will provide funding, under certain conditions to be fulfilled by the countries concerned, for cross-border vaccination of Lithuania and Estonia. The land borders for Latvia (1,150 km) and Lithuania (1,273 km) are approximately twice those of Estonia (633 km). The West and North of Estonia is bordered by the sea, while the coastline in Latvia is shorter and in Lithuania even shorter. It should be noted that the South of Estonia has been rabies-free since 2007, demonstrating the importance of cooperation between neighbouring countries during vaccination campaigns. Despite a favourable situation in South Estonia for the past four years in all large areas bordering its 339–km-long Southern border, Latvia still reported cases close to this border with Estonia (one case recorded in 2010 was located approximately 10 km from the border). Natural barriers are obviously important to prevent the reintroduction of rabies from neighbouring countries. However, this factor may explain the presence of isolated cases close to the land borders, as observed in summer 2009 and in winter 2011 in Estonia, but not the persistence of scattered cases observed on Latvian and Lithuanian territory.
Few analyses have been published regarding the cost-effectiveness of OV among wildlife for assessing rabies elimination. A study has estimated the costs associated with rabies epidemics (vaccination of domestic animals, reinforcement of epidemiological networks, support for rabies diagnosis, animal and economic losses, clinical observation of animals which have bitten humans, prophylactic vaccination and post-exposure treatment of humans) with those of oral vaccination campaigns (cost of vaccine baits and their delivery, follow-up to ensure the efficacy of vaccination) [48]. The benefits in terms of cost of wildlife vaccination were obtained after the fourth year of the programme, the highest costs being the preventive vaccination of pets and prevention in humans. In Estonia, an analysis of costs revealed that 62% of costs are channelled into purchasing oral vaccines, whereas costs associated with the vaccination of domestic animals and post-exposure prophylaxis represented 12% of total costs. In 2010, a 35% decrease in costs for parenteral vaccines for domestic animals was reported due to the successful results of rabies control in wildlife. It should be noted that the decrease in vaccination number concerned livestock in particular; furthermore, a new regulation dated July 2009 authorised a booster vaccination every two years instead of annually. The number of post-exposure courses of treatment has also decreased since 2005. We hypothesise for the coming years a continued decrease in costs for most of the expenses required for rabies control (vaccine and distribution material and services, parenteral vaccination of domestic animals, post-exposure prophylaxis and laboratory analysis).
It was also considered worthwhile to compare the cost-effectiveness of OV in the three Baltic States. We selected the 2006–2011 period because before then, no OV campaigns had been regularly conducted throughout these countries. Data were obtained using the same source of published reports (EC funding programmes). Our analyses showed that the mean yearly number of baits used per square kilometre of vaccinated area was the lowest in Estonia and in Lithuania (39 baits/km2), followed by Latvia (45 baits/km2). A higher bait density was used in Latvia. Furthermore, during the 2006–2010 period, a comparison between EC funding and the areas newly uninfected, i.e. the cost required to free an area from rabies, was also the lowest for Estonia (93 euros/km2), followed by Lithuania (122 euros/km2) then Latvia (143 euros/km2). It should be noted that the cost of the SAG2 baits used in Estonia is higher (around 0.83 euros per bait in 2010 for SAG2) than that of SAD Bern/SAD B19 (0.50 euros per SAD Bern/SAD B19 bait based on EU price indications in different programmes). Rabies cases were still scattered throughout Lithuania and Latvia in 2010, whereas rabies incidence dropped quickly and dramatically as early as 2006 in Estonia with fewer baits and without any re-infection of freed areas (except the small part at Eastern border). The Estonian strategy, leading to rabies elimination, is thus clearly more advantageous in terms of cost-effectiveness.
It is costly to keep a country rabies-free when neighbouring countries are still infected. Freuling et al. [49] have evaluated the cost of a vaccination belt to prevent the re-infection of EU countries by infected non-EU countries. Based on two campaigns per year, a bait density of 30 baits/km2 per campaign and a distribution by fixed-wing aircraft with a flight path distance of 500 m, the annual cost would come to 10–16 M euros [50]. Despite the expansion of the red fox and raccoon dog populations [1], Estonia's experience clearly demonstrated that a density of 20 baits/km2 is sufficient to control rabies.
It is generally recommended to perform four vaccination campaigns (i.e. two years of OV) after the last rabies case diagnosis [10], [50]. In Estonia, a buffer zone was established in 2011 (covering a total of 9,325 km2) with Russia and Latvia (Figure 6). The width of the buffer zone between Estonia and Russia was 30 km in areas where natural barriers exist, such as Lake Peipsi and the Narva river, or 50 km in the event of a mainland border. The width of the buffer zone between Latvia and Estonia was 20 km in most cases, but extended to 40 km in areas where rabies cases were diagnosed near the borders. The EU co-financing in Estonia for 2012 will include routine vaccination in buffer-zone (the same than that of 2011) twice a year and also emergency funds for vaccination in the event of the occurrence of residual rabies foci (emergency vaccination in 8,000 km2). OV will be pursued twice a year in these higher-risk areas to maintain a sufficient level of immunity among raccoon dogs and foxes. A bait density of 20 baits/km2 should be sufficient. Previous examples of successful long-term elimination of rabies in Western Europe have demonstrated that rabies control also requires coordination with neighbouring countries [10]. Effective collaboration with Latvia and Lithuania, based on annual meetings on rabies issues and day to day contacts for coordination of vaccination activities, is an important element of the program's success and will be maintained. A partnership with Russia to create an EU-financed vaccination belt between the two countries is being considered. Continuous passive rabies surveillance will be carried out throughout the Estonian territory and will be reinforced along the borders with Russia and Latvia. In well identified high risks areas, active surveillance could be used as a complement to passive surveillance [51]. OV efficiency will be verified only in the buffer zone. Should there be a re-emergence of rabies cases, OV will be rapidly initiated around the site of the event.
Estonia has implemented a rabies control program since autumn 2005. Rabies cases have not been detected for almost 4 years, with the exception of four cases all very close to the South-Eastern border. The example of rabies control in Estonia illustrates how rabies may be quickly and successfully eliminated through successive oral vaccination campaigns among wildlife by strictly following current recommendations from the EC [10], WHO [6] and the OIE [50]. The vaccination strategy and a potent vaccine are key factors to success.
For Estonia to reach a rabies-free status, oral vaccination campaigns among wildlife in buffer zones close to infected neighbouring countries were carried out in 2011 and will be continued until these infected countries also become rabies-free.
|
10.1371/journal.pgen.1002528 | Discovery of a Modified Tetrapolar Sexual Cycle in Cryptococcus amylolentus and the Evolution of MAT in the Cryptococcus Species Complex | Sexual reproduction in fungi is governed by a specialized genomic region called the mating-type locus (MAT). The human fungal pathogenic and basidiomycetous yeast Cryptococcus neoformans has evolved a bipolar mating system (a, α) in which the MAT locus is unusually large (>100 kb) and encodes >20 genes including homeodomain (HD) and pheromone/receptor (P/R) genes. To understand how this unique bipolar mating system evolved, we investigated MAT in the closely related species Tsuchiyaea wingfieldii and Cryptococcus amylolentus and discovered two physically unlinked loci encoding the HD and P/R genes. Interestingly, the HD (B) locus sex-specific region is restricted (∼2 kb) and encodes two linked and divergently oriented homeodomain genes in contrast to the solo HD genes (SXI1α, SXI2a) of C. neoformans and Cryptococcus gattii. The P/R (A) locus contains the pheromone and pheromone receptor genes but has expanded considerably compared to other outgroup species (Cryptococcus heveanensis) and is linked to many of the genes also found in the MAT locus of the pathogenic Cryptococcus species. Our discovery of a heterothallic sexual cycle for C. amylolentus allowed us to establish the biological roles of the sex-determining regions. Matings between two strains of opposite mating-types (A1B1×A2B2) produced dikaryotic hyphae with fused clamp connections, basidia, and basidiospores. Genotyping progeny using markers linked and unlinked to MAT revealed that meiosis and uniparental mitochondrial inheritance occur during the sexual cycle of C. amylolentus. The sexual cycle is tetrapolar and produces fertile progeny of four mating-types (A1B1, A1B2, A2B1, and A2B2), but a high proportion of progeny are infertile, and fertility is biased towards one parental mating-type (A1B1). Our studies reveal insights into the plasticity and transitions in both mechanisms of sex determination (bipolar versus tetrapolar) and sexual reproduction (outcrossing versus inbreeding) with implications for similar evolutionary transitions and processes in fungi, plants, and animals.
| Fungal gene clusters mediate sex determination, natural product synthesis, and metabolic functions. Eukaryotic organisms share features of gene cluster formation including translocations, inversions, gene conversion, and suppressed recombination. The C. neoformans/C. gattii mating-type (MAT) locus spans a single >100 kb gene cluster encoding >20 genes, many involved in sex. We examined MAT gene cluster evolution in model and pathogenic Cryptococcus species. MAT was characterized from two closely related species, T. wingfieldii and C. amylolentus, and is organized into two unlinked gene clusters on different chromosomes. MAT organization in these species provides insight into evolutionary transitions from tetrapolar to bipolar mating systems involving fusion of physically unlinked sex-determinants into one contiguous region. These sex determination transitions occurred concomitantly with the origin of the pathogenic species complex from the last common ancestor shared with tetrapolar non-pathogenic species. We discovered a tetrapolar sexual cycle in C. amylolentus that generates recombinant meiotic progeny, many of which are infertile. Fertile progeny are biased towards one parental mating-type (A1B1) and may be an evolutionary precursor to unisexual mating of the closely related pathogenic species. This study reveals factors orchestrating gene cluster formation and sex chromosome evolution in fungi, including features shared with animals and plants.
| Sexual reproduction is ubiquitous throughout nature, generates population diversity, and has been described extensively in plants, animals, and microorganisms [1]. Sex is both costly and advantageous, and the ubiquity of sexual reproduction suggests that in general its benefits outweigh its costs [2]. In sexually reproducing populations, outbreeding is common, but inbreeding forms of sex also occur that promote clonality. Additionally, unisexual reproduction may be an adaptive virulence strategy for several microbial pathogens [3].
Fungi occur in two mating configurations: bipolar and tetrapolar [4]. In bipolar species, transcription factors that establish mating-type (MAT) are encoded by a single locus; in some examples genes encoding pheromones and their receptors are also present [4]. For mating to occur compatible cells must differ at MAT (a and α), although there are examples of bipolar fungi that also undergo same-sex mating (e.g. Candida albicans and Cryptococcus neoformans [5]). In tetrapolar species, two physically unlinked genomic regions (i.e. MAT loci A and B) control and establish cell identity. These loci are often multiallelic, and alleles must differ at both loci for sexual reproduction to occur. Bipolar mating systems support more efficient inbreeding (50%) and also outbreeding (50%), while tetrapolar systems promote more efficient outbreeding (>99%) and restrict inbreeding (25%) [6]. Ascomycetous yeasts such as Saccharomyces cerevisiae and Candida albicans are bipolar while basidiomycetous yeasts like Tremella mesenterica and Ustilago maydis are typically tetrapolar [7]. In contrast to most basidiomycetous species, Ustilago hordei, Coprinellus disseminatus, C. neoformans, and Cryptococcus gattii have bipolar mating systems [8], [9], [10], [11], [12].
C. neoformans is a haploid, dimorphic fungus that has a bipolar mating system, represented by two alleles, α and a [10]. MAT spans 100 to 120 kb, and encodes more than 20 genes, many of which are involved in mating. Comparison of the MAT gene cluster among the members of the pathogenic Cryptococcus species complex revealed that extensive rearrangements and gene conversions have occurred over time even though recombination in this gene cluster is generally suppressed [13], [14], [15], [16]. The sexual cycle and the structure of MAT in the pathogenic Cryptococcus species have been extensively examined and are well defined [4], [15]. In a laboratory setting, Cryptococcus reproduces via either opposite-sex or unisexual reproduction [5], [11], [12], [14], [17], [18]. Mating (α-a) initiates with cell-cell fusion, followed by production of a filamentous dikaryon with fused clamp cell connections, and culminates in nuclear fusion and meiosis in the basidia [4], [19]. Meiosis produces four haploid nuclei that undergo mitotic division to produce four chains of basidiospores that germinate into fertile yeasts that can mate with a partner/parent of the opposite mating-type. The major differences in α-α unisexual reproduction is that a monokaryon (instead of a dikaryon) forms, mating can involve two genetically distinct isolates (α1-α2) or two genetically identical genomes (α1-α1), and the resulting meiotic spore products are all α.
Fraser et al. proposed that the ancestral form of MAT to the pathogenic Cryptococcus species was tetrapolar, with the homeodomain (HD) and pheromone/receptor (P/R) genes present in two unlinked sex-determining regions [16]. Sequential rounds of gene acquisition led to the expansion of the ancestral tetrapolar MAT loci. In this model, a chromosomal translocation event then fused the unlinked loci into a contiguous region resulting in the formation of a transient tripolar intermediate in which MAT is linked in one partner yet unlinked in the other. This unstable intermediate underwent gene conversion to link the other MAT locus alleles, one or the other homeodomain gene was lost, and MAT was subjected to multiple inversions and gene conversions events to yield the extant bipolar MAT locus of Cryptococcus 16,20.
The pathogenic Cryptococcus species form a monophyletic cluster composed of at least two but possibly as many as six species: C. neoformans var. neoformans, C. neoformans var. grubii, and the sibling species C. gattii (VGI, VGII, VGIII, VGIV) that all have the potential to infect humans and other animals [21]. A recent multi-locus sequence typing (MLST) phylogenetic study resolved the species relationships in this complex [22]. The monophyletic sensu stricto Filobasidiella clade is comprised of the pathogenic species and three closely related saprobic species: Tsuchiyaea wingfieldii, Cryptococcus amylolentus, and Filobasidiella depauperata [22], [23]. The more distantly related sensu lato sister clade Kwoniella encompasses several saprobic and one aquatic-associated species: Bullera dendrophila, Cryptococcus heveanensis, Cryptococcus bestiolae, Cryptococcus dejecticola, and Kwoniella mangroviensis [22].
Of these species that are phylogenetically closely related to the pathogenic Cryptococcus species complex, sex has recently been described for C. heveanensis and K. mangroviensis [24], [25]. Specifically, a heterothallic sexual cycle was observed in these two members of the Kwoniella clade and basidiospores associated with cruciate-septated basidia are produced during mating. Additionally in F. depauperata and T. mesenterica, the nature of sex has also been revealed in previous studies and exemplifies homothallic and heterothallic sexual cycles, respectively [12], [26], [27], [28]. The mating structures of F. depauperata resemble the basidia and basidiospores of C. neoformans and C. gattii while T. mesenterica mating products are similar to C. heveanensis and K. mangroviensis [12], [26], [27], [28], [29]. However, no sexual reproduction had been observed in either C. amylolentus or T. wingfieldii.
A recent study of C. heveanensis revealed it has a tetrapolar mating system, i.e. its sexual reproduction is governed by MAT comprised of two physically unlinked gene clusters: a multiallelic HD locus (B locus) and a P/R locus (A locus) that is at least biallelic [24]. However, it still remains unclear when the bipolar mating system in Cryptococcus pathogenic species first appeared, that is, did it emerge earlier in the common ancestor of the sensu stricto group when it split from the sensu lato group, or did it evolve later and only in the Cryptococcus pathogenic species? Given the close relationship of T. wingfieldii and C. amylolentus to the pathogenic Cryptococcus species complex, understanding their life cycles, as well as their MAT loci configurations can provide key insights into the evolution of MAT and sexual reproduction in C. neoformans and C. gattii.
In this study, we provide a detailed description of the heterothallic sexual cycle of C. amylolentus that we observed under laboratory conditions. Additionally, we characterized the MAT loci of T. wingfieldii and C. amylolentus, and discovered in both species two physically unlinked gene clusters, one encoding the HD locus and the other encoding the P/R locus. Genes within these clusters include many homologs of Cryptococcus MAT-associated genes. Furthermore, our mating assay and genetic analyses of C. amylolentus meiotic progeny showed that many meiotic progeny are sterile and one parental type is overrepresented in the meiotic products, suggesting its tetrapolar mating system deviates from the classic model. We discuss the implications of our findings in the context of the evolution of the mating type locus as well as of bipolar sexuality in the Cryptococcus species complex. Our findings also provide insights into similar evolutionary processes that drive the formation and function of sex chromosomes in algae, fish, insects, and mammals [14].
To determine the structure of MAT in T. wingfieldii, fosmid libraries were constructed from the type strain CBS7118 and probed with several genes within (MYO2, LPD1, and SXI1) or flanking (FAO1 and NOG2) the C. neoformans MAT locus. Positive clones (3F11-3A15-5J15 (P/R locus), 2B23-2K10 (HD locus), and 4E07 (FAO1), see Figure S1) were pooled and sequenced, resulting in the identification of two candidate MAT loci. The FAO1 gene lies on a distinct fosmid and appears to be unlinked or distant from MAT. The region obtained containing the P/R locus spans ∼70 kb and the region obtained containing the HD locus spans ∼40 kb (Figure 1A).
We also cloned and sequenced MAT in the saprobic yeast-like, sibling species C. amylolentus (type strain CBS6039) employing the same approach. Fosmid libraries were generated and probed with MYO2, LPD1, RPL39, and SXI1. Primers specific for MAT genes in T. wingfieldii were used to generate probes for C. amylolentus and the identity of each probe was confirmed via cloning and sequencing. Positive clones (4E01 (SXI1), 4E22 (MYO2), and 3H19 (LPD1), see Figure S2) were individually sequenced and assembled into two MAT loci. An additional fosmid (3N14 (RPL39), see Figure S2) was later identified and sequenced via primer walking. The regions that were sequenced span ∼20 kb and ∼60 kb respectively, and each contain two small sequence gaps (Figure 2A and Figures S3 and S4). The linear order of the fragments in the P/R locus was determined based on Southern blotting. Specifically, genomic DNA from CBS6039 and CBS6273 was digested with five restriction enzymes (BamHI, BglI, ClaI, EcoRI, and NcoI) and Southern blot analysis was performed with probes hybridizing to the ends of each contig in the P/R assembly of C. amylolentus. The gene content in MAT appears to be largely conserved between T. wingfieldii and C. amylolentus. However, our Southern blot analysis indicated at least two major inversions exist between the P/R regions of these two species (Figure 3).
Analysis of the MAT sequences obtained from T. wingfieldii and C. amylolentus revealed that the gene content of these regions are similar to the C. neoformans and C. gattii MAT alleles [10], [16]. In both sibling species, orthologs of both SXI1 and SXI2 are present in the HD locus implicating this as the ancestral configuration (Figure 1 and Figure 2). The orientation of the homeodomain transcription factors mirrors the organization of the paired, divergently transcribed genes, bE and bW, in the tetrapolar basidiomycete U. maydis [29], [30]. In contrast, in C. neoformans and C. gattii, only one HD gene is present and SXI1α is specific to the α allele while SXI2a is specific to the a allele. The region corresponding to the P/R locus contains the mating pheromone genes, the pheromone receptor gene STE3, and the five genes that were hypothesized to be those most recently acquired by the Cryptococcus MAT locus (LPD1, RPO41, BSP2, CID1, and GEF1). In T. wingfieldii, three pheromone genes (MFa1 and MFa3 are identical while MFa2 differs in only one amino acid) are present and share greater identity with the MFa genes of C. gattii with an identity of 80% compared to 70–75% shared with the MFα pheromone gene (Figure 4A). The P/R region in C. amylolentus differs from C. heveanensis in that the pheromone genes are located >30 kb away from STE3 whereas in C. heveanensis these genes are closely linked [24]. Moreover LPD1, STE11, ZNF1, and IKS1 are not within the P/R locus of C. heveanensis [24], while the P/R region is more extensive in C. amylolentus and spans >60 kb (Figure 2 and Figure S2). In C. amylolentus, two pheromone genes (MFa1 and MFa2 differ in only two amino acids) have been identified and share 73% identity with the MFa protein product of C. neoformans and 65–70% identity with the MFα pheromone gene. In summary, both SXI1 and SXI2 were present in the ancestral HD locus of the sensu stricto Filobasidiella clade. Thus, loss of one or the other HD gene occurred during the evolution of MAT in the pathogenic Cryptococcus species. Additionally, the five genes most recently acquired by the Cryptococcus MAT locus are linked to the ancestral P/R locus and thus appear to have been acquired into the expanding MAT A locus rather than entrapped by the MAT fusion event, in contrast to an earlier evolutionary model, suggesting a revision to the model (Figure 5) [16].
In T. wingfieldii and C. amylolentus, the FCY1 and UAP1 genes flank the 5′ end of the MAT HD locus, similar to C. neoformans/C. gattii, but FAO1 is unlinked and present elsewhere in the genome. We observed that STE11 is not present in the P/R locus but, based on PCR analysis, it is located elsewhere in the genome in both T. wingfieldii and C. amylolentus (data not shown). In the MAT locus of the pathogenic Cryptococcus species, STE11 is present. In C. heveanensis, STE11 is linked to but distant from the P/R locus and this may represent the ancestral configuration with retention in C. neoformans and C. gattii and translocation out of MAT in C. amylolentus and T. wingfieldii [24]. In T. wingfieldii, the flanking gene at the 3′ end of MAT, NOG2, was used as a probe. It was present in a single contig within the larger fosmid assembly of T. wingfieldii, but has not been linked to either the HD or P/R loci contigs. PCR analysis (using gap closure) revealed that LPD1 is linked to the P/R locus, although this gap remains to be sequenced. Interestingly, NCP1 and NCP2 are duplicated genes in T. wingfieldii and C. amylolentus but not in the pathogenic Cryptococcus species. The NCP1/2 genes are also duplicated in C. heveanensis [24], suggesting this configuration might be ancestral.
We also identified several hypothetical genes (CND06020, CND06030, CND06040, CND01650, CNBE0480, CNE02690, and CNE02670) with C. neoformans genes as the most closely related homolog in other sequenced fungal genomes. Four of these genes reside on chromosome 4 and two on chromosome 5 of C. neoformans, indicating that translocation (intra- and inter-chromosomal) events may have occurred between these two chromosomes during the evolution of MAT in the pathogenic Cryptococcus species [16], [28]. In C. heveanensis, F. depauperata, and T. mesenterica there is additional evidence for similar exchanges between chromosomes [14], [24], [28].
A considerable level of synteny exists across both MAT loci in T. wingfieldii and C. amylolentus, but we also observed at least two major inversion events that have occurred between the two genomes (highlighted in blue in the P/R locus, Figure 3). Comparison of each sibling species to the C. neoformans serotype D strain JEC21 revealed extensive gene rearrangements and inversions present throughout MAT (Figure S5), similar to the comparisons of MAT within the C. neoformans/C. gattii species complex. The arrangement of the MAT loci in T. wingfieldii and C. amylolentus corresponds to an evolutionary intermediate in MAT evolution in which the loci (or their linked gene repertoire) have expanded but not yet fused.
Analysis using pulsed-field gel electrophoresis and Southern hybridization demonstrated that the HD and P/R loci are physically unlinked in T. wingfieldii, as well as in both strains of C. amylolentus (CBS6039 and CBS6273). Each genome has approximately 10–12 chromosomes ranging in size from 800 kb to 2.2 Mb. Three genes were used to probe the T. wingfieldii chromosomes, two from the HD locus, SXI1 and RPL22, and one from the P/R locus, MYO2 (Figure 1B). For C. amylolentus, a total of three genes were used as probes: one from the HD locus, SXI1, and two from the P/R locus, MYO2 and ETF1 (Figure 2B). From the chromoblot analysis, the two loci are located on separate chromosomes (∼1.1 and 1.15 Mb) in both of the sibling species. That the HD and P/R loci are located on different chromosomes suggests a tetrapolar mating configuration for both sensu stricto species T. wingfieldii and C. amylolentus. Moreover, given the finding that other more distant outgroup species (C. heveanensis, T. mesenterica) are also tetrapolar [24], the most parsimonious interpretation is that the tetrapolar configuration represents the ancestral form of MAT and the bipolar state observed for the pathogenic Cryptococcus species therefore arose even more recently than revealed by previous studies of the more distantly related sensu lato species C. heveanensis [24]. Thus, the organization of MAT in the sibling species resembles key aspects of the proposed intermediates in the evolution of bipolar MAT in the pathogenic Cryptococcus species from a tetrapolar ancestor.
MAT is defined as a gene cluster (containing either HD and/or P/R genes) whose sequence is divergent between two strains of opposite mating-types. Based on the characterized structure of MAT in both species, we sought to determine which genes in each region govern and control sexual identity. The lack of additional T. wingfieldii strains has made it difficult to assess experimentally whether it has a sexual cycle and, if so, which genes are involved. Fortunately, in C. amylolentus two strains are available and this enabled our analysis of MAT and sex in this species resulting in the discovery of an extant sexual cycle (described below).
Regions that define MAT typically display polymorphisms when comparing sequences from strains of opposite mating-type while the genes that flank MAT share a much higher level of identity (≥99%). The SXI1 and SXI2 dimorphic region defines the diverged region of the MAT B HD locus in C. amylolentus. We aligned the nucleotide sequences and performed a matrix comparison for the dimorphic region (∼2 kb) spanning the SXI1 and SXI2 genes in CBS6039 and CBS6273. The diversity lies in the region between the two genes, and their divergently oriented 5′ regions span roughly 600 bp with a similarity score of 92% (Figure 6). This region encodes the N-terminal dimerization regions known to be variable and which also defines alleles in other species (please see Text S1, and Figures S10, S11, S12, S13 for further information on analyses of HD dimorphic region in meiotic progeny). Moreover, the sequence length for CBS6273 is slightly shorter than for CBS6039 at the 3′ end of the region we sequenced for the SXI2 gene. In summary, the SXI1 and SXI2 genes span ∼2 kb and define the B MAT locus in C. amylolentus. Although it is not yet clear whether there are any other sexually dimorphic regions beyond SXI1 and SXI2 (which could reflect expansion of the HD locus), our analysis based on PCR assay showed that the areas flanking the SXI1 and SXI2 genes are conserved enough between CBS6039 and CBS6273 that primers designed based on CBS6039 sequence amplify corresponding regions from CBS6273 (data not shown).
To determine whether the pheromone receptor gene STE3 lies within the A P/R mating-type locus, we performed Southern blot analysis using genomic DNA from the two strains of C. amylolentus. The STE3 PCR product derived from CBS6039 was used as a probe, and only hybridized to the lanes containing CBS6039 DNA with no hybridization to CBS6273 (Figure 4B). This analysis provides evidence that the STE3 gene differs between the two C. amylolentus strains and the pheromone receptor gene is also linked to mating-type. Extensive additional Southern and PCR data (summarized in Figures S2, S3, S4) document that the sequence divergent region of the P/R locus spans more than 60 kb encompassing multiple genes (mating pheromone genes, STE3, STE12, and STE20 among others). This contrasts with C. heveanensis in which the P/R locus is more restricted, STE3 and the MF pheromone genes are closely linked, and the LPD1, STE11, ZNF1, MYO2, and IKS1 genes are linked to but not within MAT [24]. In conclusion, in C. amylolentus a tetrapolar mating system with physically unlinked HD and P/R loci appears to define mating-type identity, and the P/R locus has expanded considerably compared to C. heveanensis, revealing an evolutionary intermediate in the transition from the tetrapolar to bipolar state that is even more closely related to the pathogenic species complex.
We conducted phylogenetic analysis of several genes that are located within the MAT locus of C. neoformans (CID1, ETF1, GEF1, LPD1, STE3, STE12, STE20, SXI1, and SXI2). This analysis included C. neoformans var. neoformans, C. neoformans var. grubii, and C. gattii representatives from the pathogenic species cluster [10] and the closely related sibling species C. amylolentus and T. wingfieldii, as well as the outgroup species C. heveanensis and T. mesenterica [24]. Based on the phylogeny of the species within the C. neoformans pathogenic species cluster, these genes can be classified into three different groups: species specific (CID1, GEF1, LPD1), mating-type specific (ETF1, STE3, STE12, STE20), and mating-type unique genes (SXI1, SXI2) (Figure 7 and Figures S6 and S7).
The species- specific phylogeny of CID1, GEF1, and LPD1 is consistent with the hypothesis that this region has been recruited into the MAT locus of C. neoformans during the transition from a tetrapolar to a bipolar mating system. For the sex-unique genes in the C. neoformans species complex, SXI1 and SXI2, SXI2 showed a considerably higher level of polymorphism between the two alleles from the CBS6039 and CBS6273 C. amylolentus isolates (Figure S7). ETF1 might have gained its mating-type specific divergence in the C. neoformans species complex after the common ancestor of the species complex split from the other sibling species while STE3, STE12, and STE20 all have mating type specific phylogenetic patterns within the C. neoformans species complex. In C. amylolentus, PCR primers designed based on CBS6039 sequences were only able to amplify these genes from CBS6039, but not from CBS6273, indicating the existence of considerable polymorphisms between the two alleles of CBS6039 and CBS6273 for each of these three genes. This is consistent with the mating type specific pattern observed within the C. neoformans species complex. Additionally, for STE3 and STE12, the clusters of C. amylolentus and T. wingfieldii are more closely related to the MATa alleles of C. neoformans species complex, suggesting a possible common origin of these alleles, as well as an early involvement of the STE3 and STE12 genes in the evolution of mating type determination.
Following definition of the mating-type locus for both sibling species, we sought to identify a sexual cycle for C. amylolentus and T. wingfieldii to determine whether the A, B, or both A and B MAT loci control sexual reproduction. It was previously thought that both of these sibling species were asexual [31]; however, we discovered an extant heterothallic sexual cycle for C. amylolentus. We conducted mating assays and found the following optimal conditions: V8 pH = 5 solid medium with incubation for one week or longer at room temperature in the dark. The cross between C. amylolentus strains CBS6039 and CBS6273 produced hyphae with fused clamp connections and aseptated basidia terminating in four long individual spore chains (please see further discussion on strains CBS6039 and CBS6273 in Text S1, and formal description of mating in Materials and Methods section), similar to matings in C. neoformans and C. gattii. Sterigmata were not observed (Figure 8A–8F). A marked, obvious feature is the shape of the spores which are ellipsoid in the pathogenic species [32] whereas C. amylolentus spores are round and similar in size to yeast cells. Crosses of either C. amylolentus strain with T. wingfieldii were infertile. Because there is only one strain of T. wingfieldii available, T. wingfieldii might be fertile in the presence of a suitable partner, similar to the two interfertile C. amylolentus strains, or it could be a sterile isolate.
In C. amylolentus, we observed that the periphery of some mating patches contains a mixture of both monokaryotic hyphae and sectors in which mating occurs to produce dikaryotic hyphae indicative of sexual reproduction. The dikaryotic sectoring phenotype is present in most mating patches and also serves as a visual assay for mating. The structures produced during the sexual cycle of C. amylolentus were visualized in greater detail by microscopy. The four spore chains are each very long consisting of >15 (quantified by counting 10 individual basidia) spores per chain and clamp cell connections are visible by light microscopy and SEM (Figure 8A–8F). Based on fluorescence microscopy with Hoechst 33258 or Sytox green, dikaryotic hyphae and both uni- and occasional bi-nucleate spores were observed (Figure S8A–S8D). In the Filobasidiella lineage, C. neoformans and C. gattii produce both dikaryotic (heterothallic) and monokaryotic (homothallic) hyphae while F. depauperata produces only monokaryotic hyphae. The presence of dikaryotic hyphae in C. amylolentus provides evidence that opposite-sex mating occurs during the sexual cycle [11], [12]. Additionally, the presence of two nuclei in some basidiospores could result from either a mitotic nuclear division in the spore or packaging of two nuclei into some spores (as occurs in pseudo-homothallic species) [33].
Interestingly, the cap of the spore chain represents a quartet of basidiospores. These spores are the oldest in the spore chain and remain tightly attached to each other. Younger spores in the four spore chains remain attached to the preceding and following spores in the chain but often not to their meiotic siblings in the other three spore chains. Thus, the quartet spore cap appears to tether the ends of the spore chains together. This feature has not been described in the pathogenic Cryptococcus species. In summary, microscopic examination of mating structures in C. amylolentus has revealed both shared hallmarks with sexual reproduction in the pathogenic Cryptococcus species and novel features.
To determine if recombination occurs, and to further assess whether the mating system of C. amylolentus is tetrapolar or bipolar, we performed microdissection of random progeny (F1 set 1) and individual spore chains (F1 set 2) followed by molecular genotyping analysis for both MAT markers and a genome-wide set of RAPD markers. We designate the CBS6039 parent as A1B1 and the CBS6273 parent as A2B2, according to the designation used for a tetrapolar mating system and our findings, assigning A as the P/R locus and B as the HD locus (as in T. mesenterica, C. heveanensis, and U. maydis [24], [26]).
For F1 set 1 (F1S1), a total of 40 spores were dissected and 28 (70%) germinated (Tables S1 and S2). The progeny were all haploid based on FACS analysis with C. neoformans as reference (data not shown). Genotyping using MAT markers and RAPD markers revealed that most of the progeny inherited all of the parental alleles from CBS6039 (A1B1) (Tables S1 and S2) and did not appear to be meiotic recombinants. Of the 28 progeny, three (11%) did show recombination within the P/R locus (#17, 27, and 28), whereas only one additional progeny (3.5%) exhibited reassortment between the P/R and HD loci (#18). We hypothesize that this is likely due to the dissection of a mixture of yeast cells, blastospores (mitotic pre-meiotic cells produced by budding from the hyphae or clamp cells), and basidiospores (meiotic sexual spores) [5], [34], which are all morphologically similar for this species. Similar to C. neoformans, in C. amylolentus blastospores can be generated from the clamp cell, and the following repeated mitotic events tend to produce a cluster of cells at the hyphal septa. This may explain why we did not observe an equal distribution of markers from the two parental strains among the blastospores, as they could have been mitotic products from one common parental blastospore. That many isolates in F1S1 could be blastospores is also supported by analysis of the mitochondrial genome segregation (as shown below) that revealed a majority of this progeny set possess nuclear and mitochondrial genomes inherited from different parents. Remarkably, 22 (78%) of the progeny are sterile and unable to undergo sexual reproduction with either parent or their F1 siblings. It is interesting that progeny that appear to be derived from blastospores are, for unknown reasons, frequently sterile.
To analyze meiotic basidiospores specifically, we dissected F1 set 2 (F1S2) from four well-resolved individual spore chains (one chain each from four different basidia). The germination frequency was 91% (31/34), and 58% (18/31) of the progeny were sterile with both parents (Table 1). All of this progeny set were also haploid based on FACS analysis (data not shown). Molecular analysis of this set using the same six MAT A or B genes revealed that 64.5% (20/31) of the progeny resembled one or the other parent (A1B1 or A2B2) while the other ∼35% exhibit evidence of recombination within the P/R locus and/or between the HD and P/R loci (i.e. A1B2 or A2B1 progeny) (Table 1). In contrast to the first F1 progeny set (F1S1), genotyping of the spore chain derived progeny set (F1S2) using RAPD markers revealed extensive recombination (Table 2). Linkage analyses clustered markers analyzed in this study into several linkage groups, indicating independent inheritance of markers (data not shown). In addition, analysis of the markers implemented in this study revealed that for each marker, the two parental alleles were equally inherited across the entire progeny set (Table 2). Specifically, for each marker, the percentages of the CBS6039 allele ranged between 35% and 71%, which did not show any significant bias toward one parental allele (chi-square test, P>0.05). Similarly, the percentage of the CBS6039 allele that each progeny inherited ranged from 25% to 80%, and again these values reflect equivalent inheritance of alleles from either parent (chi-square test, P>0.05). Moreover, we observed that meiotic recombination in C. amylolentus resulted in the generation of new combinations of alleles in the progeny given the multiple genotypes present in the different spore chains analyzed. The observed high level of recombination and equivalent inheritance of the two parental alleles support the conclusion that meiosis occurs in C. amylolentus.
We discovered that some of the progeny that are sterile with either parent are in fact interfertile with other progeny. Specifically, of the 59 F1 progeny (mixture of blastospores and basidiospores), there are 14 A1B1 F1 progeny that are fertile with the A2B2 parent CBS6273, one A2B2 progeny that is fertile with the A1B1 parent CBS6039, and one that is fertile with both parents. Among the spore chain derived (F1S2), progeny #13 (A2B2) was found to be able to mate with progeny #3 and #24 (A1B1). Successful mating was also observed between some MAT recombinant progeny. Specifically, successful mating was observed when A1B2 progeny (F1S2 #10 and #16) and A2B1 progeny (F1S1 #18, F2 #1, #2, and #5) were co-cultured together (with the exception of mating between F1S2 #16 and F1S1 #18). None of these MAT recombinant progeny mated with either parent, further confirming that C. amylolentus possesses a tetrapolar mating system (Figure 9).
Because only 32% (19/59) of the progeny are fertile in both progeny sets, we assessed whether fertility increases with an additional sexual cross or mitotic passage. Even after several passages on YPD, the sterile phenotype remained stable (data not shown). We crossed F1S2 progeny #3 and CBS6273 to generate a backcross progeny set (F2). Interestingly, most of the basidia in the cross were barren and if spore chains were present, the number of spores per chain was significantly reduced when compared to matings between the parental strains. We were successful in dissecting spores from two individual spore chains. The germination rate was 54% (6/11) and all of the progeny were fertile (50% with the CBS6039 parent and the remaining A2B1 progeny are interfertile with the F1S2 progeny #10 and #16 (Table 1)). All of the progeny examined are haploid with the exception of F2 #4, which is diploid by FACS yet remains self-sterile (data not shown).
In summary, taken together our genotyping data indicates that meiotic recombinants are present among the sexually produced progeny and our evidence is that the sexual cycle of C. amylolentus conforms to a modified tetrapolar mating system in that 1) sterile progeny are also frequently produced, and 2) the ratio of the four mating types is unbalanced.
To assess the mitochondrial inheritance pattern during sexual reproduction of C. amylolentus, SNPs were first identified in two mitochondrial genes, NAD4 and NAD5, between the two parental strains, CBS6039 and CBS6273, by PCR amplification and sequencing. Of the 65 progeny screened, no intra- or inter-genic recombination between the two genes was observed, and all of the progeny (with the exception of two from F1S1) typed as the CBS6273 (A2B2) parent (Table S3). The two progeny (F1S1 #13 and #16) that contain the A1B1 mitochondrial genome are likely dissected parental yeast cells, because they also both possessed A1B1 alleles at all of the other markers that were typed. For the other nuclear non-recombinant progeny that type as the A1B1 parent, the fact that they have the A1B1 nuclear genome and the A2B2 mitochondrial genome suggests that they descend from blastospores produced after cell-cell fusion and a result from cytoduction of the CBS6039 nuclear genome and CBS6273 mitochondrial genome. All other progeny that are derived from meiotic basidiospores contained a recombinant nuclear genome paired with the mitochondrial genome exclusively from the A2B2 parent (CBS6273). These results demonstrate that mitochondria are uniparentally inherited from the A2B2 parent during C. amylolentus sexual reproduction, similar to C. neoformans in which mtDNA is inherited uniparentally from the a parent [35], [36], [37], [38].
The current study extends the previous analyses of the MAT locus in the pathogenic Cryptococcus species to the closest known species, T. wingfieldii and C. amylolentus. To determine the structure of MAT in both species, we cloned and sequenced the HD and P/R loci. Due to their close phylogenetic relatedness [22], characterization of MAT has provided key insights into the evolution of MAT and revealed important aspects of the transition from an ancestral tetrapolar to a bipolar mating system in C. neoformans and C. gattii [13], [14], [39], [40].
A previous phylogenetic analysis using a six-gene multi-locus sequencing (MLS) approach identified the most closely related species to the pathogenic Cryptococcus species complex [22]. This analysis identified the sensu stricto (closely related) and sensu lato (more distantly related) species that provide unique vantage points to address questions such as: when and how did the bipolar mating system evolve? And what, if any, affects does the emergence of bipolar mating systems have on the pathogenesis of C. neoformans and C. gattii? Previous studies on a more distantly related sensu lato species, Cryptococcus heveanensis, revealed it to be tetrapolar [24]. The key advances presented here provide additional critical insights. First, as sensu stricto strains, C. amylolentus and T. wingfieldii are much more closely related to the pathogenic species C. neoformans/C. gattii than is C. heveanensis; hence the transition to bipolarity in the pathogens was even more recent than could be concluded based on the studies of C. heveanensis alone. Second, by providing additional tetrapolar outgroup species, we can conclude that the transition was from tetrapolar to bipolar, not vice versa. Third, the P/R locus is much more expanded in C. amylolentus compared to C. heveanensis, providing further insights on the evolution of the MAT and this key step in the process. Furthermore, the tetrapolar mating system in C. amylolentus showed indications of deviation from the classic tetrapolar model in that many MAT loci recombinant progeny are sterile and progeny that resemble one parent at the MAT loci dominate the progeny population. Moreover, the organization of MAT in these sibling species mirrors key aspects (gene acquisitions, chromosomal rearrangements, etc.), which shaped the evolution of the mating-type locus in the pathogenic Cryptococcus species complex.
Previous analysis resolved the phylogeny surrounding the pathogenic Cryptococcus species cluster and revealed that T. wingfieldii and C. amylolentus are sibling species, the closest relatives of the pathogenic species, and members of the Filobasidiella clade [22]. The MAT loci of T. wingfieldii and C. amylolentus share overall synteny, with two major inversion events present between the P/R loci of the two species (Figure 3). For this analysis, the type strain, the only isolate of T. wingfieldii available, was employed. Two strains of C. amylolentus are available and we characterized MAT for the type strain CBS6039 and representative sequences for CBS6273. The two MAT loci of T. wingfieldii and C. amylolentus are physically unlinked and present on different chromosomes (Figure 1B and Figure 2B). The MAT assembly for C. amylolentus is similar to T. wingfieldii in that both homeodomain transcription factors are present and opposite in their orientations, similar to the paired, divergently oriented bE and bW genes in U. maydis. Several other key genes (SPO14, RPL22, and CAP1) are present and these lie within MAT in C. neoformans but appear to lie outside of MAT in C. amylolentus. The configuration of the HD genes in the sibling species provides evidence that the ancestral form of the HD locus contained both SXI1 and SXI2, similar to tetrapolar mating systems in other basidiomycetes, and that loss of one or the other of the HD genes punctuated the formation of a bipolar mating system.
Of the >20 genes identified in the HD (B) and P/R (A) loci of the sibling species, we determined which genes define MAT. Because only one strain of T. wingfieldii is available, we were unable to establish which of the genes in the B and the A loci are MAT-specific. By comparing sequences from the two strains of C. amylolentus, we determined that the MAT-specific region in the HD locus is likely restricted to the ∼3 kb SXI1 and SXI2 dimorphic region. The divergence is present in the 5′ regions of SXI1 and SXI2, similar to recent findings on the B MAT locus alleles of C. heveanensis [24]. This is also consistent with findings in other fungi where the N-terminal regions of the homeodomain proteins are typically variable and heterodimerization only occurs when compatible (or different allelic versions) of the proteins are brought together promoting activation of genes required for sexual development [30], [41].
We also sought to define the extent of the sex-specific region in the MAT A locus. Our extensive Southern and PCR analysis document that the P/R locus has been expanded to encompass >60 kb in C. amylolentus, including the STE3 and MF pheromone genes that lie >30 kb apart in contrast to their close linkage in the P/R MAT A locus of C. heveanensis (Figures S2, S3, S4). In addition, several genes encompassed within this expanded C. amylolentus P/R locus are linked to but outside the defined P/R locus of C. heveanensis [24]. Thus, one of the two MAT loci has expanded in C. amylolentus but the two remain unfused.
We also report the discovery of sexual reproduction in C. amylolentus. Fortunately, the only two strains of C. amylolentus available in the world are of opposite mating-type and fertile, enabling us to define the sexual cycle for C. amylolentus. Mating structures in C. amylolentus resemble those observed in C. neoformans, and differed from C. heveanensis, consistent with its closer phylogenetic relationship with C. neoformans than with C. heveanensis.
Mating in C. amylolentus produces many sterile progeny, suggesting that sexual reproduction may pose a risk in which not all of the progeny produced are fertile. Although the underlying mechanism(s) causing sterility in the C. amylolentus progeny is not clear, there are several possible explanations. First, it is possible that aneuploids (1N+1) are generated during meiosis that could be sterile. FACS analysis of the examined progeny suggested that all of the progeny are haploid with the exception of a single diploid (F2 progeny #4), but FACS is not sensitive enough to detect 1N+1 aneuploids. Employing comparative genomic hybridization (CGH) of the C. amylolentus parental strains with the sterile progeny will be necessary to address the issue of possible aneuploidy generated during mating. Second, meiosis is mutagenic and sexual reproduction may also increase transposition in the genome. The resulted mutations and/or the insertion of transposons in MAT or elsewhere might result in sterility. Third, the increased sterility among progeny could be due to sex induced silencing of repetitive elements within MAT and linked fertility genes [42] or damage to MAT caused by gene conversion events. Sex induced silencing requires the RNAi machinery. However it is not known yet if C. amylolentus possesses these genes. The C. amylolentus genome sequence will allow this question to be answered. Additionally, we cannot exclude the possibility that the sterility observed among the progeny is due to divergence/incompatibility between the mating machineries of the two C. amylolentus strains, or to nuclear-mitochondrial incompatibility that has been observed in other yeasts [43].
From the genotyping analysis, it is evident that extensive recombination occurred among the progeny produced by sexual reproduction. Additionally, we observed a 1∶1 segregation pattern of the two parental alleles in the progeny population. This segregation data and the high level of genetic exchange in the progeny (especially the F1S2 spore chain derived progeny set) provide strong evidence that meiosis occurs within the basidium during sexual reproduction. Additionally, RAPD analysis revealed that in some spore chains from the F1S2 and the F2 progeny sets, more than four genotypes are present in a single chain. There are several possible explanations. In C. neoformans, meiosis typically gives rise to four meiotic products and it was recently shown that a single meiotic event occurs in each basidium [44]. In C. amylolentus, more than one meiotic event could occur in the basidium. However, this would have to involve post meiotic nuclear fusion and a second round of meiosis. In this case, up to eight genotypes could be produced from one basidium. Also, high gene conversion events favoring some alleles over others could result in a non-Mendelian inheritance pattern and skew the resulting genotypes in each individual spore chain.
Another possible explanation for the observed >4 genotypes/basidium that we favor is the presence of aneuploids in the progeny population. The RAPD markers employed in this analysis differentiate the two parental strains by the presence or absence of a PCR product. If progeny are aneuploid for one or more chromosome, they could appear unique and differ from the two parental strains. In this aneuploidy model we expect the basidiospores from one basidium to share four common genotypes with the exception of a few rarer genotypes (potential aneuploids). This model is consistent with our RAPD data (Table 2 and Table S2) in which several spore chains contain four distinct major genotypes and several anomalous minority genotypes that are closely related to one of the four consensus majority genotypes in a given spore chain (Figure S9). One limitation is that we are currently unable to score the heterozygous state of the aneuploids, which can be detected by co-dominant markers such as PCR-RFLP and CGH, and this provides fertile ground for future studies.
Micromanipulation of the individual spore chains representing F1S2 and the F2 progeny generated progeny that are recombinant at the MAT loci (A1B2 and A2B1), and these MAT recombinant progeny are inter-fertile, but cannot mate with either of the two parental strains, proving that C. amylolentus has a tetrapolar mating system. However, among those isolates that were fertile the A1B1 genotype was overrepresented, whereas the other three genotypes were underrepresented (Table 1 and Table S1).
MAT in the pathogenic Cryptococcus species evolved from an ancestral tetrapolar system with physically unlinked B and A loci and these loci fused into a large bipolar MAT locus. In C. amylolentus, the structure of MAT indicates a tetrapolar mating system with allelic diversity in both the B and A loci. Although evidence from T. wingfieldii, C. amylolentus, C. heveanensis [24], and C. disseminatus [9] suggests that MAT evolved from a tetrapolar to a bipolar system, an alternative hypothesis could be just the opposite: namely that the ancestral form of MAT was bipolar and instead evolved into a tetrapolar mating system in these species. In such a scenario, a bipolar locus could have suffered a chromosomal break resulting in the formation of physically unlinked HD and P/R loci in a derived rather than ancestral tetrapolar fungal species. In this model, the tetrapolar state would then be ancestral in some species and derived in others. We do not favor this alternative model and the one we propose (Figure 5) instead illustrates the evolution of the bipolar MAT in the pathogenic Cryptococcus species from an ancestral tetrapolar system. The evidence adduced now for three sibling species supports the more parsimonious model that C. amylolentus, T. wingfieldii, and C. heveanensis all reflect a common, shared ancestral tetrapolar state rather than multiple independent derived states.
MAT evolution in fungi has defined a continuum of transitions in modes of sexual reproduction from outcrossing tetrapolar multiallelic systems to bipolar biallelic systems that promote inbreeding to unipolar uniallelic same-sex mating that promotes extreme inbreeding and clonality [15], [45]. Aside from bipolar and tetrapolar mating systems, some deviations from these classic mating systems have been reported recently. For example, a pseudo-bipolar mating system has been recently found in the red yeast Sporidiobolus salmonicolor [46], [47]. The authors found that in this species, mating is normally bipolar and governed by a large continuous MAT locus with the A and B regions located at either end. However, meiotic recombination may occur between the MAT locus alleles, generating novel mating types, and thus increasing MAT allele number and evolutionary rates for some MAT genes.
Results from our studies illustrate features of both the transition from tetrapolarity to bipolarity in the closely aligned saprobic and pathogenic Cryptococcus species, and also the emergence of sexual reproduction in which one mating-type has an advantage resulting in a higher proportion of fertile progeny of one mating-type (A1B1) that might have ultimately led to the emergence of unisexual same-sex mating in C. neoformans. In conclusion, C. amylolentus and C. heveanensis have physically unlinked HD and P/R loci and this arrangement further supports tetrapolarity as the ancestral configuration, and that the transition to bipolarity occurred recently and concomitantly with the emergence of the pathogenic C. neoformans/C. gattii species cluster.
These studies on the molecular events leading to the fusion of two unlinked sex determining regions of the genome in the ancestral tetrapolar state to the derived bipolar mating systems mirror aspects in the hypothesized origin of sex chromosomes of more complex multicellular eukaryotes, including plants, insects, fish, and mammals. Namely, Ohno hypothesized that sex determinants arise on an autosome, and then gradually capture this chromosome, which evolves to become a sex chromosome [39]. These steps include the original emergence of the sex determinant, the recruitment of other genes that function in sex to the incipient sex chromosome, and rearrangements and the acquisition of repetitive elements that lead to two sexually dimorphic chromosomes. The transition from two unlinked sex determinants in tetrapolar fungi to two linked sex determinants in bipolar fungi, and the fact that this transition has occurred repeatedly and independently, provides further support for the hypothesis that sex determinants arise at distant genomic locations and then become linked through gene movement or chromosomal translocations in both mating type loci and sex chromosomes.
The two strains of C. amylolentus, CBS6039 and CBS6273, and the one T. wingfieldii isolate, CBS7118, were obtained from the Centraalbureau voor Schimmelcultures (CBS) Fungal Biodiversity Centre in the Netherlands. Both CBS6039 and CBS6273 were originally isolated from insect frass in South Africa, while CBS7118 was originally isolated from rubber sheet in Indonesia. All species were grown and maintained on yeast extract-peptone-dextrose (YPD) medium at 24°C. Mating assays were performed on V8 medium pH = 5 in the dark and also at 24°C. Random spore dissection was performed on YPD medium as previously described [44]. Spore chain dissection was performed by first transferring a well separated spore chain onto a drop of zymolyase on YPD, and after incubation at 24°C for 15 minutes, individual spore from the spore chain was dissected as previously described [44].
To isolate genomic DNA from T. wingfieldii and C. amylolentus, cells were cultured in 50 ml of liquid YPD shaking overnight at 24°C. The pellets were then lyophilized overnight and the CTAB method of fungal DNA isolation was performed as described before [24]. Plasmid DNA from positive TOPO clones was extracted using the QIAprep Spin Miniprep Kit (Qiagen, Valencia, CA), fosmid DNA was isolated using a modified miniprep protocol, and DNA from the shot-gun sequencing libraries was extracted using the DirectPrep96 Miniprep Kit (Qiagen, Valencia, CA). Additionally, progeny DNA was isolated using a modified miniprep protocol and colony lifts were performed to isolate DNA from individual colonies in each fosmid library according to the protocol described in [28].
We designed degenerate PCR primers using the online computer program, COnsensus-DEgenerate Hybrid Oligonucleotide Primer (CODEHOP, http://blocks.fhcrc.org/codehop.html) to identify MAT specific genes in T. wingfieldii. The primers consist of a relatively short 3′ degenerate core and a longer 5′ non-degenerate consensus clamp designed by multiple sequence alignments [48]. We aligned sequences for two flanking genes (FAO1 and NOG2) and two recently acquired MAT genes (RPO41 and LPD1) from C. neoformans var. neoformans and var. grubii, C. gattii, U. maydis, and C. cinerea to design the degenerate PCR primers (see Table S4 for primer information). PCR was performed on genomic DNA isolated by the CTAB extraction method as template and products were separated by gel electrophoresis. Products with the strongest ethidium bromide-staining signal were then gel extracted using the QIAquick Gel Extraction Kit (Qiagen, Valencia, CA) followed by transformation into E. coli using the TOPO-TA cloning Kit (Invitrogen, Carlsbad, CA). Plasmid DNA was purified from transformants and then sequenced. For C. amylolentus, degenerate primers were not used. Instead, primers from T. wingfieldii were directly used to amplify MYO2, LPD1, SXI1, and SXI2 in both C. amylolentus strains (see Table S4 for primer information).
We employed the CopyControl Fosmid Library Production Kit (Epicentre, Madison, WI) to generate fosmid libraries for T. wingfieldii and C. amylolentus strain CBS6039. At least 2.5 µg of CTAB isolated genomic DNA was randomly sheared using a 200 µl small bore pipette tip and sheared DNA was end-repair converted to blunt 5′ phosphorylated ends using End-Repair Enzyme Mix, dNTPs, and ATP. We then separated the end-repaired DNA overnight using a Contour-clamped Homogenous Electric Field (CHEF) on a CHEF DR-II apparatus (Bio-Rad, Hercules, CA). The following conditions were used: 1- to 6-second switch time, 6 V/cm, 14°C for 14–15 hrs in 0.5X TBE. The size-fractionated DNA, 25 to 40 kb fragments, was recovered by gel extraction and the DNA was precipitated with sodium acetate and ethanol. The precipitated insert DNA was then ligated into the CopyControl pCC1FOS cloning-ready vector and incubated overnight at 24°C. The ligated DNA was packaged in phage particles and plated on E. coli phage-resistant cells (EPI100-T1R plating strain) overnight at 37°C (detailed protocol can found at http://www.epibio.com/item.asp?ID=385). Approximately 16,000 fosmid clones were picked into 96-well plates and transferred to 384-well plates for long-term storage at −80°C. The 384-well plates were replicated onto high-density filters for hybridizations using the MAT genes.
Positive fosmid clones were sequenced using the shot-gun sequencing method described by Metin et al. [24]. Six fosmids were pooled and sequenced to generate the assembly for T. wingfieldii and four fosmids were individually sequenced to generate the assembly for C. amylolentus strain CBS6039. Sequencing reactions were performed using Big Dye chemistry v3.1 (Applied Biosystems, Foster City, California, United States) and analyzed on an Applied Biosystems 3730xl capillary sequencer in the Biological Sciences Sequencing Facility at Duke University. For each library, approximately ∼1200 sequence reads were imported into UNIX using Phred and Phrap to assemble the sequences into larger contigs of overlapping sequence [49], [50], [51]. To close gaps in the assemblies, we designed primers from contig ends using Primer 3 (http://frodo.wi.mit.edu/primer3/). The GenBank accession numbers for T. wingfieldii are HM368525 (HD locus) and HM368524 (P/R locus). The GenBank accession numbers for the HD locus and the three P/R contigs in C. amylolentus CBS6039 are: HM640220 (HD locus), HM640221 (RPL39-MYO2), HM640222 (LPD1-STE12), and HM640223 (GEF1-MFA). The GenBank accession numbers for genes from C. amylolentus CBS6273 are: HM640224 (SXI1), HM640225 (SXI2), HM640226 (GEF1), HM640227 (LPD1), and HM640228 (ETF1).
To determine the ploidy of the two C. amylolentus and one T. wingfieldii strains, we cultured the isolates on YPD medium for 2 days at 24°C. Each isolate was processed for flow cytometry as previously described [5], [52] and analyzed using the FL1 channel on a Becton-Dickinson FACScan. The ∼20 Mb genome of C. neoformans/gattii was used as a reference for ploidy determination (including haploid and diploid controls).
To isolate chromosomal DNA of C. amylolentus and T. wingfieldii, spheroplasts were generated following the spheroplasting protocol for C. neoformans and C. gattii [53]. The plugs containing spheroplasts were lysed at 55°C for at least 24 hrs in lysing solution (0.5 M EDTA/10 mM Tris-Cl (pH = 10) and 1% Sarcosyl) and then loaded onto a PFGE apparatus and separated for approximately 5 days on a CHEF DR-II apparatus (Bio-Rad, Hercules, CA). The following conditions were used: Block 1: 75- to 150-second switch time, 4 V/cm, 13°C for 30 hrs and Block 2: 200 to 400-second switch time, 4 V/cm, 13°C for 60 hrs in 0.5X TBE. The gel was then stained in ethidium bromide for 15 minutes, destained for an hour, and visualized using a UV lamp. The chromosomal DNA was blotted overnight onto Hybond (Amersham, Piscataway, NJ) membranes in 20X SSC using standard protocols. The membrane was then hybridized to MAT gene probes generated by PCR. We also performed Southern blot analysis on genomic DNA from C. amylolentus that was digested with EcoRV, PstI, BamHI, or NotI. The digested DNA was separated on an agarose gel and probed with the RPL22 gene probe amplified from C. amylolentus, with primers designed for T. wingfieldii (see Table S4 for primer information).
We compared sequences from the HD locus of T. wingfieldii to those of C. amylolentus by employing a matrix comparison (or dot plot) analysis. To generate each dot plot, we employed the Molecular Toolkit's online nucleic acid dot plots program (http://www.vivo.colostate.edu/molkit/dnadot/). The parameters for the dot plot analyses were as follows: the window size was 51 and the mismatch limit was 6. We also employed the bioinformatic software, Artemis Comparison Tool Release 8 (http://www.sanger.ac.uk/resources/software/) to generate comparison plots across MAT of T. wingfieldii to C. amylolentus and both sibling species compared to C. neoformans serotype D strain JEC21 [54]. The input file was created using WebACT (http://www.webact.org/WebACT/home) with the Blastn algorithm [55].
Phylogenetic analysis was performed on coding sequences using MEGA 5 [56]. To determine the phylogenetic relationship, the Neighbor-Joining method based on the Kimura 2-parameter model was employed [57]. For statistical support, 500 replicates were performed and bootstrap values were calculated.
We performed Southern blot analysis using standard protocols on genomic DNA from C. amylolentus digested with BamHI, BglI, ClaI, EcoRI, or NcoI. The digested DNA was separated on an agarose gel and probed with the STE3 gene, stripped (0.1% SDS and 0.1X SSC in boiling water, 3 times for 15 minutes each), and probed with the contig ends from the P/R assembly in C. amylolentus amplified by PCR (see Table S4 for primer information).
Standard description: Filobasidiella amylolenta Findley & Heitman sp. nov.
Etymology: The epithet is chosen to be identical with that of C. amylolentus (Van der Walt, D.B. Scott & Klift) Golubev 1981 [58]. Heterothallic fungus. Hyphae dikaryotic, clamped connections fused. Aseptate basidia, 3–5 µm diameter, terminating in four chains of basidiospores. Basidiospores are aerial, round, and 2–2.5 µm in diameter.
Holotype: Mounted teleomorph is paired cultures of C. amylolentus type strain, CBS6039T (A1B1) crossed to CBS6273 (A2B2) on V8 medium (pH = 5). These strains were originally isolated from insect frass in South Africa [58]. A slide preparation of mating structures, basidia and basidiospores, is deposited in the USDA's Systematic Mycology and Microbiology Laboratory in Beltsville, Maryland (deposit number: BPI 881008). Strains CBS6039 (mating-type A1B1) and CBS6273 (mating-type A2B2) should be designated as the ex-type strain and the isotype strain, respectively, for the teleomorph Filobasidiella amylolenta.
Latin description: Filobasidiella amylolenta Findley & Heitman sp. nov.
Fungus heterothallicus. Hyphae dikaryoticae, fibulis fusis. Basidia aseptata, 3–5 µm lata, quatuor catenas basidiosporarum producentia. Basidiosporae aeriae, globosae, 2–2.5 µm diametro.
Spores and yeast cells were cultured on slides coated with V8 pH = 5 medium for one week or longer to allow production of mating structures. The slide was first washed with phosphate buffered saline (PBS) followed by staining the cell wall using a solution of Calcofluor white (fluorescent brightener 28 F-3397; Sigma) for 15 minutes. Slides were rinsed with PBS and fixed for 15 minutes in fixing solution (3.7% formaldehyde and 1% Triton-X100 in PBS). After permeabilization of the fungal cells, nuclear content was examined by staining with Sytox green (Molecular Probes) for 30 minutes. Slides were washed with PBS and a cover slip was applied to the slide for observation. In addition to staining spores and yeasts, mating filaments were also stained. Agar pieces were removed from mating plates and washed several times with PBS. Calcofluor white was added directly to the agar piece for 30 minutes, followed by washing with PBS, and fixing for 45 minutes. After permeabilizing samples, filaments were washed with PBS and stained with 1 mg/ml Hoechst 33258 (Invitrogen, Carlsbad, CA) overnight at 4°C. The next day, samples were washed with PBS, a thin slice of the agar (containing the mating filaments) was removed using a razor blade and a mounting solution containing anti-fade (Invitrogen, Carlsbad, CA) was added to the agar slice on a slide. The slides were sealed with nail polish and stored at 4°C in the dark after microscopic evaluation. All staining was performed at 24°C, unless otherwise noted. SEM was performed on C. amylolentus matings incubated on V8 pH = 5 medium for 2 weeks. The specimen was prepared and analyzed as described in [28]. Microscopy was performed with an Axioskop 2 plus upright microscope (Zeiss). Images were captured using an AxioCam MRm camera. Scanning electron microscopy was performed and viewed on a JEOL JSM 5900LV (JEOL U.S.A., Peabody, MA) SEM at 15 kV.
Microdissection of spores (random or individual spore chains using zymolyase (Zymo Research Corp., Orange, CA, USA)) was performed on YPD medium incubated at 24°C for two days to allow spores to germinate.
Genotyping of the MAT loci was achieved using a set of PCR markers (RPL39, GEF1, and STE3) and PCR-RFLP markers (SXI1 (enzyme EcoRV), SXI2 (enzyme RsaI), and ETF1 (enzyme DdeI)). To genotype other genomic regions, we used a set of 20 RAPD markers (Table S4). Linkage analyses indicated 18 of these 20 markers are not derived from the C. amylolentus MAT loci, with exception of markers Pi_Random_24_No.2 and JOHE22656_No.1, which were positioned in the same linkage group with HD markers. Non-MAT-association of nine of these 18 markers were further confirmed by cloning and sequencing of the polymorphic bands, as none of them was MAT specific sequence (data not shown). We designate the CBS6039 parent as A1B1 and the CBS6273 parent as A2B2, according to the designation used for a tetrapolar mating system and our findings, assigning A as the P/R locus and B as the HD locus (as in T. mesenterica, C. heveanensis, and U. maydis [24], [26]). Recombination was scored according to marker exchange for the P/R and/or HD locus. Recombination frequency among RAPD markers was inferred using program MapMaker.
These genotyping data was further analyzed using program MapMaker to generate genetic linkage groups. Additionally, the UPGMA clustering method implemented in the software MEGA 5 was used to analyze the genetic relationships among F1S2 progeny isolated from the same basidium.
|
10.1371/journal.pntd.0002602 | Septins of Platyhelminths: Identification, Phylogeny, Expression and Localization among Developmental Stages of Schistosoma mansoni | Septins are a family of eukaryotic GTP binding proteins conserved from yeasts to humans. Originally identified in mutants of budding yeast, septins participate in diverse cellular functions including cytokinesis, organization of actin networks, cell polarity, vesicle trafficking and many others. Septins assemble into heteroligomers to form filaments and rings. Here, four septins of Schistosoma mansoni are described, which appear to be conserved within the phylum Platyhelminthes. These orthologues were related to the SEPT5, SEPT10 and SEPT7 septins of humans, and hence we have termed the schistosome septins SmSEPT5, SmSEPT10, SmSEPT7.1 and SmSEPT7.2. Septin transcripts were detected throughout the developmental cycle of the schistosome and a similar expression profile was observed for septins in the stages examined, consistent with concerted production of these proteins to form heterocomplexes. Immunolocalization analyses undertaken with antibodies specific for SmSEPT5 and SmSEPT10 revealed a broad tissue distribution of septins in the schistosomulum and colocalization of septin and actin in the longitudinal and circular muscles of the sporocyst. Ciliated epidermal plates of the miracidium were rich in septins. Expression levels for these septins were elevated in germ cells in the miracidium and sporocyst. Intriguingly, septins colocalize with the protonephridial system of the cercaria, which extends laterally along the length of this larval stage. Together, the findings revealed that schistosomes expressed several septins which likely form filaments within the cells, as in other eukaryotes. Identification and localization demonstrating a broad distribution of septins across organs and tissues of schistosome contributes towards the understanding of septins in schistosomes and other flatworms.
| Schistosoma mansoni is one of the causative agents of schistosomiasis, a neglected tropical disease affecting over 230 million people in the developing world. Research on new therapies for this parasitic disease has been facilitated by the recent publication of a curated draft sequence of the schistosome genome. Here, we describe proteins from the septin family found in the genome of S. mansoni. The septins are increasingly recognized as central components of the cytoskeleton of eukaryotic cells. They are linked to numerous cellular functions, although the precise role(s) of these proteins is not fully understood. Schistosome septins were seen in the miracidium and sporocyst larval stages, on superficial structures, within epidermal plates and in muscles. Notably, septins were prominently expressed in the germ cells of larval stages of the blood fluke. In addition, septins were ubiquitously immuno-localized throughout the organs and tissues of the schistosomulum stage of the parasite. This is the first report on septins in schistosomes; these proteins are broadly distributed among organs and tissues of the parasite where they likely perform diverse functions. Identification and localization demonstrating a broad distribution of septins across organs and tissues of schistosome contributes towards the understanding of septins in schistosomes and other flatworms.
| Septins comprise a family of evolutionarily highly conserved cytoskeletal proteins [1]. Absent from higher plants but otherwise ubiquitous in eukaryotes [2], [3], septins have been well characterized in human cells and model invertebrates including Caenorhabditis elegans and Drosophila melanogaster [4]–[6]. The septin family belongs to the guanosine triphosphate (GTP)ase superclass of P-loop nucleoside triphosphate (NTP)ases [1]. It was first identified due to defective cell-cycle progression in yeasts [7]. The functions attributed to septins are expanding, but span cytokinesis [7], vesicle trafficking [8], vesicle fusion [9], axonal guidance and migration [10], diffusion barriers, scaffolds [11]–[13], pathogenesis [14], [15] and others [16]. Septin function generally depends on self-assembly into hetero-oligomeric complexes, which assemble subsequently into higher-order structures such as filaments and rings [17], [18]. The septins have been consolidated as new cytoskeleton components [17].
The diversity and the number of septin-encoding genes diverge among species, ranging from one in algae [2] to 13 in humans [19]. Based on phylogenetic analysis, the metazoan septins can be classified into four groups, termed SEPT6, SEPT7, SEPT2 and SEPT3 [20]. Septins form filaments, which are composed of hetero-oligomeric complexes of septins from different groups. It has been postulated that each position of the hetero-oligomeric complex is specifically occupied by a septin group member and those cannot be replaced by member of another group [4]. Diversity among human members within most of the groups allows a multiplicity of potential complexes of human septins, with the permutation of different members of a same group in each of the hetero-oligomeric positions of the complex. However, an exception is the SEPT7 group which displays a single representative in the human genome and therefore its single member may not be replaced [4].
Schistosomiasis is considered the most important of the human helminth diseases in terms of morbidity and mortality (see [21]). Unusual among flatworms, schistosomes are dioecious, with sexual dimorphism and division of labor between sexes in the adult developmental stage [22]. Draft genome sequences for the three major species of schistosomes parasitizing humans are available [23]–[25]. Among these genomes, we identified putative septin-encoding sequences of Schistosoma mansoni and reported here four schistosome septins termed SmSEPT5, SmSEPT10, SmSEPT7.1, and SmSEPT7.2, based on sequence identity with numbered human orthologues. Phylogenetic analyses in tandem with expression profiles of transcripts among developmental stages point to structural roles for these septin-like proteins in this pathogen. Confocal imaging revealed tissue-specific and/or ubiquitous localization of septins, suggesting specialized functions in schistosomes, and in flatworms at large, in addition to cytokinesis. To our knowledge, this is the first report of septins in any member of the phylum Platyhelminthes, or indeed in any Lophotrochozoan - a major evolutionary branch of the Bilateria [26].
Mice infected with S. mansoni were obtained from the Biomedical Research Institute (BRI), Rockville, MD and housed at the Animal Research Facility of the George Washington University Medical School, which is accredited by the American Association for Accreditation of Laboratory Animal Care (AAALAC no. 000347) and has an Animal Welfare Assurance on file with the National Institutes of Health, Office of Laboratory Animal Welfare, OLAW assurance number A3205-01. All procedures employed were consistent with the Guide for the Care and Use of Laboratory Animals. Maintenance of the mice and recovery of schistosomes were approved by the Institutional Animal Care and Use Committee of the George Washington University. Procedures used for the production of antibodies were performed in accordance with the National Research Council's guide for care and use of laboratory animals [27].
Biomphalaria glabrata snails and Swiss-Webster mice infected with the NMRI (Puerto Rican) strain of S. mansoni were supplied by Drs. Fred Lewis and Matt Tucker, Biomedical Research Institute, Rockville, MD under NIH-NIAID contract HHSN272201000005I. Developmental stages of schistosomes were obtained as described [28]–[30]. In brief, adult developmental stages of the worms were recovered from infected mice by portal perfusion. Eggs were isolated from livers of schistosome-infected mice and newly hatched miracidia obtained by hatching these eggs. Primary sporocysts were obtained by transferring miracidia into sporocyst medium, as described [28] and cultured for two days [28]. Cercariae released from infected snails were snap frozen at −80°C or transformed mechanically into schistosomula which were cultured in Basch's medium [31] at 37°C under 5% CO2 in air.
Coding regions deduced in the genome of Schistosoma mansoni [23], [32] were used as a database to identify septin genes through the tBLASTn program, using all human septins as queries. Four putative orthologues of septin were identified: Smp_041060, Smp_060070, Smp_003620 and Smp_029890. The multiple sequence alignment of the GTPase domain from these four S. mansoni septins with septins from Homo sapiens, Caenorhabditis elegans, Drosophila melanogaster, Strongylocentrotus purpuratus and Ciona intestinalis was accomplished using ClustalX2 [33]. Additional alignment was performed with GTPases domains from several platyhelminths using the same approach. Phylogenetic analyses were performed using a Bayesian inference method implemented in MrBayes (v3.1.2) [34]. All analyses were run using default parameters, except by the use of the command “prset aamodelpr = mixed”, which allows the use of a mixture of amino acid models with fixed rate to estimate the appropriate model for the analysis. Analyses were stopped after 1,000,000 generations, with samplings every one hundredth generation. Tree information was summarized utilizing the “sumt burnin = 2500”, which discards the first 250,000 generations. In all cases, the measured potential scale reduction factor (PSRF), obtained using the “sump burnin = 2500” command, was equal to 1, indicating a convergence of the analysis. Amino acids models chosen by the program for each tree were: Tree 1 (Figs. 1, 2), WAG (posterior probability = 1.0); Tree 2 (Figure S1), WAG (posterior probability = 0.877) and Jones (posterior probability = 0.123). The resulting tree together containing the posterior probability for each branch was visualized using TreeView [35].
Full length transcripts encoding the septins SmSEPT5 and SmSEPT10 of S. mansoni were amplified by PCR using the following primers: SmSept5 forward primer, 5′-GCTAGCATG GCA AAT ATT CCG CGT TTT GG-3′; SmSept5 reverse primer 5′-GGATCCTCAAGACGCTTGTTGACCAGTTAC-3′; SmSept10 forward primer 5′-GCTAGCATGACTGCAGATGTTCTAAAAGCATTG-3′; SmSept10 reverse primer ACTAGCTGTACTCTCGTCAGGATCCTTATTTCC-3′ (restriction enzyme sites underlined). Reverse transcription from total mRNAs from mixed sex adult schistosomes (BH, a Brazilian strain) was accomplished and cDNA served as template for PCRs using the primers above. Amplicons of expected sizes were ligated into pTZ57R/T (Thermo Scientific), integrity of the inserts confirmed by nucleotide sequencing (3130 Genetic Analyzer, Applied Biosystems), and the septin-encoding sequences sub-cloned into the expression vector pET28a(+) (Novagen), which introduces a His-Tag at the N-terminus of the polypeptide. Recombinant septins were expressed in E. coli Rosetta (DE3) strain cells transformed with pET28 constructs, with expression induced by IPTG at 0.4 mM in LB medium for 16 h at 18°C with shaking.
Rosetta cells were lysed by sonication in 50 mM Tris-HCl pH 8.0, 800 mM NaCl, 10% glycerol, 10 mM β-mercaptoethanol (buffer A), after which lysates were clarified by centrifugation at 20,000 g for 30 min. Supernatants were fractionated by affinity chromatography on Ni-NTA resin (Qiagen) equilibrated in buffer A. Immobilized proteins were eluted in 50 mM Tris-HCl pH 8.0, 300 mM NaCl, 500 mM imidazole, 10% glycerol, 10 mM, β-mercaptoethanol (buffer B). Subsequently, eluates were separated through a column of Superdex 200 10/300 GL resin (GE Healthcare Life Sciences) fitted to a liquid chromatography system (AKTA purifier, GE Healthcare Life Sciences). Purity of eluates was assessed by Coomassie stained SDS-PAGE.
Polyclonal antibodies against recombinant SmSEPT5 and SmSEPT10 were raised in mice. Anti-septin immunoglobulins were isolated from mouse sera by affinity chromatography using immobilized recombinant schistosome septins conjugated to HiTrap NHS resin (GE Healthcare Life Sciences). Western blot analysis was employed to examine the specificity of antibodies to recombinant SmSEPT5, SmSEPT10 (each at 0.5 µM) and proteins extracted from mixed sex, adult worms (50 mg). After resolution by SDS-PAGE, proteins were transferred to nitrocellulose using the Trans-blot Semi-dry Transfer cell (Bio-Rad), 30 min at 10 V. Membranes were washed three times (15 min each) with TBS-T (10 mM Tris, 150 mM NaCl, 0.1% Tween-20) and incubated with blocking buffer (5% nonfat milk powder in TBS-T) for 2 h at 4°C. Incubation with the primary antibody, anti-SmSEPT5 or anti-SmSEPT10 diluted 1∶1,000 in TBS-T was performed for 2 h at 4°C, followed by washes as above. Thereafter, membranes were probed with goat anti-mouse IgG (whole molecule)-alkaline phosphatase (Sigma-Aldrich), diluted 1∶5,000 in TBS-T, for 2 h at 4°C, and washed as above. Signals were developed using Bio-Rad's Alkaline Phosphatase Conjugate Substrate Kit after which membranes were photographed.
Total RNA was recovered from developmental stages of schistosomes using the RNAqueous-4PCR system (Ambion). Any residual DNA in the RNA was removed by digestion with DNase (TurboDNase, Ambion). RNA concentration, purity and integrity were determined by Nanodrop 1000 spectrophotometer and Agilent 2100 Bioanalyzer; cDNA was synthesized from 150 ng RNA using the iScript cDNA Synthesis Kit (Bio-Rad). Quantitative polymerase chain reaction (qPCR) employing iQ SYBR Green Supermix (Bio-Rad), each primer at 0.3 µM in 20 µl reaction volume, was performed in a thermocycler (iCycler, Bio-Rad) fitted with real time detector (Bio-Rad iQ5). Septins-specific primers that spanned predicted exon junctions were designed as follows: SmSept5 forward primer, 5′-GGAACTGGCTTTGAGGCTATTG-3′; SmSept5 reverse primer, 5′-TGTTCTTGCATTTTACTCATTAGTTGTTG-3′; SmSept10 forward primer, 5′-CGACGTCAACGCTTAATCGA-3′; SmSept10 reverse primer 5′-CTTTAACACGCTGAACAAACATTTG-3′; SmSept7.1 forward primer, 5′-GGGTTTTGTGTTCAATCTTATGATTACT-3′; SmSept7.1 reverse primer, 5′- GATGGACCAGGATAATCAGTGTTG-3′, SmSept7.2 forward primer 5′-CGCGTTTCGATGATTACATATCTG-3′; SmSept7.2 reverse primer 5′- GGAGCAATAAAGTAAATGCATGCA - 3′. Efficiency of the PCR for each pair of septin specific primers was estimated by titration analysis to be 100%±5 [36] (not shown).The qPCRs were performed in triplicate followed an initial denaturation at 95°C for 3 min and 40 cycles of 30 sec at 95°C and 30 sec of 55°C. The specificity of the PCR product was verified by a melting curve: 1 min at 95°C, 1 min at 55°C and a ramp from 55 to 95°C with an increasing rate of 1°C/min. Absolute quantification was undertaken using copy number standards, i.e. 10-fold serial dilutions of each septin clone. Copy number of each clone dilution was calculated through the relationship between the molecular mass of the clone and the Avogadro constant. Absolute copy number of septin transcripts was estimated by interpolation of the sample PCR signals from a standard curve [36]. Biological replicates were performed. In addition, relative quantification was undertaken in order to evaluate the expression of the four septin genes within developmental stages of S. mansoni. S. mansoni glyceraldehyde 3-phosphate dehydrogenase (SmGAPDH; GenBank M92359), forward primer, 5′-TGTGAAAGAGATCCAGCAAAC-3′; reverse primer, 5′-GATATTACCTGAGCTTTATCAATGG-3′ was employed as a reference gene, with these PCRs carried out as above. The E−ΔCt method, a variation of the Livak method that incorporates the amplification efficiency values (E) for each pair of primers, was employed to determine the expression of septins relative to SmGAPDH, within each sample, i.e. each developmental stage analyzed [37]. Bioinformatics analyses were performed using RNA-seq reads from libraries of adult worms and of cercariae [32]. A tally of the RNA-seq reads aligning to the transcripts encoding the four septins was compiled based on outcomes of a blastn search, to assess relative abundance of each septin.
Developmental stages (miracidium, sporocyst, cercaria and schistosomulum) of S. mansoni were dispensed in tissue culture medium into cell culture inserts incorporating polyethylene terephthalate track-etched membranes with a pore size of 8 µm (BD Falcon, BD Biosciences, Durham NC), mounted in wells of plastic 24-well tissue culture plates. Schistosomes were fixed in 4% paraformaldehyde (PFA) by diluting 16% PFA (EMS, Electron Microscopy Sciences) in 1× phosphate-buffered saline (PBS) for 1 h at 4°C. Subsequent steps were performed on a laboratory shaker. Worms were permeabilized with Triton X-100 in PBS (0.2%) for 60 min at 25°C, followed by three washes in PBS with 0.05% of Tween-20 (PBS-T). The blocking step was carried out overnight at 4°C in PBS containing 5% normal goat serum (NGS), followed by incubation with the primary antibody (1∶50 dilution) for 2 d at 4°C. Samples were washed three times with PBS-T and incubated in 5% NGS for 20 min at 25°C as a second blocking step. Anti-mouse IgG conjugated to Alexa Fluor 633 Goat (Invitrogen) was added to the blocking solution to a final dilution of 1∶300, after which samples were incubated in the dark for 90 min at 37°C. The samples were subsequently stained for 30 min at 25°C with 4′,6-diamidino-2-phenylindole (DAPI) at 300 nM and Alexa Fluor 568 phalloidin (Invitrogen) at 165 nM in PBS containing 1% bovine serum albumin (BSA) for 30 min at 25°C. After samples were air dried, they were mounted in Fluoromount-G (EMS) on glass slides.
Confocal images were obtained using a Carl Zeiss LSM 710 system, which includes a Zeiss Axio Examiner Z1 microscope and a Quasar 32-channel spectral detector. Samples were scanned sequentially using a Plan-Apochromat 63×/1.40 Oil DIC objective. For acquisition of signals from the DAPI channel, targets were excited with a 405 diode laser line and emission was filtered in a band between 410 and 585 nm. Immunolabeling (Alexa Fluor 633) was revealed by excitation with a diode 633 laser line, with emission recorded between 638–747 nm. Phalloidin labeling of actin filaments was excited with a 561 diode laser and emission recorded from 572 to 630 nm. Optical confocal sections were generated by adjusting the pinhole to one Airy unit using the most red-shifted channel, producing an optical section of ∼0.7 µm in all channels. Confocal images were captured in sequential acquisition mode to avoid excitation bleed-through, particularly apparent with DAPI. Image frames measured 1024×1024 pixels with a pixel dimension of 0.132 µm. Images manipulation was undertaken with the assistance of Zen 2009 software (Carl Zeiss). Manipulations were limited to adjustment of brightness, cropping, insertion of scale bars and the like; image enhancement algorithms were applied in linear fashion across the image and not to selected aspects. Control images were adjusted similarly.
Four genes, Smp_041060, Smp_060070, Smp_003620 and Smp_029890, encoding putative septins were identified in the S. mansoni genome by interrogating the database in a tBLASTn search with all described human septin protein sequences as queries. Eventual discrepancies between gene predictions and actual transcript data were assessed. Utilizing the database of S. mansoni ESTs for a BLASTn search with each of four newly predicted S. mansoni septins, we observed that Smp_041060 included 5′ residues that did not correspond entirely with any EST and ESTs AM042809 and AM043866 exhibited a different 5′ terminus. These sequences were aligned to assemble a putative sequence for this transcript. Confirmation of the existence of this transcript was investigated by reverse transcription PCR utilizing primers flanking the full-length coding sequence (CDS) of the putative gene, followed by nucleotide sequencing. Sequences of the other three predicted transcripts were similarly confirmed. The sequences of the full-length CDS of SmSEPT5, SmSEPT7.1, SmSEPT7.2 and SmSEPT10 have been assigned GenBank accessions KC916723, KC916724, KC916725 and KC916726, respectively.
Sequence alignment of the four polypeptides predicted to be encoded by these transcripts using BLASTp with the 13 human septins as a database indicated that two of the S. mansoni septins, Smp_003620 and Smp_060070, displayed higher identity with the human septin SEPT7 (47% and 55% identity, respectively). Hence we termed the putative proteins SmSEPT7.1 and SmSEPT7.2. The other S. mansoni septins, Smp_029890 and Smp_041060, displayed higher identities with human septin 10 (65% identity) and septin 5 (57% identity); there have been named SmSEPT10 and SmSEPT5, respectively. The four schistosome septins have predicted molecular masses of 48–58 kDa and display hallmarks of the septin family: a highly conserved GTPase domain containing the G1 (GXXXXGKS/T), G3 (DXXG) and G4 (XKXD) motifs [38], a septin unique region [39], a variable N-terminus, and a C-terminal region predicted to form a coiled coil structure (COILS program [40]) (Figure 1).
A phylogenetic tree established using Bayesian inference of a multiple alignment of the conserved GTPase domains of the four S. mansoni septins with septins from selected deuterostomes and protostomes revealed that the orthologous S. mansoni septins SmSEPT 5, SmSEPT10, and SmSEPT 7.1 with SmSEPT 7.2 clustered in the groups SEPT2, SEPT6 and SEPT7, respectively, with strong statistical support (Figure 2). Proteins from these groups comprise a known human hetero-oligomeric septin complex [18]. Examination of the phylogenetic tree presented in Figure 2 indicated, with high posterior probability, that all the septin groups are monophyletic and their branches have an origin in the center of the tree, predicting that divergence of all septin groups preceded the protostome-deuterostome split. It is noteworthy that schistosomes, like the other protostomes sampled, lacked septins of the SEPT3 group, suggesting that this gene was lost early in this branch of evolution. The phylogram also indicated with strong statistical support that deuterostome genes encoding groups SEPT7 and SEPT6 form a monophyletic branch. This suggested that only a single copy from each of these families was present in the last common ancestor of deuterostomes and protostomes. Gene duplications that resulted in several copies of septins from group SEPT6 in humans likely occurred after the divergence of the two lineages. A peculiar scenario was evident in the SEPT2 group, in which septins of protostomes and deuterostomes did not segregate in the branch structure. This suggests that at least one event of gene duplication in this family preceded the divergence between species analyzed here. The only gene duplication observed among the four new schistosome septins was in the SEPT7 group. Curiously, a similar duplication did not occur in the orthologous human group, where only one form of SEPT7 is known [4].
Access to genome sequences of several other species of the phylum Platyhelminthes [41] facilitated analysis of putative septin sequences among flatworms at large. Phylogenetic analyses of the tapeworms Echinococcus multilocularis, E. granulosus, Taenia solium and Hymenolepis microstoma revealed the presence of four septin genes that cluster into the same groups of schistosome septins (Figure S1), indicative of a conservation of septin structures among trematodes and cestodes.
The expression profile of septins in developmental stages of the S. mansoni was investigated by quantitative PCR. At the outset, absolute quantification was employed to normalize septin expression among different developmental stages [42]–[45]. (Gene expression analysis based on normalization to a reference gene, by contrast, may be challenging in the absence of accurate information on the reference gene expression throughout the developmental stages analyzed in the present study.) The expression profile of the four septin genes exhibited similar trends among the developmental stages (Figure 3). This outcome was confirmed in a biological replicate (Figure S2).Given the propensity of septins to form hetero-filaments [18], this coordinated expression of all septin groups suggested that functional filaments of septins in schistosomes may be composed of multiple septin proteins. In order to further investigate the coordinated expression of the four septin genes, relative quantification was undertaken using SmGAPDH as a reference gene. The relative expression levels of the four septins were detected within each individual stage. In concordance to the findings with absolute quantification PCR (Figure 3), the relative abundance of the four septin transcripts was similar among the stages studied (Figure S3). Additionally, the relative contribution of each septin gene in the adults and cercarial stages of S. mansoni, was assessed by interrogation of publicly available RNA-seq data for these two stages [32]; a similar pattern of transcript abundance was apparent (Figure S4). Together, analyses by absolute and relative qPCR and of RNA-seq libraries [32] indicated that schistosome septins exhibit coordinated expression during the development cycle of the parasite.
Immunolocalization analyses were undertaken using affinity purified antibodies raised against two of the four S. mansoni septins, SmSEPT5 and SmSEPT10. At the outset, western blots were performed with two recombinant schistosome septins and lysates of adult schistosomes to ascertain the specificity of the antibodies. Minimal cross-reactivity was apparent even to excessive quantities of septins. Anti-SmSEPT5 immunoglobulin recognized SmSEPT5 strongly and SmSEPT10 weakly. In similar fashion, anti-SmSEPT10 immunoglobulin recognized SmSEPT10 strongly but only weakly recognized SmSEPT5 (Figure S5, panels A, C). Moreover, only single bands reacted in soluble lysates of adult schistosomes, incubated with anti-SmSEPT5 or anti-SmSEPT10 (Figure S5, panels B, D), indicating negligible cross-reactivity at physiological concentrations of the targets. The same trend of expression was revealed by the imaging analysis regardless of the antibody used - anti-SmSEPT5 or anti-SmSEPT10, consistent with the proposition that schistosome septins form hetero-oligomeric complexes. Representative images labeled with anti-SmSEPT5 and anti-SmSEPT10 in two developmental stages are presented in the Figures S6 and S7 to illustrate the similarity of the localization profiles. Images from samples incubated in the secondary antibody only showed minimal, though marginally detectable, signals in the septin channel (Figure S8).
In parallel to the immunolocalization with anti-septin antibodies, filaments of schistosome actin were labeled with phalloidin, a ligand of fungal origin that binds actin polymers. Two- and 14-day-old schistosomula exhibited similar immunofluorescence profiles (Movies S1, S2). Septin fibers could be identified at the worm surface, along with actin fibers (Figure 4). Septins were also evident in deeper layers, usually at the periphery of the cells (Figure 5). Three-dimensional renderings of confocal images including signals from the DAPI, phalloidin and septin probes revealed ubiquitous expression of septins (Figure 6). By contrast, actin was more prominent in muscle layers and the gut (Figure 6B).
Miracidia and sporocysts cultured for two days were permeabilized and probed with anti-septin immunoglobulins, followed by incubation with an Alexa Fluor 633-conjugated secondary antibody. Confocal imaging allowed the sampling of increasingly deeper (internal) layers of these stages of the blood fluke which, in turn, precisely localized septins in organs and tissues (Movie S3; Figures S9, S10). The superficial layers of the miracidium expressed septins on the epidermal plates (Figure 7), which contain the cilia of the motile larva. Images of superficial layers of the sporocyst revealed colocalization of actin and septin in the longitudinal and circular muscle layers (Figure 7D). Moreover, septins were prominent in optical sections of germs cell in miracidia (Figure 8A) and two-day-old sporocysts (Figure 8B). Colocalization of septin and actin was observed in the superficial optical sections of these larval stages, though not in deeper sites.
Robust staining of septin was seen in the cercaria along the protonephridial ducts that extend laterally down both sides of the larva; characteristic flame cells are located at the termini of the canals (Figure 9, arrows; Movie S4). The protonephridium consists of a network of osmoregulatory tubules comprising the excretory system of the larva and extends nearly the entire length of the body [46], [47]. Septins appear to occur at the collecting tubules of this osmoregulatory system (Figure 9); the role of septins in these structures remains to be elucidated.
Septins are cytoskeleton components formed from hetero-oligomeric complexes which assemble into higher-order structures such as filaments and rings. Whereas septins are evolutionarily conserved and widely distributed among eukaryotes, until now septins have not been reported in schistosomes or indeed in any flatworm. Four schistosome septins are reported here; they are members of three discrete septin groups, with a duplication verified in SEPT7 group. A similar gene organization was apparent in the genomes of four cyclophyllidean tapeworms including E. multilocularis. The conserved gene organization between trematodes and cestodes indicates an essential role of these proteins across the phylum.
The flukes and tapeworms investigated displayed members of the same three septin groups described previously in some other invertebrates, including Drosophila melanogaster and Caenorhabditis elegans. The SEPT3 group appears to be absent from these taxa [20], [48], suggesting that a deletion occurred in a common ancestor. Despite the absence of one of the groups, the functional assembly of septins into heter-oligomeric structures is maintained [49], [50]. Phylogenetic analysis involving nine metazoans [20], including the sea urchin Strongylocentrotus purpuratus and the sea squirt Ciona intestinalis which, as representatives of the echinoderms and non-vertebrate chordates, respectively, occupy informative phylogenetic positions with respect to the evolution of mammals, revealed that they share orthologues of all human septin groups. This suggests that these four septins groups evolved before the appearance of the vertebrates and supports the hypothesis of a deletion event in some lineages of the invertebrates. Moreover, schistosomes exhibited only one isoform for groups SEPT2 and SEPT6 in contrast to the situation in humans where several isoforms of the groups occur. Septin configuration in schistosomes may resemble an ancestral arrangement, whereas successive duplications and mutations have endowed the human isoforms with a more specialized role(s).
Real-time PCR analyses indicated that transcription of genes encoding the four schistosome septin was maintained at approximately the same ratio throughout the developmental cycle. This suggested that a septin heterocomplex might exhibit a similar composition in the diverse morphological stages of the schistosome, although deeper investigation will be required to precisely define the proportion of the different group members of the septin heterocomplex in S. mansoni. Confocal imaging revealed that muscle layers were rich in actin and septin, suggesting cell specific co-expression of these cytoskeleton elements. Their co-localization was more evident in the sporocyst, which undergoes shedding of ciliated epidermal plates and the emergence of a new tegument during infection of the snail [51]. Co-localization of actin and septin is well-known in mammals [52], [53]. Rapid-freeze, deep-etch immuno-replica electron micrographs reveal associations of septins with actin-based membrane skeletons of kidney cells [54]. Septins can be recruited to an actin bundle through the interaction with the adaptor protein anillin of Xenopus laevis [53]. BLASTp searches using X. laevis anillin allowed the identification of a homologue from S. mansoni (XP_002576415.1; 43% similar over 177 positions). Analysis using the CDD tool at NCBI (not shown) revealed an anillin pleckstrin homology (PH) domain (cd01263) in the schistosome orthologue, suggesting that it performs a similar role.
Confocal immunolocalization micrographs revealed septins in the ciliated epidermal plates of the miracidium. Septins are required for ciliogenesis and constitute a diffusion barrier at the base of the ciliary membrane in mammalian cells and Xenopus embryos [49], [55]. The cilium of the miracidium exhibits microtubules with a 9+2 pattern [56], typical of eukaryotic motile cilia [57]. Since the cilium is a highly conserved organelle in eukaryotes [57], septins in the epidermal plates of the miracidium likely display a similar organization to the septins found in cilia of mammalian cells, and may perform a similar role. The prominent expression of septins in germ cells of both the miracidium and the mother sporocyst and absence of phalloidin staining in this tissue together indicated that actin and septins of schistosomes do not always act in concert. In like fashion, earlier reports indicate that actin was not detectable in germ cells when miracidia and sporocysts were probed with phalloidin [47], [58], [59]. A pattern of staining for β-tubulin in the germ cells has been reported [48] that is similar to the localization of septins in germ cells described here, which together suggests a cellular co-expression of these filaments as well. Associations of septins and microtubules are well known [60]–[63]. In D. melanogaster, Peanut, SEPT1 and SEPT2 have been identified in male germ cells [64]. Likewise, septins are known from the mammalian germ cells. Sept4 null male mice are sterile due to immotile spermatozoa with defective annulus [65], [66], and diminished expression of SEPT12 transcripts is evident in the testicles of infertile men [65]. Moreover, septin 12 plays a key role in terminal differentiation of germ cells in both humans and mice [67]. Septins are best known for their role in cytokinesis [9]–[14] and we speculate that the role of septins in miracidial and sporocyst germ cells might be related to the mitotic activity of these cells.
Schistosomula cultured for two or 14 days showed ubiquitous localization of septin in contrast to a more restricted distribution in the other stages. Ubiquitous septin localization has also been reported for some human septins [68]; it is feasible that septins are involved in some similar functions in human and schistosome cells and tissues. To conclude, this is the first description of septins of a schistosome or any platyhelminth. Four septins were identified and they were differentially expressed among developmental stages of the blood fluke. Confocal imaging indicated that schistosome septins undertake specialized roles in this pathogen. Detailed evaluation of schistosome septins can be expected to clarify the relationship of this category of proteins with cellular and physiological functions and to deliver deeper understanding of schistosome physiology and anatomy and its roles in the host-parasite relationship.
|
10.1371/journal.ppat.1000096 | Vaccinia Virus E3 Protein Prevents the Antiviral Action of ISG15 | The ubiquitin-like modifier ISG15 is one of the most predominant proteins induced by type I interferons (IFN). In this study, murine embryo fibroblast (MEFs) and mice lacking the gene were used to demonstrate a novel role of ISG15 as a host defense molecule against vaccinia virus (VACV) infection. In MEFs, the growth of replication competent Western Reserve (WR) VACV strain was affected by the absence of ISG15, but in addition, virus lacking E3 protein (VVΔE3L) that is unable to grow in ISG15+/+ cells replicated in ISG15-deficient cells. Inhibiting ISG15 with siRNA or promoting its expression in ISG15−/− cells with a lentivirus vector showed that VACV replication was controlled by ISG15. Immunoprecipitation analysis revealed that E3 binds ISG15 through its C-terminal domain. The VACV antiviral action of ISG15 and its interaction with E3 are events independent of PKR (double-stranded RNA-dependent protein kinase). In mice lacking ISG15, infection with VVΔE3L caused significant disease and mortality, an effect not observed in VVΔE3L-infected ISG15+/+ mice. Pathogenesis in ISG15-deficient mice infected with VVΔE3L or with an E3L deletion mutant virus lacking the C-terminal domain triggered an enhanced inflammatory response in the lungs compared with ISG15+/+-infected mice. These findings showed an anti-VACV function of ISG15, with the virus E3 protein suppressing the action of the ISG15 antiviral factor.
| Modification of proteins by ubiquitin (UB) and ubiquitin-like proteins (UBL) represents a key regulatory process of innate and adaptive immune responses. Interferon-stimulated gene product 15 (ISG15) is a member of UBL molecules that can reversibly be conjugated to proteins mediating considerable antiviral response. In turn, many viruses, including poxviruses, have evolved strategies to block the antiviral and inflammatory effects of innate immune responses to keep cells alive until virus replication is completed. Here, a novel viral immune evasion mechanism that inhibits ISG15-dependent antiviral pathway is described. Vaccinia virus (VACV) pathogenesis in ISG15+/+ versus ISG15−/− mice is linked to the virus E3 protein, blocking the activity of ISG15 through its C-terminal domain. This effect was independent of PKR activation. ISG15 controls the inflammatory response regulating cytokine levels. Our findings support a new strategy for poxviruses to evade the host antiviral response through interaction of the virus E3 protein with ISG15.
| Type I interferons (IFN-α and -β) serve a critical role in antiviral innate immunity and in modulating the adaptive immune response to infection and tumor development [1]. In response to infection or Toll-like receptor agonists, IFN is produced and consequently leads to the up-regulation of hundreds of IFN-stimulated genes (ISG) [2],[3]. One of the most highly induced genes is ISG15 that encodes a small UBL protein of 17 kDa that forms covalent conjugates with cellular proteins [4]. ISG15 is composed of two domains, each of which carries high sequence and structural similarity to UB (33 and 32% for the N- and C-terminal domains, respectively) [5],[6].
ISG15 conjugation (ISGylation) to substrate proteins occurs in a manner similar to UB conjugation by utilizing activating, conjugating and ligating enzymes to facilitate the addition of ISG15 to specific lysine residues [7]. The ISG15 activating enzyme is ubiquitin E1 like protein (UBE1L), and the E2 enzyme for UB conjugation, UbcH8, also recognizes ISG15 [8],[9],[10]. ISG15 is removed from conjugated proteins by an ISG15-specific protease, UBP43 (USP18 (UB-specific protease 18)) [11],[12],[13]. UB as a central cellular regulator and UB-mediated proteolysis also plays a regulatory role in the immune system [14],[15]. While the degradation by the proteosome generally depends on poly-UB conjugation, protein modification by ISG15 does not typically cause substrate degradation [16]. Instead it may alter the subcellular localization, structure, stability or activity of targeted proteins [17]. A large number of cellular proteins that are associated with cellular cytoskeleton, stress response and chromatin remodelling were identified as ISG15 targets. ISG15 also targets proteins that play a role in the innate antiviral response, including: PKR, MXA, STAT1, JAK1 and RIG-I [18]. ISGylation of these antiviral molecules may regulate their activity during viral infection.
ISG15 expression is almost undetectable under normal conditions but is strongly up-regulated during viral infections such as human cytomegalovirus (HCMV), herpes simplex virus (HSV), Sindbis virus (SV) and hepatitis C virus (HCV) [19],[20],[21],[22],[23],[24]. It has been speculated that the ISG15 up-regulation following viral infection is involved in different strategies of the antiviral response [25],[26]. Some viruses have developed specific strategies to counteract the activity of the IFN-stimulated genes (ISGs). The influenza B virus protein NS1B binds ISG15 and blocks protein ISGylation [27]. Furthermore, constitutive expression of ISG15 in type I IFN receptor knockout (KO) mice confers potent antiviral activity against SV. This evidence suggests that ISGylation is important for protecting cells from viral infection [21].
Previously, using cDNA microarrays we described up-regulation of ISG15 after infection of HeLa cells with the attenuated VACV strains MVA and NYVAC, an effect not observed after infection with the virulent strain WR [28],[29],[30]. Furthermore, the attenuated mutant VVΔE3L that lacks the viral early protein E3 also produces an increase in the ISG15 mRNA levels [31]. VVΔE3L is a virus that only replicates in IFN-incompetent systems exerting IFN antagonist activity [32], is nonpathogenic in the mouse model, and provides protection against a wild-type virus challenge [33],[34]. The E3 protein represses the host cell antiviral response by multiple mechanisms, including inhibition of both PKR and RNase L, two enzymes induced by IFN and whose activation triggers a global inhibition of protein synthesis and virus replication [35],[36],[37] through the phosphorylation of eIF-2α (for PKR) and breakdown of RNA (for RNase L). Significantly, once activated both PKR and RNase L produced upregulation of ISG15 messenger levels [38],[39]. E3 also blocks induction of IFN-α/β through inhibition of phosphorylation of the transcription factors IRF3 and IRF7 [40],[41] and prevention of NF-κB activation [42].
The biological significance of ISG15 mRNA induction in cultured cells after infection with the VACV mutants and its repression by the virulent WR is not known. Here we have investigated the role of ISG15 as an anti-VACV immunity factor using in vitro and in vivo systems based on MEFs and mice lacking ISG15. While in MEFs the yields of WR were slightly different between ISG15+/+ and ISG15−/− infected cells, the non-replicating VVΔE3L in ISG15+/+ cells grew more than one log better in ISG15−/− cells. Biochemical analyses showed that the E3 protein interacts with ISG15 through its carboxyl terminal domain. Repression of ISG15 with siRNAs or expression of ISG15 by a lentivirus vector in ISG15 null cells indicate that VACV replication can be controlled by ISG15 and that E3/ISG15 protein interaction is independent of the presence of PKR. In mice lacking ISG15, VVΔE3L induced stronger pathogenesis than in ISG15+/+, an effect similarly triggered by a C-terminal deletion mutant (VVE3LΔ26C). Our findings reveal a novel VACV strategy to counteract the IFN antiviral response through interaction of the virus E3 protein with ISG15.
We and others have previously described upregulation of ISG15 transcript from gene expression profiles of HeLa cells infected with the attenuated VACV strains, VVΔE3L [31], MVA [28] and NYVAC [30]. This up-regulation was not observed in HeLa cells infected with the virulent WR [29]. Here, we have validated the transcriptional changes in ISG15 mRNA levels after VACV infection by real-time RT-PCR. As shown in Table S1, ISG15 mRNA levels at different times postinfection (p.i.) were enhanced in HeLa cells infected with the mutant viruses compared to the virulent WR, in agreement with the microarray data (not shown).
To correlate changes in ISG15 protein levels, we analyzed by immunoblot the levels of ISG15 in WR- or MVA- or NYVAC- or -VVΔE3L or uninfected MEFs. In agreement with the results of real time RT-PCR obtained in Hela cells, a clear increase in ISG15 protein levels was also observed in VVΔE3L- or MVA-infected MEFs cells at 6 and 16 hpi (Fig. 1A). The increase was less apparent after NYVAC infection probably because overall protein synthesis is more severely inhibited by NYVAC than MVA [30]. Moreover, the increase of ISG15 protein levels after VVΔE3L or MVA infection required de novo protein synthesis as its accumulation was prevented by cycloheximide treatment discarding the possibility that infection might increase ISG15 protein levels by enhancing protein stability (not shown). It should be noted that there is an increase in the conjugation of ISG15 to its target proteins after VVΔE3L, but reduced in levels after MVA infection (Fig. 1A). The findings of Fig. 1 establish a clear up-regulation of ISG15 by the attenuated VACV mutants.
Since the increase in ISG15 in Hela cells correlated with the attenuated phenotype of several VACV strains, we next examined the role of ISG15 in VACV replication using primary MEFs derived from ISG15+/+ and ISG15−/− mice. While the cytopathic effect (CPE) observed in ISG15−/− after WR infection (0.1 PFU/cell, 24 h) was similar to ISG15+/+ cells (Fig. 2A, upper panels), the CPE in ISG15−/− cells after VVΔE3L infection was markedly increased with respect to that observed in ISG15+/+ cells (Fig. 2A, lower panels). The virus titers for WR were slightly increased in ISG15−/− compared to ISG15+/+ cells (Fig. 2B), while the yields of VVΔE3L were increased in the ISG15−/− compared to ISG15+/+ cells (about 25-fold higher). The increase in virus titers correlated with increase in cellular mortality, as shown in Fig. 2A. The findings of Fig. 2 suggest that E3 expression might be suppressing ISG15 function. To define the breath of the E3 anti-ISG15 activity we analyzed the role of another antiviral factor, PKR, using MEFs derived from PKR−/− mice. Both the difference in CPE and virus yields between VVΔE3L infected ISG15−/− and PKR−/− cells were clearly distinct (Fig. 1A), indicating that the in vitro replication of VVΔE3L in ISG15−/− is a process independent of PKR. With PKR+/+ cells the CPE and virus yields were similar as for ISG15+/+ cells (not shown).
To provide further evidence for a VACV antiviral role of ISG15, we used siRNA to specifically block ISG15 mRNA production. Using siPORT Amine as a transfection reagent, MEFs were transfected with two specific ISG15 siRNAs (siRNA1 or siRNA2), or with a specific GAPDH siRNA (positive control) or with a scrambled siRNA (negative control). Twenty four hours after transfection cells were infected with WR or VVΔE3L (0.1 PFU/cell), and ISG15 expression, CPE and virus titers were evaluated during the course of infection. As shown in Fig. 3A, the two ISG15 siRNAs decreased the expression of ISG15 by over 80% after 24 h of transfection (Fig. 3A). The decrease in ISG15 protein levels was accompanied by an enhanced CPE in VVΔE3L infected ISG15+/+ cells (Fig. 3B); the difference in CPE was less clear in WR-infected cells. In addition, viral titers were enhanced in silenced ISG15+/+ cells infected with WR or VVΔE3L (Fig. 3C). We also performed ISG15 mRNA inhibition with ISG15 siRNAs in PKR−/− cells and found no changes in CPE and virus yields for WR and VVΔE3L infected cells in comparison to the results obtained in PKR+/+ cells (not shown), indicating that the function of E3 protein is independent of PKR activity.
Transfection of GAPDH siRNA or a scrambled siRNA followed by infection with WR or VVΔE3L had no significant effect in either ISG15 protein level, CPE or virus production, indicating the specificity of ISG15 function. We checked that GAPDH protein levels were decreased only in the siRNA-GAPDH transfected cells, as measured by Western blot analysis (not shown). The results of Fig. 3 revealed that suppression of ISG15 protein levels leads to enhanced replication of WR and of VVΔE3L, further supporting an anti VACV role of ISG15.
To further extend the role of ISG15 expression in VACV replication, we tested whether ectopic ISG15 expression in ISG15−/− cells leads to inhibition of WR or VVΔE3L viral growth. Using pRVISG15-ires-GFP, an optimized retroviral vector that expresses efficiently ISG15 in transduced cells, we evaluated the CPE (see Fig. S1) and viral growth of WR or VVΔE3L. Viral titrations showed that retroviral transduction of the ISG15 gene in ISG15−/− cells result in inefficient VVΔE3L viral production, whereas non-transduced ISG15−/− cells infected with VVΔE3L were able to produce infectious viral particles (Fig. 4C, left panel). Furthermore, yields of VACV infectious virus decreased in the ISG15-transduced cells in comparison to those that do not express ISG15 (Fig. 4C, right panel). These findings demonstrate that the absence of ISG15 is essential for the productive infection of VVΔE3L and the increase in VACV production in murine cultured cells.
To characterize the effect of ISG15 over-expression in WR or VVΔE3L replication in cells with endogenous ISG15, a similar approach was carried out as before but with retroviral transduction of ISG15+/+ cells, followed by measurements of WR or VVΔE3L viral growth. As shown in Fig. 4, panels C–D, both viruses showed a decrease in infectious virus production in correlation with higher ISG15 expression levels. This experiment supports that ISG15 is a negative regulator of VACV replication.
It has been described that the influenza B virus protein NS1B inhibits ISGylation after binding through its amino terminal domain to ISG15 [27],[43]. To test whether E3 protein, that contains a similar domain to NS1, is able to bind to ISG15 protein, we performed immunoprecipitation (IP) assays in PKR+/+ cells using the following different viruses: WR, MVA, VVΔE3L lacking the entire E3L gene, and two deletion mutants VVE3LΔ83N and VVE3LΔ26C with truncated versions of E3L gene at the N and C-terminus [44]. After ISG15 IP, the entire E3 protein binds efficiently to ISG15 and the N-terminal mutant Δ83N, that lacks 83 aminoacids and the PKR binding domain, binds efficiently to ISG15 (Fig. 5A). In contrast, the C-terminal mutant Δ26C, that lacks 26 aminoacids and the ability to bind dsRNA, does not bind ISG15 (Fig. 5B). We also performed the reverse IP using an anti-E3 antibody and only the E3 protein that lacks the C- terminus fails to be immunoprecipitated (Fig. 5B, left panel). When IP was performed without antibody or using a pre-immune serum as a control, no interaction was observed (Fig. 5B, central panels). To analyze if RNA was involved in E3/ISG15 interaction, we treated the IP complex with RNase just before its loading in the SDS-PAGE, and found that the complex was destroyed, as no interaction was observed with any of the E3 proteins from the different viruses (Fig. 5B, left panel). This result indicate that RNA, and probably dsRNA, has a role as a linking component in the interaction between E3 and ISG15 proteins as its degradation abolishes the binding of both proteins. ISG15/E3 protein interaction was confirmed by confocal microscopy, as WR-, or MVA- or VVE3LΔ83N-infected MEFs showed co-localization between ISG15 and E3, while VVΔE3L- or VVE3LΔ26C-infected MEFs did not (Fig. 5C). In addition, we also studied if the presence of PKR was relevant for this interaction by performing both IP and confocal experiments in PKR−/− cells. Both approaches indicate that the interaction between ISG15 and E3 is independent of PKR, as in its absence the entire E3 and the protein that lacks the amino terminus are able to interact with ISG15 (Fig. 5D). The findings of Fig. 5 reveal that ISG15 binds the E3 protein in a PKR-independent manner and that binding requires the C-terminal domain of E3 spanning the RNA-binding site, which suggests that dsRNA acts as a linker.
To further show that ISG15 is a biological relevant antiviral molecule against VACV infection, we next evaluated the susceptibility of ISG15−/− mice to the virus. Thus, we infected by the intraperitoneal (i.p) route ISG15−/− or ISG15+/+ mice with WR at 2×107 or with the attenuated VVΔE3L at 108 PFU/mouse and scored for prominent indicators of viral pathogenesis (weight loss and mortality). While there was reduced weight loss in ISG15−/− mice, survival was similar between both groups (Fig. 6, panel A and B). However, after VVΔE3L infection the ISG15−/− mice displayed signs of disease within 2 days, characterized by ruffled fur and lack of activity, and 25% of the animals died within 1 to 2 days. Half of the mice infected with VVΔE3L appeared sick at 4 days p.i., and 75% recovered after 7 to 8 days p.i. (Fig. 6C–D). We did not observe virus yields for ISG15−/− or ISG15+/+ mice infected with VVΔE3L in liver or spleen, while virus titers were easily obtained in WR infected mice (not shown).
Since the inflammatory response might explain the rapid signs of illness in VVΔE3L infected ISG15 KO mice, we measured serum cytokine levels (IL-6, TNF-α, IL-10, MCP-1, IFN-γ, and IL-12 p 70) at early times post infection. In ISG15+/+ mice, IL-6 levels were similar in serum from WR- or VVΔE3L- or mock-infected mice (Fig. 6E). In contrast, ISG15−/− mice infected with VVΔE3L had an 8-fold increase in serum levels of IL-6 compared with those infected with WR (P<0.01; Fig. 6E). There were no changes in levels of other cytokines analyzed between the groups (not shown).
We also examined the extent of protection conferred in animals pre-immunized with WR or VVΔE3L by i.p route. Thus, pre-immunized mice (as in Fig. 6) were challenged by i.n route with WR at 2×107 PFU/mouse. In the case of ISG15+/+ and ISG15−/− mice pre-immunized mice with WR, the challenge had little effect on weight loss and signs of illness, clear signs of protection (Fig. 7A–B). However after WR challenge, VVΔE3L pre-immunized ISG15−/− mice developed weight loss and signs of sickness which was not observed in ISG15+/+ mice (Fig. 7C–D). These findings revealed a reduced protection to challenge with WR conferred by VVΔE3L pre-immunized KO mice, indicating limited adaptive immune response triggered by VVΔE3L infection of ISG15 −/− mice.
Since VVΔE3L pre-immunized ISG15 KO mice developed disease transiently after i.n WR challenge and the upper respiratory tract is a natural route for variola virus infection, we next evaluated disease progression in ISG15+/+ and ISG15−/− mice after i.n inoculation with WR and several E3L deletion mutants (VVΔE3L; VVE3LΔ83N and VVE3LΔ26C). The highly attenuated MVA strain was included as control. Infected mice were scored for prominent indicators of viral pathogenesis (weight loss, signs of illness and mortality). After WR or VVE3LΔ83N infection, no significant differences in weigh loss and signs of illness were observed between ISG15−/− and ISG15+/+ mice, although slight differences in weight loss were observed between both groups of mice when inoculated with a lower dose of WR (Fig. 8A, upper panel). Mortality was higher in mice infected with WR, as all mice died within 7 days in the case of WR, while infection with VVE3LΔ83N caused 25% mortality (Fig.8A–B). However, clear differences were observed after i.n. inoculation of VVΔE3L or VVE3LΔ26C. While ISG15+/+ mice did not show signs of illness at any times p.i, ISG15−/− mice infected with VVΔE3L or VVE3LΔ26C showed disease as revealed by clear signs of illness as soon as 2 days and 25% of the animals died. About half of the mice infected with VVΔE3L or VVE3LΔ26C appeared sick at 4 days p.i, but 75% of them recovered after 7 to 8 days p.i, (Fig. 8C).
To analyze the status of ISGylation in the infected mice, lungs were homogenized and conjugation of ISG15 to its target proteins was determined by Western blot. While, ISG15−/− mice do not express ISG15 (Fig. 7A) and lungs homogenates from ISG15+/+ mice had similar amounts of the ISG15 protein, conjugation of ISG15 to its targets proteins is enhanced in lung extracts from mice infected with VVΔE3L or VVE3LΔ26C (Fig. 9A). Similar result was also observed in MEFs infected in vitro with VVΔE3L where high levels of ISG15 conjugates were clearly observed (Fig. 1A). These findings suggest that E3 blocks conjugation of ISG15 to its target proteins by its carboxy-terminal domain.
The presence of VACV proteins, as determined by Western blot, was more evident in lungs of ISG15−/− in comparison to ISG15+/+ mice infected with WR or the deletion mutant viruses (Fig. 9B). As expected, appearance of virus in lungs correlated with the presence of viral proteins in these tissue extracts (Fig. 9B–C). These results indicate that although the absence of ISG15 has no effect in the mortality of the mice at a high dose of WR inoculation (5×106 PFU/mice; Fig. 8A, middle and lower panel), it has an effect in the replication of the WR and E3L mutant viruses, as seen by the different amount of viral protein and virus titers in lung tissues of ISG15−/− versus ISG15+/+ mice (Fig. 9B–C).
Histological examination of lung tissue showed that ISG15+/+ animals infected with the different mutant viruses had no inflammatory cells infiltrating the lung parenchyma. In contrast, lung sections obtained from ISG15−/− mice infected with VVΔE3L or VVE3LΔ26C presented severe inflammation with alveolar wall thickening and infiltration of inflammatory cells (see enlarged sections in Fig. 10). This phenotype was not observed in WR- or VVE3LΔ83N- infected ISG15−/− mice (Fig. 10). This result indicates that in ISG15−/− mice, pathogenesis and development of an inflammatory response is mediated by the absence of E3 virus expression. This phenotype was maintained after VVΔE3L and VVE3LΔ26C pointing to E3 as a major VACV molecule involved in virus evasion of the IFN-defense ISG15 antiviral protein.
Pro-inflammatory and IFN-stimulated genes (ISGs) represent essential components of the innate immune response to viral infection (40). Upon viral entry into cells, ISG induction occurs in two waves: acute, IFN-independent induction of a subset of ISGs and delayed, IFN-dependent induction via the production of IFN-α/β during the initial phase. In many viral infections, IFN-independent ISG induction is mediated by the IRF-3 phosphorylation, homodimerization, and nuclear translocation. Activated IRF3, in turn, induces the expression of type I IFN genes, whose products trigger strong induction of a subsets of ISGs, including IFN-β which after its release and ligand-binding to its receptor then initiates IFN-dependent ISG induction via the IFN receptor and JAK/STAT signaling pathways. IFN-inducible enzymes, like the 2.5 OAS/RNAse L system, PKR, and M×, are the best characterized proteins that mediate antiviral action of IFN. Another protein, ISG15, was first identified as an IFN-stimulated gene whose expression is induced strongly by IFN-α/β treatment and can be detected at low constitutive levels in cells [45]. ISG15 modifies several important molecules and affects type I IFN signal transduction; ISG15 expression is markedly increased following viral infection (14, 30, 49), and many viruses encode inhibitors of the IFN-transduction pathway or specific inhibitors of ISG to avoid deleterious effects triggered by these cytokines.
Among animal viruses, the poxvirus family contains a large array of genes which are used by the virus to evade host immune responses. VACV encodes multiple proteins that interfere with complement regulatory proteins, with many cytokines and chemokines, with TLRs (Toll like receptors) and signal transduction pathways, with apoptosis, and others [46]. One of the VACV proteins with strong inhibitory activity of IFN-induced pathways is E3 [47]. E3 represses the host cell antiviral response inhibiting both PKR and RNaseL, which trigger global inhibition of protein synthesis and virus replication [35],[36],[37]. In addition, E3 blocks the activation of IRF3 [40],[41], and effectively prevents the first wave of type I IFN synthesis. E3 has two domains, an N-terminal involved in the direct inhibition of PKR, its nuclear localization, and Z-DNA binding [34],[48],[49],[50], and the C-terminal that contains the dsRNA-binding domain required for IFN-resistance and for the broad host range phenotype of the virus [44],[51]. It has been described that VACV lacking E3 (VVΔE3L) replicates in PKR or RNaseL deficient cells [40].
Through the use of microarrays we identified the gene ISG15 as being induced in the course of infection of human cells with different strains of VACV [28],[29],[30]. The attenuated mutant VVΔE3L also produces an increase in ISG15 messenger levels [31]. The reason for the induction of ISG15 mRNA levels by attenuated viruses (Fig. 1) is probably due to the activation of several cellular signal transduction cascades and of host transcription factors [28],[30],[31]. Since MVA and NYVAC strains contain the E3L gene, this upregulation may be independent of E3 expression with induction being likely due to the increase in IFN-β levels.
In this study we showed that a VACV mutant lacking E3, which cannot grow in ISG15 WT cells, is able to replicate both in MEFs cells derived from ISG15 KO mice or in ISG15 silenced cells. In addition viral titers also increase in the absence of ISG15 indicating that ISG15 has an essential role against infection of VACV. During infection of MEFs from ISG15 KO or ISG15 depleted cells, the presence of E3 enhances viral production since the WR titers are greater than those after infection with the VVΔE3L mutant virus. One explanation of this phenotype is that the mutant virus lacking E3 triggers apoptosis through PKR activation which, in turn, reduces virus production as previously described [52]. The role of ISG15 in VACV replication was also supported by the more abundant VACV infectious virus and viral proteins produced in lungs of ISG15−/− mice compared with lungs of ISG15+/+ mice after infection with WR or with the E3L deletion mutant viruses (Fig. 9B–C).
While the depletion of ISG15 has an effect on VVΔE3L mutant phenotype and restores virus growth, over-expression of ISG15 in ISG15−/− murine cells using a retroviral transduction system revert the restricted VVΔE3L viral growth. Over-expression of ISG15 also reduces markedly WR titers reinforcing the idea that ISG15 plays a role in the control of VACV replication.
Inhibition of ISG15 function by VACV is likely due to its interaction with VACV E3, as shown by IP and confocal analyses. This interaction with ISG15 occurs independently of PKR, through the C-terminal region of E3 and requires RNA. We have shown that ISG15 controls the in vitro replication of VACV in a PKR-independent manner, as WR and VVΔE3L titers do not increase in murine PKR−/− cells in comparison to those observed in PKR+/+. While in murine cells VVΔE3L is able to replicate in a PKR-independent manner, as also described in MEFs lacking RNase L [53], in human HeLa cells with PKR expression suppressed by siRNA, the mutant virus is able to grow [40]. The differences in cell origin might explain the distinct effect of the IFN system in the control of VACV replication.
The mutant VVΔE3L virus that was able to replicate in ISG15−/− MEFs (Fig. 2) did not replicate in ISG15 KO mice (Fig. 7), but surprisingly infection with VVΔE3L provokes sickness and mortality only in ISG15−/− mice. This was probably related to the strong inflammatory response triggered by the mutant in ISG15 KO mice, as observed by the increased levels of IL-6 in serum (Fig. 4E). Although the biological relevance of this observation remains to be established, it can be suggested, in view of the functions assigned to ISG15 in the innate immune response [54], that this molecule plays a role as regulator of IFN-triggered innate responses during VACV infection. It will be of interest to know the type of innate response triggered in ISG15 KO mice infected with VVΔE3L.
The inability of VVΔE3L to cause significant disease in WT mice is presumably due, at least in part, to induction of type I IFN that, in turn, leads to up-regulation of antiviral proteins, such as PKR and 2–5 OAS/RNaseL system. The mutant virus that lacks the C terminus of E3L gene involved in the dsRNA sequestration (VVE3LΔ26C) is completely attenuated in WT mice; however deletion of the N terminus (VVE3LΔ83N) reduces pathogenesis 500- to 5,000-fold [33],[48]. We extended these in vivo studies using the i.n route and compared WR and E3L mutant viruses in WT and ISG15 KO mice. We observed that only after VVΔE3L or VVE3LΔ26C inoculation (i.n), the mortality of mice was increased by the absence of ISG15 (Fig. 6). In the case of WR or VVE3LΔ83N, there were no differences in mortality of both viruses in ISG15 KO in comparison to WT mice although viral replication was enhanced in the lungs of ISG15−/− in comparison to ISG15+/+ mice. One explanation is that ISG15 is made non functional after infection with these viruses, because the carboxy-terminal domain of E3 binds to ISG15 and blocks its activity. These observations are in correlation with the reduced presence of conjugates in lungs of WT mice infected with WR or VVE3LΔ83N, compared with infection by VVΔE3L or VVE3LΔ26C. Although lung homogenates presented similar amounts of ISG15 protein, the conjugation of ISG15 to its targets proteins, was greatly enhanced after infection with VVΔE3L or VVE3LΔ26C (Fig. 7). This evidence suggests that inhibition of conjugation of ISG15 is mediated by E3 and this inhibition requires the presence of the dsRNA binding domain. Similar result was also observed in MEFs infected in vitro with VVΔE3L where high levels of ISG15 conjugates were clearly observed (Fig. 1A). The cause of mortality of ISG15 KO mice after infection with VVΔE3L or VVE3LΔ26C was a massive inflammation of lungs with alveolar wall thickening and infiltration of cells (Fig. 8). These results indicate a role of ISG15 in the control of an inflammatory response by regulating cytokine levels.
Cytokine and chemokine release occurs rapidly in response to virus infection, with the aim of recruiting inflammatory leukocytes in order to limit virus replication and spread, and to induce adaptive immunity. However, prolonged expression of chemokines in the context of viral infections may be detrimental to the host. We find that in the absence of ISG15, infection with VVΔE3L produces an increase of IL-6 that correlates with short-term morbidity and complications that include pulmonary function abnormalities. Although the mechanisms of this up-regulation remains to be established, it can be speculated in view of the functions assigned to ISG15 that it might be involved in the regulation of cytokine signal transduction, through the stabilization of specific signalling components that facilitate the development of a correct innate immune response. In this sense a family of intracellular proteins called suppressors of cytokine signalling (SOCS) are essential for the regulation of cytokine expression having a critical role in the regulation of the innate response. Considering that SOCS-1 and SOCS-3 negatively regulate the IFN-induced signal cascade, and VACV E3 protein inhibits the type I IFN response, it is possible that E3 or other viral proteins may regulate the IFN response by affecting SOCS protein expression regulating the ISG15 activity by an unknown mechanism. We have previously demonstrated that although WR provokes a general downregulation of cellular mRNAs, there are a discrete number of human genes that are induced selectively during the course of VACV infection. A variety of these upregulated genes encode different members of the SOCS family [29] indicating that probably VACV may modify SOCS protein expression to manipulate the cytokine pathway and the antiviral host response. This strategy may be used to reduce the efficacy of innate and acquired immune responses to infection. However, WR modification of cytokine or chemokine responses may also be a mechanism to recruit new targets for infection, or provide new niches for infection. It has been described that over-expression of HCV core protein inhibits IFN signalling and induces SOCS-3 expression. SOCS-1 and SOCS-3 proteins have been reported to inhibit IFN-induced activation of the JAK-STAT pathway and expression of antiviral proteins, such as M×A [55].
There are similarities between the functions of VACV E3 and the NS1 dsRNA-binding protein of influenza virus. NS1 blocks IRF3 phosphorylation and IFNβ mRNA induction [56]. In addition, NS1 is an inhibitor of PKR, suggesting that dsRNA sequestration is a strategy used by both RNA and DNA viruses to evade the IFN induction and action [57]. Furthermore NS1B binds and blocks ISG15 protein inhibiting the ISGylation. The region of the NS1B protein that is required for this inhibition includes the domain that binds dsRNA. VACV may have a similar mechanism of influenza NS1 to evade ISG15 action as well.
We conclude that the cellular ISG15 protein has an essential role in VACV replication, acting as a negative feedback regulator of the cytokine signalling pathway and regulating in this way the innate response. VACV has therefore developed a mechanism to counteract this antiviral host response through E3. Because VVΔE3L is not lethal to ISG15+/+ mice lacking PKR, RNase L, and M×1 [40], there must be an additional IFN-induced antiviral pathway(s) effective against viruses, in which ISG15 should play an essential role. Understanding the host responses triggered by ISG15 and virus mechanisms of escape is necessary for development of therapies against important human pathogens.
HeLa cells (ATCC) were cultured in Dulbecco's medium (DMEM) supplemented with 10% newborn bovine serum (NCS) and antibiotics (Gibco, http://www.invitrogen.com). ISG15−/− cells and their wild type counterpart were generated by Osiak et al [58] and cultured in DMEM with 10% fetal calf serum (FCS). PKR−/− and their wild type counterpart [59] were cultured in Dulbecco's modified Eagle's medium with 10% FCS. VACV wild-type Western Reserve strain (WR) was grown on monkey BSC-40 cells (African green monkey kidney cells), purified by sucrose gradient banding as described (24) and titrated in BSC-40 cells. MVA and NYVAC, as well as VACV mutant of E3L were grown in Baby hamster kidney cells (BHK-21), sucrose purified and titrated in BHK-21 cells by immunostaining as previously described [60]. VACV constructs deleted of E3L (VVΔE3L), of the first 83 N-terminal amino acids of E3L (VVE3LΔ83N), or of the last 26 C-terminal amino acids of E3L (VVE3LΔ26C) were kindly provided by B. L. Jacobs (University of Arizona, USA) [44],[61].
RNA (1 µg) was reverse-transcribed using the Superscript first-strand synthesis system for reverse transcription-PCR (RT-PCR) (http://www.invitrogen.com). A 1∶40 dilution of the RT reaction mixture was used for quantitative PCR. Primers and probe sets used to amplify ISG15 was purchased from (http://www.appliedbiosystems.com). RT-PCR reactions were performed according to Assay-on-Demand, optimized to work with TaqMan Universal PCR MasterMix, No AmpErase UNG, as described [28]. All samples were assayed in duplicate. Threshold cycle (Ct) values were used to plot a standard curve in which Ct decreased in linear proportion to the log of the template copy number. The correlation values of standard curves were always >99%.
Cells were grown in 96-well plates to confluency and infected with different VACV or VVΔE3L viruses at the indicated multiplicity of infection (MOI) from 0.01 to 10 PFU/cell. At 24 hours post-infection (hpi), the medium was removed and cytolysis was determined by crystal violet staining as described previously [62]. The percentage of viable cells was calculated assuming the survival rate of uninfected cells to be 100%.
Murine embryonic fibroblasts (MEFs) were infected in 6-well plates with WR, or MVA, or NYVAC, or VVΔE3L and collected at indicated hpi in lysis buffer (50 mM Tris-HCl pH 8.0, 0.5 M NaCl, 10% NP40, 1% SDS). Equal amounts of protein lysates (100 µg) were separated by 14% or 8% SDS-polyacrylamide gel electrophoresis (SDS-PAGE), transferred to nitrocellulose membranes and incubated with antibodies, anti-ISG15 [58], -actin (http://www.sigmaaldrich.com), -E3 (kindly provided by B.L. Jacobs) followed by peroxidase-conjugated mouse or rabbit secondary antibodies. For the in vivo measurement of ISG15, parental or ISG15−/− mice were infected with WR, VVΔE3L, VVE3LΔ83N or VVE3LΔ26C at the multiplicity indicated. Lung samples were homogenized and mixed with SDS loading buffer and boiled for 10 min before Western blot analysis. ISG15 expression was detected as previously described with a rabbit antiserum against ISG15 [58] followed by peroxidase-conjugated rabbit secondary antibodies. Blots were developed using ECL (http://www.amersham.com).
Confluent MEF cells grown in 100 mm plates were treated with mouse IFN-α (100 units/ml) during 10 hrs and infected at 3 PFU/cell for 16 h with the recombinant viruses indicated and cells were collected and lysed and clarified supernatant was incubated with 20 µg of anti-mouse IP beads (http://www.ebioscience.com) previously incubated with a rabbit antiserum against ISG15 [58] or against E3 respectively. Immunoprecipitates were analyzed by SDS-PAGE followed by immunoblot with the antibody anti-E3 (kindly provided by B.L. Jacobs). The RNase treatment consisted in an incubation of the IP extracts with 10 µg of RNAse for 15 min at room temperature.
The origin of ISG15−/− mice has been described [58]. ISG15−/− and control wild type (WT) C57/BL-6 mice ( 6 to 10 weeks old) were immunized i.n in 25 µl PBS with VACV at 5×106 or 105 PFU/mouse or with VVΔE3L or VVE3LΔ83N or VVE3LΔ26 at 5×108 PFU/mouse. The i.p inoculation was with VACV at 2×107 or VVΔE3L at 108 PFU/mouse in 200 µl PBS. Animals were sacrificed at various times post-inoculation and spleen, liver, ovaries and lungs were removed, washed with sterile PBS, and stored at −70°C. Serum was obtained by retro-orbital bleedings 3 hours post-inoculation and was allowed to clot 1 hour at 37°C; after leaving samples at 4°C overnight, they were spun down in a microcentrifuge, and serum removed and stored at −20°C.
Secreted IL-6 from serum of mice infected i.p with WR or VVΔE3L at the indicated days was measured with the quantitative human IL-6 (BD Biosciences) according to the manufacturer's instructions. Captured IL-6 was quantified at 450 nm with a spectrophotometer. Triplicate samples were measured in two independent experiments. Alternatively, serum cytokine levels were analyzed for IL-6, TNF, IL-10, MCP-1, IFN-γ, and IL-12 p 70 by using the cytometric bead array mouse inflammation kit as indicated by the manufacturer (http://www.BDBiosciences.com).
ISG15-synthetic siRNA (s79221 and s79223), scrambled siRNA (used as a negative control; s4390843) and GAPDH siRNA (used as a positive control; s4390849) were purchased from Applied Biosystems and resuspended in RNase-free H2O. Transfection of siRNAs targeting each mRNA was carried out according to the manufacturer's instructions with some modifications. Murine ISG15+/+, PKR+/+ and PKR−/− embryonic fibroblasts were plated in 12-well plates 18 to 24 h before transfection. On the day of transfection, RNA-lipid complexes were introduced into each well of cells (20 nM RNA) by using siPORT Amine transfection reagent (http://www.ambion.com). The effect of specific siRNAs on target protein abundance was assessed by Western-blot. Twenty-four hours after transfection, siRNA-treated and non-treated control cells were mock-infected or infected with different WR or VVΔE3L viruses at 0.1 PFU/cell and CPE were visualized by phase-contrast microscopy at the indicated times p.i.
ISG15+/+ and ISG15−/− MEFs were transduced with high-titer viral supernatants corresponding to the pISG15-ires-GFP retroviral vector obtained as described [63]. Supernatants were collected at 48 h after transfection, filtered through a 0.45-µm-pore-size filter, and supplemented with complete DMEM medium +10% FCS before addition to growing MEFs. This protocol was repeated each 12 hours three times in presence of polybrene. The transduction efficiency was evaluated by Western-blot. Twenty-four hours after retroviral infection treated and non-treated control cells were mock-infected or infected with different VACV or VVΔE3L viruses at 0.1 PFU/cell and and CPE were visualized by phase-contrast microscopy at the indicated times p.i.
PKR+/+ and PKR −/− embryonic murine fibroblasts cultured on coverslips were infected with the viruses indicated. At 16 hpi cells were washed with phosphate-buffered saline (PBS), fixed with 4% paraformaldehyde (PFA) and permeabilized (10 min, room temperature) with 0.1% Triton X-100 in PBS, washed, and blocked with 20% bovine serum albumin (BSA) in PBS. Cells were incubated (1 h, 37°C) with anti-ISG15, -E3 (mouse antibody kindly provided by B. Moss); coverslips were washed extensively with PBS and further incubated (1 h, 37°C) with ToPro (http://www.molecularprobes.com) and appropriate fluorescein- or Texas Red-conjugated isotype-specific secondary antibodies. After washing with PBS, coverslips were mounted on microscope slides using Mowiol (http://www.calbiochem.com). Images were obtained using a Bio-Rad Radiance 2100 Confocal Laser microscope (http://www.biorad.com).
Formalin-fixed lung from mice mock-infected or infected with WR, VVΔE3L, VVE3L•83N or VVE3L•26 was resected, sectioned and stained with both hematoxilin and eosin as previously described [64]. |
10.1371/journal.pgen.1004622 | Evidence for Widespread Positive and Negative Selection in Coding and Conserved Noncoding Regions of Capsella grandiflora | The extent that both positive and negative selection vary across different portions of plant genomes remains poorly understood. Here, we sequence whole genomes of 13 Capsella grandiflora individuals and quantify the amount of selection across the genome. Using an estimate of the distribution of fitness effects, we show that selection is strong in coding regions, but weak in most noncoding regions, with the exception of 5′ and 3′ untranslated regions (UTRs). However, estimates of selection on noncoding regions conserved across the Brassicaceae family show strong signals of selection. Additionally, we see reductions in neutral diversity around functional substitutions in both coding and conserved noncoding regions, indicating recent selective sweeps at these sites. Finally, using expression data from leaf tissue we show that genes that are more highly expressed experience stronger negative selection but comparable levels of positive selection to lowly expressed genes. Overall, we observe widespread positive and negative selection in coding and regulatory regions, but our results also suggest that both positive and negative selection on plant noncoding sequence are considerably rarer than in animal genomes.
| Selection affects patterns of genomic variation, but it is unclear how much the effects of selection vary across plant genomes, particularly in noncoding regions. To determine the strength and extent of selective signatures across the genome, we sequenced and analyzed genomes from 13 Capsella grandiflora individuals. Because C. grandiflora has experienced a large, stable effective population size, we expect that selection signatures will not be overly distorted by demographic effects. Our analysis shows that positive and negative selection acting on new mutations have broadly shaped patterns of genomic diversity in coding regions but not in most noncoding regions. However, when we focus only on noncoding regions that show evidence of constraint across species, we see evidence for strong positive and negative selection. In addition, we find that genes with high expression experience stronger negative selection than genes with low expression, but the extent of positive selection appears to be equivalent across expression categories.
| Determining the amount of positive and negative selection and how it varies across the genome has wide-ranging implications for understanding genome function and the maintenance of genetic variation [1]. Current evidence suggests that both positive and negative selection are common in coding and some noncoding sequences in several model systems [2]–[8]. However our understanding of genome-wide selection in plants remains relatively limited [6], particularly in noncoding regions.
One key question concerns the extent to which both positive and negative selection act in noncoding regions of the genome compared with coding regions [2], [5]–[8]. For example, it has been suggested that the majority of adaptive evolution may occur in noncoding regulatory regions, where new mutations may have fewer deleterious pleiotropic effects [9], [10] but see [11]. Halligan and colleagues [8] showed that there have been many more adaptive substitutions in noncoding DNA than in coding regions in house mice, although adaptive substitutions in coding regions may experience stronger positive selection. Moreover, studies in Drosophila species and vertebrates have found that, although noncoding regions as a whole are generally less conserved than coding regions, there is more functional noncoding sequence than constrained coding sequence by a considerable margin [2], [12].
Comparing these results to noncoding selection across plant genomes is of particular interest because it has been hypothesized that in plants, regulatory evolution may occur more often through gene duplication than cis-regulatory change [13], possibly leading to lower levels of functional constraint and positive selection on plant noncoding DNA. Consistent with this prediction, Haudry and colleagues [14] recently compared the genomes of nine Brassicaceae species, and showed that approximately one quarter of the conserved sites in the Arabidopsis thaliana genome were in noncoding regions, a much smaller fraction than found to date in studies of vertebrates and Drosophila. However, the strength of selection on these noncoding sites, the extent of species-specific selection in noncoding regions, and the extent of positive selection in noncoding regions compared with coding regions have not been quantified.
While the strength of selection is expected to vary between coding and noncoding sequence, it also varies between genes. Gene expression level is one of the major determinants of rates of nonsynonymous evolution in coding regions in many species [15]–[17], including plants [18]–[21]. Variation in the strength of selection on genes could reflect differences in the relative importance of gene products for organism fitness, or it may simply relate to inherent properties of expression [22]. For example, deleterious mutations that cause misfolding or mis-interaction have more opportunity to interfere with cellular function when they occur in high expression genes [23]–[25]. Regardless of the underlying selective mechanisms, the negative correlation between expression level and nonsynonymous divergence could reflect relaxed purifying selection in lowly expressed genes, increased positive selection in lowly expressed genes, or both.
Here, we use population genomics to quantify the strength of both positive and negative selection inside and outside of coding regions and within highly and lowly expressed genes in a species-wide sample of 13 outbred Capsella grandiflora individuals. C. grandiflora is an obligately outcrossing member of the Brassicaceae family with a large effective population size (Ne∼600,000) and relatively low population structure [26], [27]. We estimate the strength of negative selection by fitting polymorphism data to a model of the distribution of negative fitness effects of mutations. We then quantify the contribution of positive selection to divergence in C. grandiflora using two complementary approaches: an extension of the McDonald-Kreitman test [28] and an analysis of neutral variation linked to lineage-specific fixed substitutions [29]. Our results demonstrate that both positive and negative selection are pervasive in coding regions, 5′ and 3′ untranslated regions (UTRs), and constrained noncoding regions of the C. grandiflora genome, but also that a large proportion of noncoding DNA may evolve neutrally. In addition, we find stronger negative selection in high expression genes compared to low expression genes, suggesting that differences in negative selection drive differences in rates of molecular evolution.
We sequenced 13 outbred C. grandiflora individuals (26 sampled haploid chromosomes; ∼140 Mb genome assembly) sampled from across the species' range in northern Greece using single-end Illumina GAII sequencing (Table S1). The resulting 108 bp reads were mapped to the Capsella rubella reference genome [30] using the Stampy aligner resulting in a median coverage of 34 reads per sample per site. Genotypes were called using the Genome Analysis Toolkit's Unified Genotyper [31]. After filtering for quality and depth (see Methods), we were left with ∼27 million sites, ∼1.5 million of which were single nucleotide polymorphisms (SNPs) (Table S2). Sites from across the genome were identified as 0-fold degenerate, 4-fold degenerate, intronic, 5′ UTR, 3′ UTR, or intergenic, based on the annotation of the C. rubella reference genome [30]. To avoid comparing sites that do not have equivalent mutation profiles, we excluded sites in coding regions that were neither 4-fold nor 0-fold degenerate. After filtering, our analysis includes 30–40% of coding and noncoding sites, except in intergenic regions where only approximately 10% of sites are retained due to the higher repeat content in these regions and the removal of highly repetitive pericentromeric DNA (Figure S1).
Consistent with previous estimates made using a much smaller set of loci (257 Sanger-sequenced loci) and a different range-wide sample [32], average nucleotide diversity at 4-fold degenerate sites (Watterson's Θw) was 0.022 and there was evidence for an excess of rare variants genome-wide at 4-fold degenerate sites compared with the standard neutral model (Tajima's D = −0.512). Introns (Θw = 0.020) and intergenic regions (Θw = 0.019) showed only slightly lower levels of nucleotide diversity than 4-fold degenerate sites, suggesting that the large majority of sites in these regions are effectively neutral, or subject to comparable levels of purifying selection as 4-fold degenerate sites. 5′ and 3′ UTRs showed a much stronger diversity reduction (Θw = 0.015 and 0.014 respectively), while 0-fold degenerate nonsynonymous sites showed the strongest reduction (Θw = 0.005).
Neutral diversity at 4-fold degenerate sites near centromeric regions was elevated on most chromosomes, similar to observations made in A. thaliana [33], Arabidopsis lyrata [34], [35] and Medicago truncatula [36], (Figure S2). As with these other species, this effect is not obviously caused by higher mutation rates, since divergence between Capsella and Neslia is not clearly elevated in these regions (Figure S2). Although elevated error rates in repetitive regions may contribute to high diversity, our observation of high diversity in these regions is still apparent after extensive filtering (see Methods). This increase in neutral diversity in pericentromeric regions may reflect a weakening of background selection in regions of low gene density, as recently shown in models of background selection applied to Arabidopsis [37]. Furthermore, diversity generally declines towards the ends of the chromosomes, potentially reflecting the stronger effects of background selection and/or selective sweeps in regions of relatively low recombination but high gene density, where the effects of linked selection are expected to be strongest. Consistent with these interpretations, we see an increase in diversity in regions of low coding density (Figure S3).
We also examined individual heterozygosity in sliding windows along each chromosome. A number of individuals showed large stretches of homozygosity indicative of biparental inbreeding (Figure S4 and Figure S5). Consistent with these regions reflecting local biparental inbreeding, no such regions are found in our sample that is derived from a between-population cross, called AXE. These regions of identity-by-descent (IBD) in our data highlight that, despite being self-incompatible and obligately outcrossing, local biparental inbreeding can still generate excess homozygosity in stretches across the genome. To avoid biased estimation of species-wide allele frequencies in these regions, we subsampled the data to treat all IBD regions as haploid rather than diploid sequence for the purposes of allele frequency estimation, although treating these regions as diploid does not qualitatively change our conclusions (Figure S6).
In order to quantify the amount of negative selection acting on different categories of sites, we used the methods of Eyre-Walker & Keightley [1] to compare the allele frequency spectrum (AFS) and divergence of various site categories to those for 4-fold degenerate sites, which are putatively neutral (Figure 1A). Consistent with the patterns of diversity described above, negative selection is generally much stronger in coding regions than noncoding regions (Figure 1A). This pattern is most clearly seen in 0-fold degenerate sites, the only site category with a sizable fraction of sites in the strongest category of negative selection (41%). Of the noncoding categories, UTRs show much stronger negative selection than other regions. In C. grandiflora ∼55% of both 5′ and 3′ UTRs are under moderate levels of purifying selection (Nes>1), but a considerably larger fraction of UTR sites are effectively neutral (45%) than 0-fold degenerate sites (14%). Additionally while the UTRs and CNSs (see below) show a signal of strong purifying selection (Nes>10), they experience less strong selection than 0-fold degenerate sites.
Genome-wide, we estimate that the proportion of intergenic sites that are nearly neutral approaches 100% and that approximately 70% of intronic sites are effectively neutral. Furthermore, bootstrapping results suggest that there is not significant support for less than 100% of intronic sites being effectively neutral. The large confidence intervals around estimates of selection on intronic sites may be due to strong selection at splice site junctions [14] coupled with typically weak to no selection outside of splice junctions. To test for selection near splice junctions, we quantified selection acting on the first and last 30 bp of each intron separately from sites in the middle of introns. While 100% of sites in the middle of introns are estimated to be effectively neutral, only 68% of sites in junctions are, suggesting that our wide confidence intervals around intronic sites can be partially explained by variance caused by sampling sites in these different regions between bootstraps. These generally low estimates of Nes in (non-junction) intronic and intergenic sites imply a general lack of purifying selection in most noncoding regions, a lack of sensitivity to detect small proportions of selected sites, and/or nearly equivalent purifying selection to synonymous sites.
Although our analysis suggests very low levels of purifying selection in noncoding regions other than UTRs and splice junctions, these global analyses may miss signatures of purifying selection on a small proportion of noncoding sites. One candidate set of sites that may have different signatures of selection are conserved noncoding sequences (CNSs); these are regions that show evidence of cross-species conservation, and are therefore prime candidates for functional noncoding sequences subject to selection. We identified CNSs across nine Brassicaceae genomes, following the implementation in Haudry et al. [14]. For this study, we used the Capsella genome as a reference for alignment, but excluded Capsella when identifying CNSs in order to avoid circularity when quantifying selection from diversity [8]. This method allows our analysis of selection on noncoding sites using polymorphism to be more independent of the comparative analysis. When we look at only these conserved regions in our C. grandiflora sample we see a small proportion of effectively neutral sites (28%) compared to the noncoding regions as whole, suggesting that the majority of CNS sequences are subject to purifying selection (Figure 1). However, estimates suggest that CNSs are generally under weaker purifying selection than nonsynonymous (0-fold) sites and experience primarily weak and intermediate purifying selection (Figure 1).
Although CNSs as a whole retain a considerable proportion of effectively neutral sites, it is of interest to examine whether particular classes of CNS show stronger selection. To examine differences between categories we quantified selection on the different types of CNSs separately (Figure S7). In most categories, about 25% of sites are nearly neutral, a slightly stronger signal of purifying selection than when we pool all CNSs. Intronic CNSs have a larger proportion of effectively neutral sites than other categories, in agreement with the general neutrality of intronic sites (Figure 1). In contrast, small noncoding RNAs (sncCNSs) have a stronger signal of selection than the other CNS categories. However, the number of sites used to make the AFS for each of these categories varies substantially (Table S2), and our sample of sncCNSs has very little polymorphism (155 segregating sites). Nevertheless, despite the wide confidence intervals, sncCNSs still show a significantly (p<0.001) smaller fraction of sites that are nearly neutral (Nes<1) than the pooled CNSs, which could be due to strong selection for sequence specificity to obtain the proper secondary structure important for RNA activity [38]. This effect is consistent with sncCNSs showing a higher degree of conservation across the Brassicaceae [14] and having traceable orthologs in other plants.
We used the approach of Eyre-Walker and Keightley [28] to estimate the proportion of fixations driven by positive selection (α) and the rate of positive selection (ω) while taking into account the effect of slightly deleterious mutations, which can bias estimates of positive selection downwards. To do this, we estimated divergence using whole genome alignments of C. rubella, A. thaliana, and Neslia paniculata (estimate of 4-fold synonymous divergence Ks between C. rubella and N. paniculata is Ks = 0.14). Because the large majority of noncoding sites are estimated to be effectively neutral, and because of alignment concerns between species in unconstrained noncoding regions, we focus our estimates of positive selection on 0-fold degenerate sites, CNS sites, and UTRs. We found that 0-fold degenerate sites show a very high proportion of divergence driven by positive selection (Figure 1B; α = 0.417) and estimates of the rate of adaptive substitution relative to synonymous substitution (Figure 1C; ω = 0.08). Similarly, UTRs and CNS sites show evidence for positive selection (Figures 1B and 1C). These results generally suggest widespread positive selection in both nonsynonymous and functional noncoding genomic regions.
If many of the amino acid changes between C. grandiflora and its nearest relatives are due to recent, strong positive selection from new mutations, we expect to see the signature of selective sweeps: reduced neutral diversity surrounding amino acid fixations [39], [40]. We tested for this signature by measuring the proportion of 4-fold degenerate sites in each window that were polymorphic (referred to hereafter as ‘4-fold diversity’) in non-overlapping 1 kb windows surrounding fixed replacement (n = 60,378), and silent (n = 83,812) substitutions in C. grandiflora. We found that 4-fold diversity surrounding fixed replacement substitutions was lower than 4-fold diversity surrounding fixed silent substitutions in the 4 kb window surrounding substitutions (Figure 2A). This result was robust to various window sizes from 500 kb to 2 kb (Figure S8) and a one-tailed test for reduced 4-fold diversity around replacement sites was significant (p<0.01 for 2 kb on either side of the substitution).
Patterns of diversity may be distorted by elevated mutation rates surrounding substitutions [39], which would increase diversity and divergence in C. grandiflora. Consistent with this prediction, divergence at 4-fold degenerate sites (‘4-fold divergence’) is elevated around synonymous and replacement substitutions (Figure 2B). To control for elevated mutation rate, we divided diversity by divergence at 4-fold degenerate sites (subsequently referred to as ‘4-fold diversity/divergence’). We observed a reduction in 4-fold diversity/divergence around replacement substitutions compared to silent substitutions, demonstrating that the signature of recurrent sweeps is not an artifact caused by variation in mutation rate (Figure 2C, p<0.01 for 1 kb on either side of the substitution).
An analogous test for selective sweeps around fixations in noncoding regions is challenging because the test depends on accurately identifying interspersed functional and neutral sites, a difficult task in noncoding regions [8]. Instead, we compared 4-fold diversity and divergence around fixed substitutions in CNS regions (n = 12,578) with 4-fold diversity and divergence around fixed substitutions in non-conserved intergenic, intronic, and UTR regions (n = 117,178). Interestingly, there is a reduction in both 4-fold diversity and divergence surrounding fixed substitutions in CNSs compared to non-conserved noncoding regions (Figure S9). It is not clear why 4-fold divergence decreases around CNS substitutions; it is possible that in genomic scans for conserved regions, large-scale constraint might span both coding and noncoding sequence, causing non-independence and reducing divergence at 4-fold degenerate sites near CNSs. However, there is still a reduction in 4-fold diversity/divergence around fixed substitutions in CNSs compared to those in non-conserved intergenic regions, consistent with the action of recurrent selective sweeps (Figure 3A).
The observed reduction in diversity/divergence around CNS substitutions could also reflect the action of background purifying selection; sites closer to CNSs may experience a reduction of neutral diversity due to greater purifying selection on mutations in CNSs. This effect is not a problem for comparisons between replacement and silent substitutions because they are interspersed within the same exons, so diversity and divergence around these sites experience the same background selection. To ensure that the reduction in diversity/divergence surrounding CNS substitutions compared to non-conserved noncoding substitutions is not due to differences in background selection between CNS and intergenic sites, we compared neutral diversity and divergence surrounding CNSs that contain at least one fixed substitution to neutral diversity and divergence around those that do not. There is a reduction in neutral diversity/divergence surrounding CNSs containing a fixed substitution (n = 12,884) compared to CNSs without fixed substitutions (n = 41,212), suggesting that this signature of recurrent sweeps is not driven by background selection specific to CNSs (Figure 3B).
We measured expression levels of all expressed genes using RNA extracted from leaf tissue of 10 of the 13 C. grandiflora individuals. Genes were sorted by mean expression level and split into four equally sized groups, which will be referred to as “high”, “mid-high”, “mid-low”, and “low” expression genes. We calculated polymorphism within C. grandiflora and lineage-specific divergence from N. paniculata and A. thaliana for sites within these genes. As expected from previous studies, dN/dS is considerably lower in high expression genes (0.15) than low expression genes (0.22). In addition, dN/dS is negatively correlated with expression level across all genes (correlation coefficient = −0.051, p<0.001).
To test whether the strength of negative selection differs between expression categories we compared the allele frequency spectra of sites in different expression categories. Replacement polymorphisms in high expression genes show a stronger skew towards rare alleles than those in low expression genes (Figure S10). In addition, a larger proportion of replacement sites are invariant in high expression genes (98.9%), than in low expression genes (97.8%), consistent with stronger negative selection. Comparisons of the distribution of fitness effects show that high expression genes have a much smaller proportion of effectively neutral sites (6.8%) than low expression genes (16%, randomization test [28], p<0.001) (Figure 4A).
Increased divergence in low expression genes relative to high expression genes could also be caused by increased positive selection in low expressed genes compared to highly expressed genes. To test this possibility, we calculated α and ω as described above. High expression genes have a significantly higher value of α (0.66) than low expression genes (0.42, p<0.01) but the ω value for both classes is similar (high: 0.11, low: 0.10, p = 0.38), suggesting that the rate of positive selection does not differ between high and low expression genes (Figure 4B, 4C). The difference in α between the two categories likely reflects the reduction in the number of weakly deleterious and effectively neutral mutations that are able to fix due to stronger purifying selection in high expression genes compared to low expression genes, causing a higher proportion of those amino acids that do reach fixation to be positively selected.
In this population genomic survey of C. grandiflora, we demonstrated that positive and negative selection contribute to DNA sequence variation in protein-coding regions, UTRs, and CNSs. We also showed that differences in divergence between high and low expression genes are due to increased negative selection in high expression genes, not increased positive selection in low expression genes. In addition, we found a clear signature of recurrent selective sweeps contributing to divergence in coding regions as well as CNSs. Overall, our evidence for widespread positive and negative selection in C. grandiflora is in line with expectations, given its outcrossing mating system, large Ne, limited population structure, and lack of a recent whole genome duplication [6].
In contrast, selection appears to be very rare in intergenic and (non-junction) intronic regions that are not conserved across Brassicaceae species. In particular, we cannot detect significant evidence of purifying selection in intergenic or intronic regions as a whole, suggesting that selected sites within these regions must be rare or absent. However, when we only examine CNSs, we do see evidence of selection, indicating that at least 5% of sites in intergenic regions are selected, but the DFE approach is not sensitive enough to detect selection on such a small subset of intergenic sites. This result implies that this approach is likely to also be missing lineage-specific selection when it comprises a relatively small fraction of sites, and it highlights the importance of integrating additional evidence of function (comparative and experimental) for improved quantification of selection.
The general neutrality of noncoding regions, based on population genomic analysis, is consistent with the conclusions of Haudry and colleagues [14], who used comparative genomics approaches to estimate that only 5% of noncoding bases are under selection in the Arabidopsis genome. This result contrasts with Drosophila and humans, where a relatively large fraction of selected sites are found in noncoding regions [6]. For example, in Drosophila, only 30%–70% of intronic and intergenic regions are nearly neutral [2], [28], [29]. Similarly, Halligan et. al. [8] recently used information from the DFE to infer the number of adaptive substitutions in mice both in coding and noncoding regions. They show that the majority (approximately 80%) of the adaptive substitutions in the mouse genome are in noncoding regions and suggest that they may have regulatory function. In contrast, our data show that C. grandiflora has similar numbers of adaptive substitutions in 0-fold sites (50.6 kb) and noncoding sites (21.6 kb, 3′ UTR excluding CNSs; 10.2 kb, 5′ UTR excluding CNSs; 32.7 kb, CNS; 64.4 kb total). Additionally, the width of diversity reductions surrounding replacement substitutions and substitutions in CNS regions appear comparable, suggesting that there is little evidence for a difference in the strength of positive selection on substitutions in coding regions compared to conserved noncoding regions. Our results are consistent with previous suggestions that, unlike in animals, plant genomes may contain fewer noncoding regulatory sequences subject to positive and negative selection, possibly because gene expression can be modified through frequent gene duplication and functional divergence rather than through the evolution of novel regulatory elements [13]. In future work, it would be interesting to quantify the extent to which adaptive changes in gene expression in plants occur following gene duplication relative to between-species divergence at orthologous genes.
Unlike other classes of noncoding sequence, UTRs show relatively high levels of purifying selection, likely reflective of their function in post-transcriptional regulation [41]. UTRs are also under stronger negative selection than other noncoding regions in Drosophila [2], and this result is also in line with the previous study using comparative genomics in the Brassicaceae [14]. Interestingly, we infer that a large fraction of selected sites in UTRs may be outside of CNS regions identified in between-species comparisons. In particular, using estimates of the proportion of sites under selection, we estimate that 88% of 3′ UTR and 77% of 5′ UTR selected sites are outside of conserved regions. This result suggests that there may be many species-specific (i.e., non-CNS) functional regions in UTRs and they may therefore play an important role in recent or local adaptation.
One important consideration is the extent to which our analyses are truly reflective of genome-wide patterns of selection. Despite whole genome sequencing, our analyses are restricted to approximately 20% of the genome, and only 10% of intergenic sites, largely due to the fact that a large fraction of the genome is pericentromeric, repetitive and/or surrounds insertion/deletion events. It is important to recognize that our estimates of selection apply strictly to this ‘accessible’ genome and that the extent of purifying and positive selection on the repetitive regions remains difficult to assess. Nevertheless, we would expect that our conclusions about low levels of purifying and positive selection across most noncoding regions are likely conservative with respect to these filters because a large proportion of repetitive DNA is likely to be neutral. On the other hand, rates of positive selection may be elevated in coding regions of duplicate genes filtered out of our analysis [42], suggesting that our estimates of positive selection in protein-coding regions may also be a lower bound.
A second concern is the extent to which synonymous sites are neutrally evolving. Although analysis of codon usage bias from population genetic data does suggest the action of some purifying selection on synonymous sites in this species [43], the strength of selection inferred is close to effective neutrality. Furthermore, synonymous site selection is expected to be stronger in more high expression genes [23], [44], causing us to underestimate, rather than overestimate, the difference in the strength of purifying selection compared with low expression genes. Thus, while selection on synonymous sites may bias our estimates of selection slightly downward, our general conclusions are likely to be robust to violations of neutrality. Nevertheless, more investigation of the action of selection on synonymous sites is important, particularly given growing evidence for synonymous site selection that may reflect gene regulation, in addition to codon usage [45], [46].
At synonymous sites, we see an excess of rare variants, as indicated by a negative Tajima's D. The excess of rare variants is unlikely to be explained by a high Illumina error rate, as our observed value of −0.51 is nearly identical to a previous estimate (−0.52) from Sanger-sequenced loci and a comparable geographic sampling [27]. This previous study found that, while population subdivision was low compared to other herbaceous species studied, there were still three major geographic clusters (average between-population Fst of 0.11). If we restrict our dataset to one of the three geographic regions based on these previous results, Tajima's D approaches zero (−0.16 at 4-fold degenerate sites), suggesting that the excess of rare variants at synonymous sites may be largely due to population structure.
In this study, we took advantage of the two detectable signatures expected to remain after recurrent classic selective sweeps from new mutations: 1) an excess of replacement substitutions relative to expectations based on polymorphism, and 2) reduced neutral diversity near fixed differences. Our findings strongly suggest that positive selection has been common in coding regions, UTRs and conserved noncoding regions in C. grandiflora and that classic selective sweeps contribute significantly to divergence in these regions. To our knowledge, this is the first time that the signature of recurrent selective sweeps has been observed in a non-Drosophila species, despite being tested in other species [8], [47]. Our ability to detect the signature of recurrent sweeps may be because C. grandiflora has relatively low linkage disequilibrium, increasing power.
However, many positively selected alleles may not follow the trajectory of a classic selective sweep. Soft sweeps — adaptation from an allele previously maintained in the population by mutation-selection-drift balance or the simultaneous fixation of multiple independently derived mutations at the same allele — may still increase the replacement to silent divergence ratio, but are expected to have a smaller effect on linked neutral diversity [48]–[50]. We expect that soft sweeps will also be common in C. grandiflora because of its large Ne [50], [51]. In addition, adaptation in genes that contribute to polygenic traits is often expected to occur without fixation of new mutations [52], and this will be missed by both of our tests for positive selection. These considerations suggest that both measures of positive selection are conservative and may miss many instances of positive selection acting in the genome.
Our conclusions about the prevalence of selective sweeps in C. grandiflora may seem to conflict with our observation that diversity and Tajima's D are slightly higher at 4-fold degenerate sites than intergenic sites, since frequent sweeps in coding regions should reduce diversity more strongly in sites near and within genes. There are two likely contributors to this discrepancy. First, recurrent sweeps may in fact reduce average diversity in 4-fold degenerate sites and, by using these sites to set neutral expectations, we are underestimating the strength of purifying selection in intergenic regions. Second, because recombination rates are relatively high, and intergenic regions near coding regions relatively small in Capsella, the average impact of linked selection may be similar at 4-fold degenerate sites and intergenic sequences.
Highly expressed genes diverge less than genes with low expression in many species [15]–[17], [19], [24], [53]–[55]. This pattern could be due to stronger positive selection in low expression genes or stronger negative selection in high expression genes, or both. Our results suggest that variation in divergence rates between high and low expression genes is largely due to increased negative selection in high expression genes compared to low expression genes. This result is consistent with previous studies that have suggested that new nonsynonymous mutations that cause protein mis-folding or mis-interaction will have stronger deleterious effects in high expression genes than low expression genes and that new mutations that cause mRNA mis-folding are under stronger negative selection in high expression genes than low expression genes [23]–[25]. In addition, our results agree with a similar study in Medicago truncatula that found stronger purifying selection in genes that were expressed than in genes that were not expressed [20].
Population samples for C. grandiflora represented a ‘scattered’ sample of one individual per population for twelve populations from across the geographic range in Greece, plus a thirteenth sample that was the product of a cross of two additional populations (Table S1). Plants were grown for several months at the University of Toronto greenhouse, and genomic DNA was extracted from leaf tissue using a modified CTAB protocol. Library preparation and single-end genomic sequencing were conducted at the Genome Quebec Innovation Centre at McGill University on the Illumina GAII platform. Each sample was sequenced in 2 to 3 lanes and with a read length of 108 bp.
Leaves from 10 of the 13 individuals were collected and flash frozen for RNA extraction using Qiagen's RNAeasy plant extraction kit. This RNA was sequenced at the Genome Quebec Innovation Centre, on an Illumina GAII platform with one individual per lane, generating single-end 108 bp long reads. The RNA sequence from these 10 individuals was used for the annotation of the C. rubella reference genome, as reported in [30], but the raw sequence data was reanalyzed for this study (see below).
Genomic reads were aligned to the C. rubella reference genome [30] using the Stampy aligner 1.0.13 with default settings [56]. Sites around indels were realigned using the Genome Analysis Toolkit (GATK) v1.05777 indel realigner [31]. Genotype and SNP calls were conducted using the GATK UnifiedGenotyper with default parameters [57], after aligning and genotyping the median site quality was 89 and the median individual depth across all sites was 34.
To get a rough assessment of genotyping error rates, we conducted Sanger sequencing from nine coding regions in six of our individuals. From a total of 16,389 bp of Sanger sequence, we found 8 differences between Sanger and Illumina genotypes, giving an estimated error rate of 0.00049. Three of these disagreements were due to three segregating bases at a single site, which we excluded in our GATK genotyping protocol. As we suspect several of these disagreements may be due to Sanger sequencing errors due to variation in allelic representation of heterozygotes, this provides an upper bound estimate of error rate in coding regions, although higher indel rates and repetitive sequence in noncoding DNA may lead to a higher error rate in those regions.
AFSs were generated from counts of sites in the VCF. Invariant sites were excluded from the AFS if (1) the site quality score was below 90, (2) the fraction of reads containing spanning deletions was not 0 (i.e. the ‘Dels’ value was greater than zero), or (3) any individual's read depth was less than 20 or greater than 60. Additionally, polymorphic sites were excluded, based on filters 1–3, if (4) the most likely genotype of any individual did not have a phred scaled likelihood score of 0, and if (5) the second most likely genotype had a phred likelihood score less than 40. Additionally, entire regions of the genome were filtered out of the analysis if less than 30% of the sites in a 20 kb window passed all other filters. This final filter primarily eliminated pericentromeric regions that were highly repetitive, where we were not confident in genotype calls and observed high heterozygosity.
Our data showed evidence of identity by descent (IBD) in some samples (Figure S4). We identified these regions by splitting the genome into 200 kb windows, then calculating FIS (Figure S5). If FIS was greater than 0.5, the region was flagged as IBD. Across all samples no more than 3 of these regions overlapped. For further analyses we downsampled data in other regions down to 23 chromosomes treating any region of IBD as haploid to ensure that no IBD region was sampled twice from the same individual.
We calculated lineage-specific divergence in two ways. First, we aligned the C. rubella reference sequence with sequence data from A. thaliana and N. paniculata using lastZ [58] with chaining, as previously described [14]. In order to get an estimate of divergence unique to the Capsella lineage, we called sites as diverged where A. thaliana and N. paniculata had the same nucleotide and this nucleotide differed in the C. rubella sequence. If any of the three species was missing data at a site, then that site, and sites 5 bp upstream and downstream of the site, were excluded from divergence analyses in order to avoid inflating divergence because of spurious alignments around indels.
We used a second method for calculating divergence for comparisons that included only coding sequences, particularly for the comparison of genes with different expression levels. We found orthologs between C. rubella, A. thaliana and N. paniculata genes using InParanoid [59] and MultiParanoid [60]. The peptide sequences of these orthologs were aligned using DialignTX [61], and reverse-translated into coding sequence. Whole-gene divergence at synonymous and nonsynonymous sites was calculated, using PAML [62], under a model where ω was allowed to vary in the Capsella lineage compared to other branches.
We conducted comparisons of estimates of the distribution of fitness effects using the two methods above with identical gene sets, and found a very strong concordance of results (see Figure 4 compared to Figure S11). Furthermore, while we don't predict a significant effect on results, it is important to note that the two methods also differed in how selected and nonselected classes were determined: the first distinguishes between 0-fold and 4-fold sites and discards other sites, while the second distinguishes between synonymous and nonsynonymous sites, including all data. However, both approaches gave comparable estimates of positive and negative selection.
Conserved noncoding sequences (CNS) were identified in the C. rubella genome by first obtaining whole-genome multiple alignments, using a variant of the lastZ/Multiz pipeline previously described [14], [63] and using C. rubella as the reference genome. The C. rubella genome sequence was then neutralized (bases replaced with N) and the PhastCons tool used to quantify family-wide levels of conservation. CNSs were then identified, based on extended (>12 nt) near-continuous regions of high conservation as previously described [14].
Site categories were determined based on the Joint Genome Institute's gene annotation of the C. rubella reference genome [30]. The allele frequency spectra (AFS) and divergence values were calculated for each category of sites, and DFE-alpha [28], [64] was used to estimate the fraction of sites under negative selection and α, using 4-fold degenerate sites as the neutral reference. The genome was broken up into 10 kb regions and these regions were bootstrapped 200 times to generate 95% CIs for selection on each category of sites. We tested for a significant difference in selection between the pooled set of CNSs and each individual category of CNSs using a randomization test, as in Keightley & Eyre-Walker [28], by calculating the proportion of bootstraps where selection was higher in the pooled set of CNS versus the category of interest. Because this is a two-tailed test, we report twice this proportion as the p value.
We used the multiple species alignments of orthologous genes, generated as described above, to identify silent and replacement single-nucleotide sites that were the same in A. thaliana and N. paniculata but differed in the C. rubella reference, suggesting that the substitution had most likely occurred in the Capsella lineage after divergence from N. paniculata. From these substitutions, we identified those that did not diverge between C. rubella and C. grandiflora and were fixed in C. grandiflora.
We calculated neutral diversity in sliding windows around fixed substitutions by calculating the proportion of 4-fold degenerate sites within these windows that were polymorphic in C. grandiflora (i.e., the proportion of segregating sites). Neutral divergence was measured by calculating the proportion of 4-fold sites within these windows that diverged in the Capsella lineage. Diversity/divergence was calculated by dividing diversity by divergence in each window. We conducted this analysis for windows of 500 bp, 1 kb, and 2 kb, extending 40 kb from each substitution. We chose this window size range to match analysis done in Sattath et al [39]. For each of the above measures, we bootstrapped by substitution (n = 1000) and removed the top and bottom 25 bootstraps to construct 95% confidence intervals. Following Hernandez and colleagues [47], we tested the null hypothesis that diversity/divergence around replacement and silent substitutions does not differ by calculating a one-tailed p value for each window, equal to (i+1)/(n+2) where i is the number of bootstraps in which diversity/divergence around silent sites is lower or equal to the actual diversity/divergence around replacement sites, and n is the total number of bootstraps.
To detect the effects of linked selection on noncoding DNA, we compared diversity around fixed substitutions within CNSs to diversity around fixed substitutions in non-conserved intergenic regions. To find these substitutions, we compared the multiple sequence alignments of the CNSs between C. grandiflora, N. paniculata, and A. thaliana and chose sites that differed between C. grandiflora and the other species and were fixed within C. grandiflora. Additionally, we compared neutral diversity around CNSs with at least one fixed substitution to neutral diversity around CNSs without any fixed substitutions.
Illumina sequencing generated 331,629,531 reads for 10 individuals, ranging from 31,267,774 to 35,552,133 reads per individual. This RNA sequence was mapped to the C. rubella reference genome using Tophat 1.2.0 [65], and expression level was quantified from these mapped reads using Cufflinks 1.3.0 [66]. Cufflinks standardizes expression levels by gene length and library size, returning values in units of ‘fragments per kilobase of exon per million fragments mapped’ (FPKM). We calculated the mean expression level for each gene across our 10 samples and removed those genes with <1 FKPM to eliminate genes that may have been mis-annotated. The remaining 11,564 genes were divided into four, roughly equally sized categories based on expression level: low (1–6.8 FPKM), mid-low (6.8–17.5 FPKM), mid-high (17.5–44.7 FPKM), and high (44.7–17,092 FPKM). The distribution of fitness effects, α, and ω were calculated for each gene set, using the same protocol described above. We bootstrapped each gene set by sampling genes with replacement 1000 times to generate 95% confidence intervals for selection strength. Using the same methods described for tests of differences within the CNSs categories above, we tested for a significant difference in selection strength between high and low expression genes.
|
10.1371/journal.ppat.1006684 | Preconceptual Zika virus asymptomatic infection protects against secondary prenatal infection | Pregnant women, and their fetal offspring, are uniquely susceptible to Zika virus and other microbial pathogens capable of congenital fetal infection. Unavoidable exposure to Zika virus in endemic areas underscores the need for identifying at-risk individuals, and protecting expecting mothers and their fetal offspring against prenatal infection. Here we show that primary Zika virus asymptomatic infection in mice confers protection against re-infection, and that these protective benefits are maintained during pregnancy. Zika virus recovery was sharply reduced in maternal tissues and amongst fetal concepti after prenatal challenge in mothers with resolved subclinical infection prior to pregnancy compared with mice undergoing primary prenatal infection. These benefits coincide with expanded accumulation of viral-specific antibodies in maternal serum and fetal tissues that protect against infection by the identical or heterologous Zika virus genotype strains. Thus, preconceptual infection primes Zika virus-specific antibodies that confer cross-genotype protection against re-infection during pregnancy.
| Expecting mothers are uniquely susceptible to Zika virus infection that often spreads to vital tissues of the developing fetus. Since Zika virus infection in healthy non-pregnant individuals is mostly asymptomatic, a large proportion of reproductive age women that live in Zika endemic areas have been previously infected, and cleared the infection prior to pregnancy. Here we show that primary Zika virus asymptomatic infection in mice protects against re-infection, and that these protective benefits are maintained during pregnancy. Protection in this preclinical model is mediated by circulating antibodies found at very high levels amongst individuals with resolved infection that efficiently neutralize virus infectivity. Thus, antibodies against Zika virus, naturally stimulated by prior asymptomatic infection, protect against re-infection during pregnancy, and the presence of these protective antibodies may help discriminate protected individuals from others that remain susceptible to infection. This knowledge, combined with protection primed by promising candidate vaccine formulations that also stimulate production of high level neutralizing antibodies, will help identify reproductive age women at-risk for severe infection consequences and more focused therapeutic strategies for averting infection.
| The ongoing Zika virus (ZIKV) epidemic has triggered an explosion in cases of fetal death, microcephaly and other birth defects in surviving infants with congenital infection [1–4]. These sequelae usually occur in parallel with higher and more prolonged maternal viremia for up to 15 weeks following prenatal infection [5–9]. By contrast, ZIKV infection in non-pregnant individuals is mostly subclinical or asymptomatic, and associated with only transient self-resolving viremia [10]. For example, only 19% of ZIKV IgM seropositive individuals reported clinical symptoms during the 2007 Yap Island outbreak [11]. Likewise, only 11% of individuals developed reportable symptoms despite an approximate 50% rate of newly acquired seroprevalance during the 2013–2014 French Polynesian outbreak [12]. Thus, considering ZIKV-associated morbidity is largely confined to infection during pregnancy, there is an urgent need for improved strategies for protecting against congenital fetal invasion and prenatal infection susceptibility. This urgency persists even though several ZIKV candidate vaccines have recently been shown to confer very promising protection in animal infection models [13–21], since clinical validation of efficacy and safety, especially with increasingly recognized immunological shifts that physiologically occur during human allogeneic pregnancies, have not been demonstrated.
Given the limited and non-specific clinical symptoms associated with most ZIKV infections in healthy non-pregnant individuals, a fundamental unanswered question regarding prenatal ZIKV susceptibility is whether preconceptual infection protects against re-infection during pregnancy. For some classical prenatal pathogens (e.g. varicella virus, rubella virus), maternal susceptibility and congenital fetal invasion are each efficiently overturned amongst women with resolved infection prior to pregnancy [22–24]. By contrast, protection conferred by preconceptual infection is considerably less reliable for other prenatal pathogens—that may reflect susceptibility to re-infection by immunologically discordant strains (e.g. human cytomegalovirus, influenza virus) [25, 26], or attenuated responsiveness of normally protective maternal immune components simultaneously required for sustaining fetal tolerance (e.g. Plasmodium spp., Listeria monocytogenes) [27–29]. Importantly, while prior infection has been shown to protect against lethal re-challenge in non-pregnant hosts for other flaviviruses such as West Nile virus (WNV) or St. Louis encephalitis virus [30–32], whether primary infection protects against re-infection during pregnancy for flaviviruses as a group remains poorly defined since prenatal infection has not traditionally been a dominant feature for these viral pathogens until the recent clinical emergence of ZIKV [33].
Herein, pregnancies established among genetically discordant strains of inbred mice were used to investigate how primary preconceptual ZIKV infection impacts the susceptibility of mothers and their fetal offspring to re-infection during pregnancy. We show that anti-viral immunity primed by asymptomatic infection prior to pregnancy protects against re-infection during pregnancy, with sharply reduced susceptibility in both maternal and fetal tissues. These findings have important translational implications for discriminating individuals at risk for severe prenatal infection from others with naturally acquired protective immunity, and new strategies for protecting against congenital fetal invasion.
ZIKV shares with other flaviviruses, susceptibility to innate anti-viral immunity activated by type I interferons (IFNs), and functionally overriding host cell type I IFN responsiveness is required for productive symptomatic infection [34, 35]. We exploited the selective resistance of murine cells to STAT2 degradation by ZIKV NS5 protein, reasoning that the natural innate resistance from unabated type I IFN responsiveness in mice makes this species ideally suited to probe immunological shifts primed by asymptomatic infection in non-pregnant individuals [36, 37]. In turn, temporally overriding innate protection by initiating type I IFN receptor antibody blockade immediately prior to secondary challenge creates a unique opportunity for investigating the impacts of prior asymptomatic infection on susceptibility to re-infection. Consistent with the results of recent studies [34, 35], standard laboratory C57BL/6 mice inoculated with a ZIKV clinical isolate (PRVABC59) from the contemporary Latin American outbreak showed no clinical symptoms and rapidly cleared the infection (Fig 1A and 1B). Comparatively, clinical signs of infection (i.e., hunched posture, ruffled hair, lethargy) that paralleled high levels of circulating virus were unleashed by type I IFN receptor antibody (MAR1-5A3) blockade administered beginning one day prior to infection (Fig 1A and 1B). Thus, antibody mediated type I IFN receptor blockade efficiently causes clinical and virological susceptibility to ZIKV infection despite previously described normal weight gain in mice [34].
To investigate how prior infection impacts susceptibility to re-infection, clinical symptoms and viral RNA levels after secondary ZIKV challenge amongst mice with prior asymptomatic infection were compared with primary infection in naive control animals. Interestingly, despite susceptibility conferred by type I IFN receptor blockade initiated one-day prior to infection, clinical symptoms were sharply attenuated amongst mice with prior asymptomatic infection compared with naive control mice (Fig 1C). These protective benefits paralleled significantly reduced recovery of ZIKV RNA in the blood and tissues of mice undergoing secondary challenge compared with primary infection in naive controls (Fig 1D). Together with recent studies demonstrating significantly reduced susceptibility to secondary compared with primary ZIKV infection in non-human primates [38–40], these results highlight the immunogenicity of ZIKV where even abortive primary infection can efficiently prime protection against re-infection.
A variety of adaptive immune components stimulated by ZIKV infection or candidate vaccine formulations are associated with protection. For example, polyfunctional IFN-γ plus TNF-α producing CD8+ T cells with broad ZIKV specificity primed by primary infection can overturn the susceptibility of naive recipient mice after adoptive transfer in vivo [41, 42]. Likewise, sterilizing immunity induced by live attenuated or inactivated ZIKV candidate vaccine formulations occurs in parallel with sharply expanded accumulation of IFN-γ producing ZIKV-specific CD4+ and CD4+ T cells [14, 15]. On the other hand, the same live attenuated viral strains and nucleic acid based candidate vaccines each prime high titer ZIKV-specific IgG antibodies that protect against infection in non-pregnant mice and non-human primates [13–18]. ZIKV susceptibility to antibodies is further highlighted by protection against infection in pregnant and non-pregnant mice adoptively transferred individual clones of human monoclonal antibodies that bind and neutralize ZIKV in vitro infectivity most efficiently [43]. Thus, while ZIKV-specific T cells and antibodies are each capable of protection, the adaptive immune components stimulated by subclinical infection that protect against re-infection remain undefined. Accordingly, our model of protective immunity primed by primary ZIKV asymptomatic infection was used to evaluate the relative contribution of serological compared with cellular adaptive immune components.
Consistent with efficient serological conversion after human subclinical infection [11, 12, 44], the serum of mice three weeks after asymptomatic primary infection showed robust accumulation of ZIKV-specific IgG antibody, whereas IgA and IgM titers remained at background levels found in naive control mice (Fig 2A). ZIKV-specific IgG antibodies were highly enriched for IgG2a and IgG2b subclasses, with specificity to ZIKV envelope (ENV) and NS1 proteins (Fig 2B and 2C). In turn, serum from mice three weeks after primary asymptomatic infection also efficiently neutralized ZIKV plaque formation in Vero cell monolayers, with serial dilutions of the serum showing reductions in functional activity that directly paralleled when the titer of virus-specific IgG antibody returned to background levels (loss of activity for both after 103 to 104-fold dilutions) (Fig 3A compared with Fig 2A). Interestingly, heat-inactivation did not significantly impact the neutralization potency of serum from mice with prior asymptomatic infection suggesting heat-liable serum components (e.g. complement) are not required for neutralizing ZIKV in vitro infectivity (S1A Fig).
To further investigate whether ZIKV-specific antibodies primed by asymptomatic primary infection protect against infection, the susceptibility of naive mice adoptively transferred serum from donor mice infected with ZIKV three weeks prior was evaluated. ZIKV RNA levels were significantly reduced in the serum, spleen, liver and brain of recipient mice prophylactically treated one day prior to infection with serum from donor mice with resolved primary asymptomatic infection compared with the serum of naive control mice (Fig 3B). Interestingly, Fc-γ receptor (CD16/CD32) in vivo blockade [45, 46] efficiently overturned the protective benefits of adoptively transferred serum from mice with resolved primary infection, highlighting essential roles for phagocytic host cells that take up antibody coated virus through Fc-γ receptors (Fig 3C). In contrast, depletion of CD8+ or CD4+ T cells, either individually or in combination, did not significantly alter protection against re-infection amongst mice with prior infection (S1B Fig). Thus, protection against ZIKV re-infection primed by primary infection is associated with retained accumulation of high titer viral-specific antibodies that can adoptively transfer protection to naive recipients. Together with the protective benefits of exogenously administered ZIKV-specific IgG antibodies shown in recent studies [43, 47], these findings suggest high titer ZIKV antibodies that inhibit viral infectivity can bypass the necessity for CD8+ T cells during primary infection [41, 48].
Given the sharply increased morbidity associated with ZIKV infection during pregnancy—with ensuing congenital invasion of fetal tissues, we further investigated whether protection primed by preconceptual infection persists during pregnancy. Three weeks after primary asymptomatic infection, allogeneic pregnancies were established amongst C57BL/6 (H-2b) female mice by mating with Balb/c (H-2d) males to recapitulate the natural heterogeneity between maternal and fetal antigens in outbred populations. Type I IFN receptor blockade was subsequently initiated in pregnant females at midgestation (E10.5), followed by ZIKV infection one-day later (E11.5) (Fig 4A). Interestingly, despite immunological shifts required for averting fetal rejection, the presence of expanded fetal target tissue and diminished responsiveness of maternal T cells to primary ZIKV infection during pregnancy [49], protection against secondary challenge was maintained amongst mice with prior preconceptual infection shown by sharply reduced ZIKV RNA levels in their serum, spleen, liver and brain compared with primary prenatal infection in naive control mice (Fig 4A). Importantly, congenital invasion was also efficiently averted as ZIKV RNA levels were reduced to near or below the limits of detection for most concepti (fetal plus decidual tissue) after prenatal challenge of mice with prior preconceptual infection (Fig 4A).
Protection against re-infection during pregnancy paralleled persistence of high titer circulating ZIKV IgG antibodies that efficiently neutralized ZIKV plaque formation in vitro (loss of activity for both after 103 to 104-fold dilutions) following preconceptual primary infection compared with naive control mice (Fig 4B and 4C). Interestingly, despite high levels of neutralizing antibodies in maternal serum following preconceptual primary infection, sporadic, low level placental ZIKV dissemination occurred with secondary prenatal infection (Fig 4A). These findings are consistent with breakthrough congenital invasion in mice receiving preconceptual vaccination or exogenous ZIKV-antibody transfer prior to prenatal infection [20, 21, 43]. To investigate the possibility that vertically transferred maternal antibodies may provide additional protective benefits, the levels of ZIKV-specific antibodies were compared with viral RNA levels amongst individual concepti. ZIKV-specific IgG levels were sharply increased in nearly all concepti homogenates of pregnant mice with prior preconceptual infection compared with the concepti of naive control mice, with a highly significant inverse correlation with ZIKV RNA levels amongst individual concepti scattered across multiple litters with low level breakthrough infection (Fig 4D). Likewise, ZIKV infectivity of Vero cells was significantly reduced by pre-incubation with UV-inactivated fetal tissue homogenates of pregnant mice with prior preconceptual infection compared with the concepti of naive control mice (Fig 4E). Together with recent studies demonstrating protection against congenital ZIKV transmission primed by mRNA or live attenuated viral vaccine platforms [20, 21], or amongst mice exogenously administered ZIKV human monoclonal antibody with high neutralization potency [43], these findings highlight the protective capacity of virus-specific neutralizing antibodies in overturning the natural vulnerability of mothers and their fetal offspring to ZIKV prenatal infection.
ZIKV is believed to have originated in East Africa, with subsequent mutation into unique West African and Asian variants [50]. Despite incomplete information on whether selective pressures by host immunity drive these antigenic shifts, the existence of unique ZIKV lineage strains has important practical implications for the scope of protection primed by natural infection or vaccination. For example, while purified ZIKV-specific human monoclonal antibodies can effectively neutralize both Asian and African strains, a single amino acid mutation in ZIKV ENV protein can override neutralization by individual antibody clones [43]. Furthermore, while antibodies primed by infection with the related flavivirus, Dengue (DENV), protect against secondary infection by identical viral serotype strains, they can also enhance susceptibility to re-infection by discordant DENV serotypes [51–53].
To investigate whether protective immunity primed by primary ZIKV infection extends to cross-lineage ZIKV strains, susceptibility to the original Uganda East African MR766 strain primed by prior infection with the Asian lineage PRVABC59 strain used in our preceding experiments was evaluated. We found the serum from mice with prior PRVABC59 infection efficiently neutralized plaque formation by the cross-lineage MR766 strain, and with near identical potency compared with monolayers infected with the homologous PRVABC59 virus (loss of activity between 103 to 104-fold dilutions) (Fig 5A compared with Fig 3A). In turn, the potency of MR766 neutralization by the serum of mice with prior PRVABC59 infection was not significantly impacted by pregnancy (Fig 5A).
In agreement with this cross-lineage susceptibility of MR766 to antibodies primed by prior PRVABC59 infection, ZIKV RNA levels were significantly reduced after secondary MR766 challenge amongst mice with prior asymptomatic PRVABC59 infection compared with primary MR766 ZIKV infection in naive control mice (Fig 5B). Importantly, cross-lineage protection primed by preconceptual prior infection is maintained during pregnancy shown by significantly reduced ZIKV RNA in the maternal serum spleen, liver, and brain, and amongst individual concepti (fetal plus decidual tissue) after MR766 secondary challenge in midgestation pregnant mice with preconceptual PRVABC59 infection compared with primary MR766 prenatal infection in naive control mice (Fig 5C). Significantly increased ZIKV IgG antibody titers that were inversely associated with ZIKV RNA levels were found in the tissue homogenate of nearly all concepti recovered from protected pregnant mice with preconceptual primary infection, but absent in concepti from naive control mice (Fig 5D). Thus, the broadly neutralizing capacity of serum after primary ZIKV infection in humans, non-human primates and mice [38, 40, 54], extends to cross-genotype protection against re-infection during pregnancy.
Despite identification nearly 70 years ago, ZIKV has remained a relatively obscure human pathogen until its emergence and global spread beginning in 2015. The unique propensity for congenital fetal invasion makes ZIKV distinct from other flaviviruses, and opens-up many unanswered fundamental questions for ZIKV and other pathogens that cause prenatal infection. These include whether strategies that protect against infection in non-pregnant healthy individuals remain effective despite pregnancy-associated anatomical changes that significantly expands susceptible target tissue to include fetal tissues, and increasingly recognized immunological shifts that avert maternal-fetal immunological conflict and dampen the proliferation-activation of T cells after infection during pregnancy [49, 55, 56]. Here we show preconceptual asymptomatic ZIKV primary infection protects against re-infection, and that these protective benefits are maintained during pregnancy (Fig 6). Thus, naturally acquired immunity against prenatal ZIKV infection is similar to the resistance of mothers to classical prenatal pathogens (e.g. varicella virus, rubella virus) that is also efficiently primed by preconceptual infection. A potentially unifying theme among these pathogens is diversity of protective epitopes that functionally minimizes the impacts of antigenic shifts amongst individual immune dominant microbe expressed antigens. For naturally acquired immunity to ZIKV, this notion is supported by the wide distribution of protective epitopes across spatially distinct domains of immune dominant envelope protein [13–18, 43, 57–59]. By contrast, the number of protective immune dominant epitopes is more limited for other viral pathogens (e.g. human cytomegalovirus, influenza virus) where protection after preconceptual infection is less reliable [25, 26].
These protective benefits conferred by pre-conceptual asymptomatic infection parallel sustained accumulation of ZIKV-specific antibodies in maternal serum that efficiently neutralize virus infectivity in vitro and reduce susceptibility to in vivo infection by the identical or heterologous ZIKV genotype strains. Interestingly, serum from mice primed by pre-conceptual infection efficiently neutralized ZIKV infectivity despite prior heat-inactivation, suggesting these protective benefits do not require complement or other heat-liable components associated with enhanced protection for other flaviviruses such as WNV [60]. On the other hand, antibody-mediated Fcγ receptor blockade overturned protection conferred by transfer of serum from mice with prior preconceptual infection. These results are consistent with our finding that primary asymptomatic infection selectively primes IgG2a and IgG2b antibodies with ZIKV ENV and NS1 specificity, and high affinity Fcγ receptor binding for these specific mouse antibody isotypes [61, 62]. Interestingly however, inactivation of Fcγ receptor binding for a human monoclonal IgG1 antibody with high affinity for ZIKV ENV protein does not significantly impact protection against prenatal ZIKV infection in mice [43]. Thus, further studies are needed to investigate how the necessity for Fcγ receptor can be functionally bypassed, in particular focusing on the importance of antibody isotype and/or antigen affinity, and the potential for cross-reactivity of ZIKV-specific antibodies with structurally homologous flaviviruses such as DENV that may promote antibody-dependent enhanced infectivity [63, 64].
Complexities inherent to investigating immunity and the pathogenesis of prenatal infection largely stem from the choice of experimental models that need to balance practicality with relevance to human pregnancy. Here, it is important to highlight that only humans have human placentas—with anatomical and molecular features that are not reproduced in any other species [65]. Likewise, the physiological discordance between maternal and fetal-expressed paternal antigens drive potent immunological shifts during pregnancy amongst humans and other outbred species that convey profound impacts on prenatal infection susceptibility [66]. For example, expanded accumulation of immune suppressive regulatory CD4+ T cells is highly accentuated in allogeneic compared with syngeneic pregnancies amongst inbred strains of mice [67]. In turn, expanded systemic accumulation of immune suppressive regulatory CD4+ T cells promotes maternal susceptibility to common prenatal pathogens (e.g. Listeria and Salmonella spp.), whereas dampened suppressive function of maternal regulatory CD4+ T cells fractures fetal tolerance and promotes congenital invasion in the context of allogeneic pregnancy [67–69]. Therefore, to extend the analysis of ZIKV prenatal infection that in mice has been limited to syngeneic pregnancies, MHC haplotype discordant strains of inbred mice were used for breeding to recapitulate the physiological mis-match between maternal-fetal antigens encountered in human pregnancy. Strategies that render mice innately susceptible to more prolonged and higher levels of ZIKV viremia by administration of type I IFN receptor blocking antibodies were exploited to overcome the natural resistance of murine STAT2 to degradation by ZIKV NS5 protein [34–37]. We further reasoned that unabated type I IFN responsiveness that attenuates ZIKV replication and protects against symptomatic disease makes this species ideally suited to investigate the immune response primed by preconceptual infection. Using this model where susceptibility to ZIKV can be temporally controlled by delayed administration of type I IFN receptor blocking antibody, we show even asymptomatic primary infection protects against re-infection, and that these protective benefits extend to re-infection during allogenic pregnancy.
Importantly, ZIKV primary infection in wildtype mice used to model human infections that are mostly subclinical and primarily associated with only transient self-resolving viremia does not replicate all aspects of human infection. For example, ZIKV infection in non-pregnant individuals remains asymptomatic despite presumed functional neutralization of type I IFN responsiveness though STAT2 inactivation [10, 11]. Thus, the enhanced vulnerability of type I IFN receptor deficient mice that almost uniformly develop symptomatic and often fatal infection [34, 70], suggests STAT2-indepenent, type I IFN-dependent cell activation pathways may also be important in protection against ZIKV symptomatic infection. Potential candidates are type I IFN induced STAT3 activation that overrides the pro-inflammatory effects of activated STAT1 and STAT2 in myeloid cells [71], or STAT5 phosphorylation that drives CD4+ T cell differentiation into FOXP3+ regulatory cells [72]. Nonetheless, despite this potential limitation regarding how asymptomatic primary infection is achieved, protective immunity efficiently primed by abortive infection we demonstrate in wildtype mice further highlights the immunogenicity of endogenous viral antigens recently shown with minimally-replicating live-attenuated viral strains, non-replicating inactivated virus or nucleic acid-based candidate vaccine formulations [13–18, 20, 21].
In the broader epidemiological context, antibody-mediated protection against re-infection during pregnancy primed by preconceptual ZIKV primary infection has important translational implications for new therapeutic strategies aimed at identifying at-risk individuals, and protecting expecting mothers and their fetal offspring. Considering an estimated attack rate that exceeds 90% [44], together with an ~80% rate of subclinical infection among individuals with newly acquired infection [11, 12], a majority of reproductive age women in ZIKV endemic areas likely have naturally acquired immunity against re-infection primed by resolved prior infections. Thus, despite very promising protective benefits having been recently shown for several ZIKV candidate vaccines in preclinical infection models involving non-pregnant animals or mice during syngeneic pregnancies [13–21], neither the safety of these formulations, nor their protective efficacy during pregnancies that recapitulate the heterogeneity between maternal and fetal expressed antigens representative of naturally outbred human populations have been established. Thus, our data highlighting that protection against re-infection primed by preconceptual infection are functionally retained during allogeneic pregnancy adds an important, but previously unaddressed perspective on how protection against prenatal infection can be achieved.
Our finding that primary abortive infection primes robust expansion of anti-ZIKV neutralizing antibodies that functionally persist during pregnancy and can be found amongst individual concepti proportional to their degree of protection, together with recent studies showing donor serum or purified human monoclonal antibodies can transfer protection against ZIKV infection in pregnant hosts [43, 47], points to serological screening for viral neutralizing antibodies as a practical approach for distinguishing susceptible at-risk individuals from those with naturally acquired immunity. Considering reduced placental transfer of maternal antibodies in mice compared with other mammalian species [73, 74], the 102 to 104-fold reduced ZIKV levels we demonstrate in individual mouse concepti likely underestimates the degree of protection achievable for human fetal offspring. In turn, the inverse correlation between ZIKV RNA levels amongst individual concepti following secondary infection and levels of anti-ZIKV IgG antibody suggest individual fetal offspring are near the threshold for minimal amount of passively transferred maternal antibody required for protection against congenital invasion in this model using mice rendered susceptible with type I IFN receptor blockade. Taken together, these results showing cross-lineage immunity against ZIKV prenatal infection conferred by primary abortive infection, together with protection primed by minimally replicating, attenuated ZIKV vaccine candidates [14, 20, 21], underscores the therapeutic potential of preconceptual strategies that prime accumulation of high titer neutralizing antibodies for broadly protecting susceptible reproductive age women. On the other hand, since primary ZIKV infection during early pregnancy in mice also primes the accumulation of neutralizing antibodies that are presumably similarly protective [49], averting congenital fetal invasion may require increased functional thresholds based on antibody levels and/or affinity to viral expressed antigens. Thus, important next-steps are to further investigate the degree of protection against ZIKV prenatal infection primed by preconceptual primary infection, and whether the reduced virus levels we find in maternal tissues are below the threshold required for pathological fetal infection in animals where gestational length, and placental expression of factors that influence antibody function (e.g. neonatal Fc receptor, complement regulatory protein Crry), are more representative of human pregnancy [73–75].
Experiments involving animals were performed under Cincinnati Children’s Hospital Institutional Animal Care and Use Committee (IACUC) approved protocols (Assurance Number 2013–0170). These protocols strictly adhere to recommendations described in the National Research Council’s “Guide for the Care and Use of Laboratory Animals” and American Veterinary Medical Association's "Report of the AVMA Panel on Euthanasia”.
C57BL/6 (H-2b) and Balb/c (H-2d) mice were purchased from the National Cancer Institute Charles River Laboratories (Frederick, Maryland), and maintained under specific-pathogen free conditions at the Cincinnati Children’s Hospital. For all Zika challenge studies, 6–8 week old, sex-matched C57BL/6 mice were randomly assigned to experimental groups. For experiments during gestation, 6–8 week old Balb/c males were used to sire allogeneic pregnancies in C57BL/6 females.
Zika virus (ZIKV) strains PRVABC59 (Puerto Rico, 2015) and MR766 (Uganda, 1947) were obtained from the US Center for Disease and Prevention (Atlanta, Georgia) and American Type Culture Collection (Manassas, Virginia), respectively [50]. Virus stocks were propagated and titred based on the number of plaque forming units (PFUs) in semi-confluent monolayers of Vero cells (American Type Culture Collection; Manassas, Virginia). All experiments were carried out under biosafety level 2 (BSL2) containment at Cincinnati Children’s Hospital.
For abortive infections, mice were inoculated subcutaneously in the lateral flank with 106 PFU ZIKV suspended in 100 μL sterile saline. For challenge studies, non-pregnant or allogeneic pregnant females (embryonic day 10.5) with and without primary abortive infection were administered 1 mg anti-type I IFN receptor antibody (MAR1-5A3; BE0241; BioXcell, West Lebanon, New Hampshire) intraperitoneally 24 hours prior to and on the day of infection, and were subsequently boosted (0.5 mg/dose) every five days thereafter. Mice were checked daily and assigned the following clinical disease score (1 healthy; 2 limited ruffled fur; 3 ruffled fur throughout; 4 mild lethargy; 5 limited movement; 6 moribund or uncontrolled spastic movements; 7 deceased) as previously described [76]. For serum harvest, blood from donor mice was obtained 21 days after initial infection, spun, filtered, and stored at -20°C. For heat inactivation, serum was incubated at 56°C for 30 minutes. For adoptive transfer, 300 μl serum (representing ~1/3 serum volume per donor animal) was administered i.p. to each recipient mouse one day prior to ZIKV infection. For Fcγ receptor neutralization in vivo, 250 μg of anti-CD16/CD32 antibody (clone 2.4G2, BioXcell) were administered intraperitoneally to mice day -1, 0 and 2 relative to ZIKV infection as described [45, 46]. For cell depletion, anti-CD4 (GK1.5; BE0003-1; BioXcell) and/or anti-CD8 (2.43; BE0061; BioXcell) antibodies (500 μg/mouse) were administered intraperitoneally to mice as previously described [68] one day prior ZIKV infection.
Following ZIKV infection, individual tissues were harvested and homogenized, whereas serum was collected after coagulation and centrifugation. Homogenized tissue and serum samples were extracted with the RNeasy mini kit (Qiagen, Hilden, Germany), and levels of ZIKV RNA evaluated by TaqMan (ThermoFisher, Waltham, Massachusetts) one-step quantitative reverse transcriptase PCR (qRT-PCR) on an ABI 7500 Fast instrument, using a previously described primer/probe set: forward primer, 5’-CCGCTGCCCAACACAAG-3’; reverse primer, 5’-CCACTAACGTTCTTTTGCAGACAT-3’; probe 5’/FAM/AGCCTACCTTGACAAGCAATCAGACACTCAA/NFQ-MGB/-3’ [34, 77]. Viral burden for each entire tissue was calculated by interpolation from a standard curve produced using serial 10-fold dilutions of ZIKV, and expressed on a log10 scale as number of viral copies.
Blood was collected from infected mice compared with uninfected controls and serum was isolated after coagulation and centrifugation. Each individual concepti (fetus plus placenta) was collected 3 days after prenatal infection from primed or naive pregnant females, homogenized in PBS (1 mL) and stored at -80°C. For ELISA, 96-well plates were coated overnight with purified ZIKV ENV (MBS319787) or NS1 (MBS319788) proteins, blocked with BSA (1%), incubated with serial dilutions of serum (starting at 1:10 dilution) or clarified fetal tissue homogenates (1:4 dilution). ZIKV-specific antibodies were probed with biotinylated secondary antibodies including rat anti-mouse IgA (clone 11-44-2; 13-5994-82; ThermoFisher), IgM (clone eB121-15F9; 13-5890-85; ThermoFisher), IgG1 (clone A85-1; 553441; BD Bioscience, San Jose, California), IgG2a (clone R10-15; 553388; BD Bioscience), IgG2b (clone R12-3; 553393; BD Bioscience), IgG3 (clone R40-82; 553401; BD Pharmingen), developed with streptavidin-peroxidase (554066; BD Bioscience) using o-phenylenediamine dihydrochloride as a substrate and reading absorbance at 450 nm (A450).
For viral neutralization, serial dilutions of the serum from mice with and without primary abortive infection were pre-incubated with 102 PFU ZIKV PRVABC59 or MR766 for 1 hour at 37°C. To investigate functionally neutralizing antibodies in concepti, clarified fetal homogenate was UV treated for 1 hour to inactivate any residual virus [78], and subsequently incubated with 102 PFU ZIKV PRVABC59 at a 1:100 dilution for 1 hour at 37°C. Following incubation, protection against plaque formation by each virus-antibody complex was assessed in Vero cell monolayers by first incubation at 37°C for 1 hour, followed by overlaying monolayer cells in each well with methyl cellulose (1%), and enumeration of plaques 72 hours thereafter as described [43].
All data were analyzed using GraphPad Prism software. For viral burden, levels of ZIKV RNA within each individual data set, were analyzed using the non-parametric Mann-Whitney test (two groups) or ANOVA (3 or more experimental groups). Linear regression was performed to determine correlation between ZIKA RNA and ZIKA-specific IgG levels in fetal tissue homogenates. P < 0.05 was taken as statistical significance.
|
10.1371/journal.ppat.0030107 | Epigenetic Silencing of Plasmodium falciparum Genes Linked to Erythrocyte Invasion | The process of erythrocyte invasion by merozoites of Plasmodium falciparum involves multiple steps, including the formation of a moving junction between parasite and host cell, and it is characterised by the redundancy of many of the receptor–ligand interactions involved. Several parasite proteins that interact with erythrocyte receptors or participate in other steps of invasion are encoded by small subtelomerically located gene families of four to seven members. We report here that members of the eba, rhoph1/clag, acbp, and pfRh multigene families exist in either an active or a silenced state. In the case of two members of the rhoph1/clag family, clag3.1 and clag3.2, expression was mutually exclusive. Silencing was clonally transmitted and occurred in the absence of detectable DNA alterations, suggesting that it is epigenetic. This was demonstrated for eba-140. Our data demonstrate that variant or mutually exclusive expression and epigenetic silencing in Plasmodium are not unique to genes such as var, which encode proteins that are exported to the surface of the erythrocyte, but also occur for genes involved in host cell invasion. Clonal variant expression of invasion-related ligands increases the flexibility of the parasite to adapt to its human host.
| Plasmodium falciparum is responsible for the most severe forms of human malaria. Invasion of host erythrocytes is an essential step of the complex life cycle of this parasite. There is redundancy in many of the interactions involved in this process, such that the parasite can use different sets of receptor–ligand interactions to invade. Here, we demonstrate that the parasite can turn off the expression of some of the proteins that mediate invasion of erythrocytes. Expression can be turned off without alterations in the genetic information of the parasite by using a mechanism known as epigenetic silencing. This is far more flexible than genetic changes, and permits fast, reversible adaptation. Turning on or off the expression of these proteins did not affect the capacity of the parasite to invade normal or modified red cells, which suggests that the variant expression of these genes may be used by the parasite to escape immune responses from the host. Parasite proteins that participate in erythrocyte invasion are important vaccine candidates. Determining which proteins can be turned off is important because vaccines based on single antigens of the parasite that can be turned off without affecting its growth would have little chance of inducing protective immunity.
| Invasion of human erythrocytes by merozoites of the malaria parasite P. falciparum is an essential step of the asexual blood cycle of the parasite, which is responsible for all the pathology associated with the disease. Erythrocyte invasion by merozoites of P. falciparum is relatively well characterised at the ultrastructural level [1], but the precise molecular interactions and the role of the specific parasite proteins are still poorly described. Proteins located on the surface of the merozoite probably mediate the initial, reversible contact through low affinity interactions. The next step involves reorientation of the merozoite, such that its apical end, which contains specialised organelles like rhoptries and micronemes, faces the erythrocyte. This is followed by the irreversible formation of a tight moving junction based on high affinity interactions (reviewed in [2]). The micronemal protein EBA-175 is believed to participate in junction formation by interacting with the erythrocyte surface protein glycophorin A [3], but other proteins located in the apical organelles of the merozoite are also likely to be involved. Strong candidates are the proteins encoded by the small gene families eba (also known as dbl-ebp, to which eba-175 belongs) and pfRh (also known as pfnbp or pfrbl), and recent data suggest that some members of the two families may have overlapping roles [4]. All individual members of these two gene families seem to be non-essential, as they can be knocked out without impairing parasite growth (reviewed in [2]). AMA1 and thrombospondin repeat domain–containing proteins may also play a role in the formation or migration of the junction [2]. Whatever the precise proteins involved, it is clear that there is redundancy in many of the ligand–receptor interactions between the apical end of the parasite and the erythrocyte. The particular set of receptor–ligand interactions used for invasion determine the so-called alternative invasion pathways, which are described by the sensitivity of invasion to treatment of erythrocytes with various enzymes. It is well established that both field isolates and laboratory-adapted parasite lines vary in their capacity to use the different invasion pathways [5,6].
In the next stage of invasion, the junction migrates towards the posterior end of the parasite driven by a parasite actin-myosin motor, creating an invagination in the erythrocyte membrane. The final step is the sealing of the invagination, which forms a parasitophorous vacuole where the parasite will reside until the next cycle of invasion. Although it is not well established which proteins participate in the formation of the vacuole and remodelling of the erythrocyte, it is possible that the high molecular mass rhoptry complex (RhopH complex) has a role [7], in addition to a possible role at earlier stages of invasion. Two of the components of this trimeric complex, RhopH2 and RhopH3, are encoded by single copy genes, whereas RhopH1/Clag can be encoded by five different genes of the clag gene family [8].
We recently described two essentially isogenic parasite lines both derived from the cloned parasite line 3D7 but maintained independently for several years, 3D7-A and 3D7-B. These two parasite lines differ dramatically in their capacity to use invasion pathways that permit entry into mutant and enzyme-treated erythrocytes. Remarkably, 3D7-A can invade erythrocytes sequentially treated with neuraminidase plus trypsin, which are completely resistant to invasion by 3D7-B and by all other parasite lines tested [9]. We hypothesised that the different invasion phenotypes of 3D7-A and 3D7-B parasite lines might be accounted for by differences in the expression of some invasion-related genes. In order to identify invasion-related genes under variant expression and to identify the genes responsible for the different invasion phenotypes of 3D7-A and 3D7-B, we performed a microarray comparison of the two parasite lines. While we did not identify the genes responsible for the different invasion phenotypes of the two parasite lines, our experiments led to the identification of invasion-related genes that exhibit variant expression. We also analysed the transcription of invasion-related genes in subclones of 3D7-A and demonstrated the epigenetic nature of the silencing observed for several genes.
Microarray experiments were used to compare tightly synchronised schizonts of the parasite lines 3D7-A and 3D7-B (Dataset S1). After excluding from the analysis all genes from the large var, rif, and stevor families, which are unlikely to participate in the process of erythrocyte invasion, three genes were found to be expressed at very different levels (more than 5-fold difference) between 3D7-A and 3D7-B: eba-140 (MAL13P1.60, also known as baebl), pfg27/25 (PF13_0011), and acylCoA binding protein gene (acbp) on Chromosome 14 (acbp-14, PF14_0749). The three genes were expressed at much higher levels in 3D7-B than in 3D7-A, with fold differences of 12.4, 11.2, and 7.7, respectively. Most genes presumed to be involved in erythrocyte invasion were expressed at similar levels between 3D7-A and 3D7-B, with the exception of the aforementioned eba-140 (see Figure 1 for the best-characterised genes). The 95% identical genes clag3.1 and clag3.2 were expressed at different levels in the two parasite lines, but the difference was only about 2-fold. However, the high level of identity between the two genes most likely resulted in cross-hybridization of some of the probes and consequent underestimation of the differences (see below).
eba-140 is the only gene among those expressed at very different levels between the two parasite lines that is known to participate in erythrocyte invasion [10], whereas pfg27/25 is known to have an important role in gametocyte development [11], and the function of acbp-14 is not known. We first aimed to determine whether lower eba-140 transcript abundance in 3D7-A compared to that of 3D7-B resulted in lower abundance of EBA-140 protein. Western blot and immunoprecipitation experiments revealed that EBA-140 was abundant both in schizont extracts and in culture supernatants of 3D7-B, but only present at very low levels in 3D7-A (Figure 2A and 2B). The multiple EBA-140-specific bands in Figure 2A and 2B correspond to proteolytic processing products. Furthermore, erythrocyte binding assays revealed a very different composition of proteins that bind to erythrocytes in supernatants from 3D7-A or 3D7-B (Figure 2C). A band of approximately 175 kDa with an identical mobility to EBA-175 (as determined by western blot on samples run side by side, unpublished data) and a band of approximately 152 kDa were observed for both 3D7-A and 3D7-B, but two strong bands with a mobility identical to two of the EBA-140-immunoprecipitated bands were observed only in 3D7-B.
EBA-140 was present by immunofluorescence assay (IFA) in the apical tip of 100% of EBA-175-positive segmented schizonts and free merozoites in 3D7-B, where the two proteins co-localised. The pattern was consistent with the previously described micronemal location for both proteins [10]. In contrast, only a very low percentage (about 7%) of EBA-175 positive schizonts were positive for EBA-140 in 3D7-A (Figure 2D).
To detect differences in the expression of invasion-related proteins that might have escaped microarray analysis, we compared radiolabelled culture supernatants from 3D7-A and 3D7-B by SDS-PAGE. Two abundant bands were present in 3D7-A but not in 3D7-B (Figure 3A). A very high molecular mass band corresponds to a form of PfRh2b with an insertion that is present only in 3D7-A [12], as demonstrated by western blot with anti-PfRh2b antibodies on samples run side by side (Figure 3B). The other band has an electrophoretic mobility of approximately 148 kDa and forms a doublet with a band of slightly higher mobility. Because the size of these bands is similar to the size of RhopH1/Clag, we immunoprecipitated supernatants of the two lines with an anti-RhopH2 monoclonal antibody that immunoprecipitates the whole RhopH complex. The immunoprecipitated complex contained an additional band in 3D7-A with an identical mobility to the band present in supernatants of 3D7-A but not of 3D7-B, indicating that the protein present only in 3D7-A supernatants is part of the RhopH complex (Figure 3C). Mass spectrometry analysis revealed the identity of this polypeptide as Clag3.2 (PFC0110w) (25% coverage), whereas the lower band of the doublet (also present in 3D7-B, arrowhead in Figure 3A) was identified as Clag3.1 (PFC0120w) in both parasite lines (33% coverage). Despite the 95% identity between the two proteins, the identification was unambiguous because it was based on three peptides that were specific for one or the other protein (Table 1). As expected from these results, reverse transcriptase (RT)-PCR analysis revealed that clag3.1 transcripts are present at similar levels in 3D7-A and 3D7-B schizonts, but clag3.2 transcripts are almost absent in 3D7-B (Figure 3D).
To determine whether further heterogeneity in the expression of invasion-related genes occurs in the cloned parasite line 3D7, we analysed expression of these genes in 11 subclones of 3D7-A (described in Materials and Methods). Silver-stained SDS-PAGE analysis of culture supernatants from these 11 subclones revealed that all of them expressed either Clag3.1 or Clag3.2, but none of them expressed both (Figure 4A, top panels). Each subclone had gone through at least 11 cycles of replication from a single parasite to harvesting for analysis. The subclones reflect the clonally transmitted expression pattern of the individual parasites from which they originated. This result indicates that 3D7-A is a mixture of parasites expressing one or the other protein. Mutually exclusive expression of the two genes was confirmed by semi-quantitative RT-PCR (Figure 4A, middle panels). All clones that expressed clag3.1 at high levels had only low-level expression of clag3.2 and vice versa. Furthermore, another member of the clag family, clag2, also showed clonal variant expression between the subclones, whereas clag8 and clag9 were expressed at very similar levels in all subclones (Figure 4A). Western blot analysis with antibodies specific for Clag2 and Clag3.2 confirmed that low levels of transcripts resulted in low abundance or absence of the corresponding proteins in culture supernatants (Figure 4A, bottom panels). Expression patterns remained stable over continuous culture for at least one additional month (Figure S1A). Interestingly, a stock of the cloned line HB3 at Ehime University (Japan) derived from HB3B, which had been passed through chimpanzees [13], expressed only clag3.1, but a stock at the same university derived from HB3A (prior to chimpanzee passage) expressed only clag3.2, supporting the idea of mutually exclusive expression and switching between the two genes (Figure S2). HB3 at the National Institute for Medical Research (United Kingdom) and W2mef lines only expressed clag3.1 at detectable levels (unpublished data).
We also analysed by RT-PCR the expression of members of other multigene families involved in erythrocyte invasion. All members of the eba family were expressed at similar levels in the 11 3D7-A subclones except for eba-140, which was silenced in most subclones but expressed at levels similar to that of 3D7-B in two of them (Figure 4B). Western blot analysis of schizont extracts revealed that abundance of EBA-140 protein correlated well with transcript abundance (Figure 4B, bottom panels). The pattern of expression of EBA-140 in 3D7-A subclones was fully consistent with the result of IFA experiments (Figure 2D).
Expression of some members of the PfRh family had previously been described as varying between non-isogenic parasite lines [14–16]. We found only small differences in the level of expression of these genes among our subclones, with the exception of pfRh2b that was expressed at low levels in the subclones 4D and W4–1 to W4–4 (Figures 4C and S1B). See Text S1 for an explanation of the confounding effect of small differences in the stage of the parasites and the controls developed to overcome this difficulty.
Variant expression among subclones was also observed for two genes, pfg27/25 and acbp-14, which were expressed at higher levels in 3D7-B than in 3D7-A according to the microarray analysis (Figure 4D). Analysis of the subclones revealed that these genes are also silenced in some individual parasites but expressed at levels similar to that of 3D7-B in others. acbp-14 belongs to a four-gene family [17], but expression of the other genes of this family was similar in all 3D7-A subclones (Figure 4D and Text S1).
Despite differences in the expression of several invasion-related proteins, all the 3D7-A subclones tested had very similar growth rates as determined in a one-cycle FACS-based growth assay (Table 2). Likewise, the capacity to invade erythrocytes treated with various enzymes was very similar among all the 3D7-A subclones tested and indistinguishable from that of the parental 3D7-A, but different from that of 3D7-B (see Figure 5A and Figure 5F for comparison). All the 3D7-A subclones invaded erythrocytes sequentially treated with neuraminidase plus trypsin, which are completely resistant to invasion by 3D7-B. Thus, the expression status of clag2, clag3.1, clag3.2, eba-140, acbp-14, pfg27/25, or pfRh2b did not affect the invasion pathways used by the parasites. This was confirmed by selection-based experiments (Figure S3).
We used an additional approach to confirm that silencing of eba-140 does not alter the invasion phenotype of the parasites and is not necessary for the invasion of neuraminidase plus trypsin–treated erythrocytes. We expressed eba-140 in 3D7-A parasites under the control of its own promoter on an episome. We tried three different constructs that contained 0, 797, or 1,331 bp of the region upstream from the eba-140 start codon to drive the expression of the episomal transgene (Figure 5B). Transfection with these constructs resulted in production of EBA-140 protein only when the longer version of the 5′ region was used (E140-1300), as determined by western blot (Figure 5C). The timing of expression of the episomal eba-140 was correct, because transcripts of this transgene were undetectable by RT-PCR in ring or trophozoite stages, but were highly abundant in schizonts, as observed for authentic eba-140 in 3D7-B (Figure 5D). Furthermore, episomally expressed EBA-140 co-localised with EBA-175 by IFA (Figure 5E), indicating that it is correctly located in the apical organelles. Altogether, these results indicate that this large protein can be correctly expressed from an episome. However, only 65% of EBA-175-positive schizonts were EBA-140 positive, probably due to defective segregation of the episome.
E140-1300-transfected 3D7-A parasites had an invasion phenotype indistinguishable from that of 3D7-A, but clearly distinct from that of 3D7-B (Figure 5F). Despite expressing EBA-140, these parasites invaded erythrocytes double-treated with neuraminidase plus trypsin as efficiently as 3D7-A parasites, in contrast to 3D7-B parasites that completely failed to invade them. Invasion assays were performed in the absence of drug. To rule out the possibility that only parasites that had lost the episome (and consequently were not expressing EBA-140) were able to invade double-treated erythrocytes, parasites that had invaded these erythrocytes were placed back under drug pressure and found to have survival rates similar to those invading control erythrocytes.
To determine whether there were genetic differences in the eba-140, clag3.1, or clag3.2 genes associated with their expression status, we analysed these loci in parasite lines where the genes were either expressed or not. Southern blot analysis of the eba-140 locus, covering the open reading frame (ORF), and also 4.1 kb upstream from the start codon and 3.8 kb downstream from the stop codon, did not reveal any difference between 3D7-A and 3D7-B parasites (Figure 6A and 6C). Furthermore, an eba-140-specific PCR product was amplified from genomic DNA from all 3D7-A subclones, including those that do not express the gene. Similarly, Southern blot analysis of the chromosomal region where clag3.1 and clag3.2 are located, including the region between the two genes and 2.9 kb upstream from the start codon of clag3.2 and 6.4 kb downstream from the clag3.1 stop codon, did not reveal any difference between parasite lines that only express clag3.1 at high levels (3D7-B and 4D), a parasite line that only expresses clag3.2 (10E), and 3D7-A, which is a mixture of parasites that express one or the other gene (Figure 6B and 6D). Altogether, these results rule out the possibility that low expression of eba-140, clag3.1, or clag3.2 was associated with major chromosomal rearrangements, deletions, or recombination of the two clag genes in Chromosome 3 to form a single gene.
Furthermore, direct sequencing of PCR products spanning the ORF of eba-140 and 1.3 kb upstream from its start codon did not reveal any difference between 3D7-A and 3D7-B. We also PCR amplified the full clag3.1 and clag3.2 ORFs in these two parasite lines with primers in their divergent 5′ and 3′ UTR sequences. Sequencing of these PCR products revealed no difference between the two parasite lines or compared with the published sequences for the genes, which rules out the possibility that a gene conversion event had occurred.
The two genes expressed at most different levels between 3D7-A and 3D7-B as determined by microarray analysis, eba-140 and pfg27/25, are located in the left subtelomeric region of Chromosome 13 (Figure 7A), at a distance of 89.4 and 121.8 kb from the telomere, respectively. This is suggestive of coordinated regional silencing of this subtelomeric zone in 3D7-A. None of the genes located between eba-140 and pfg27/25 or between the telomere and eba-140 is expressed at high levels in schizonts [18,19]; thus, our microarray experiments were unable to determine whether they are also differentially expressed between 3D7-A and 3D7-B. We prepared RNA from tightly synchronised ring- (11–16 h), trophozoite- (23–28 h), or schizont-stage parasites (39 h and 30 min to 44h and 30min for 3D7-A, 41–46 h for 3D7-B) and performed RT-PCR analysis of eba-140, pfg27/25, and four additional genes located in this chromosomal region (Figures 5D and 7). Two of the genes with a peak of expression in rings, gbph2 and MAL13P1.61, were expressed at slightly higher levels in 3D7-B than in 3D7-A, and the same was true for the member of the rif family, PF13_0006, which was unexpectedly expressed in schizonts. On the other hand, PF13_0076 was expressed at all stages and at similar levels in both parasite lines (Figure 7B). Thus, higher expression in 3D7-B than in 3D7-A occurred for several genes in this chromosomal region, but different genes were affected to different extents. The most marked differences were observed for genes expressed in late schizonts (eba-140 and pfg27/25).
Integration of the construct E140–0 (Figure 5B) in the eba-140 locus by a single-recombination event was achieved by cycling E140–0-transfected 3D7-A parasites for two cycles on/off drug and confirmed by Southern blot (Figure 8A and 8B). The integration of the construct resulted in the duplication of the eba-140 gene, but only one copy was preceded by sequences with promoter activity (Figures 8A and 5B). Western blot analysis revealed that integration of this plasmid resulted in expression of the eba-140 gene, though at a lower level than in 3D7-B (Figure 8C).
E140–0-transfected and -drug-cycled 3D7-A parasites were subcloned by limiting dilution. Five of the resulting subclones were analysed by Southern blot, which revealed that three of them had integrated one copy of the gene (W4–1, W4–2, and W4–5), whereas one had integrated two or more copies (W4–3) and one was wild type (W4–4) (Figure 8B). Subcloning was performed in the absence of drug pressure. RT-PCR and western blot analysis of these subclones revealed that only one of them (W4–1) expressed EBA-140 at high levels, whereas all the other subclones expressed it at low levels similar to that of the parental 3D7-A (Figure 4B). The three subclones that had one copy of the construct integrated were maintained in parallel either in the absence or presence of drug selection for 1 mo, and RNA from tightly synchronised schizonts of the resulting populations was analysed by RT-PCR. The presence of the drug resulted in a large (W4–1 and W4–2) or moderate (W4–5) increase in the abundance of transcripts of the drug resistance gene hdhfr. The increase was paralleled by a dramatic increase in the abundance of eba-140 transcripts in the clone W4–2, whereas expression of this gene was not affected in the clone W4–1, which already expressed eba-140 at high levels before drug selection, and only moderately increased in the clone W4–5 (Figure 8D). This result indicates that insertion of a gene (hdhfr) that is forced to be active (by drug selection) in the vicinity of eba-140 can cause the activation of this gene. Thus, eba-140 can be activated in situ, ruling out the possibility that undetected genetic changes were responsible for the silencing and indicating that silencing was epigenetic. Expression of the gene pfg27/25, which is more distal to the telomere than eba140 and was already active in the absence of drug in the three clones (Figure 4D), was affected to a lower extent (Figure 8D).
The process of erythrocyte invasion by merozoites of P. falciparum involves several essential, highly conserved interactions as well as dispensable, redundant interactions. Here we show that several of the genes responsible for the latter can be epigenetically silenced. Some members of the pfRh family had been shown to vary in expression between non-isogenic cloned parasite lines [14,15] and among field isolates [16]. Furthermore, switching from sialic acid–dependent into sialic acid–independent invasion in the two related parasite lines W2mef and Dd2 involved increased expression of PfRh4 [4,20]. Here, we extend the observation of variant expression to several other invasion-related gene families in isogenic parasite lines and demonstrate that silencing is transmitted epigenetically.
The comparison of the two isogenic parasite lines 3D7-A and 3D7-B revealed differences in the expression of eba-140, clag3.2, pfg27/25, and acbp-14. To determine whether further heterogeneity exists within the cloned line 3D7, we analysed the expression of invasion-related genes in 11 subclones of 3D7-A, which reflect the pattern of expression in the 11 3D7-A individual parasites from which they originated. Clonal variant expression was detected for three additional genes, clag2, clag3.1, and pfRh2b. In all cases tested, mRNA abundance reflected protein abundance. Thus, 3D7-A is a mosaic of parasites expressing different combinations of invasion proteins. It will be important to determine to what extent this mosaicism occurs in natural parasite populations.
A main feature of all the invasion-related genes showing variant expression is that they belong to small multigene families. The var genes, which are the paradigm of variant expression in Plasmodium, are also part of a multigene family, though of a much larger size. var genes, which participate in both immune evasion and cytoadhesion of infected erythrocytes, exhibit mutually exclusive expression, such that only one gene of the family is expressed at a time [21]. In the case of invasion-related multigene families, we observed mutually exclusive expression for two members of the clag family, clag3.1 and clag3.2.
We did not detect any DNA alteration associated with the active or silent state of invasion-related genes. Although we cannot exclude the possibility of minor alterations that escaped our analysis, or that regulatory regions where alterations occurred were located in distant regions of the chromosome, the most plausible explanation is that, at least in the cases of clag3.1, clag3.2, and eba-140, silencing was transmitted epigenetically. This is again reminiscent of the situation for var genes, which switch from a silent to an active state without detectable DNA alterations [21] and which are epigenetically silenced in a process that involves modifications of the chromatin structure [22,23]. Furthermore, approximately two-thirds of var genes and the majority of members of small invasion-related multigene families for which we detected variant expression are located in subtelomeric positions (Figure S4), though var genes are always more proximal to the telomere. It will be important to determine whether the subtelomeric location of these invasion-related genes is critical for their variant expression or, instead, is related to their evolution.
In addition to the absence of genetic alterations between an active and a silenced state, reversibility and region-specific rather than sequence-specific effects are hallmarks of epigenetic silencing and heterochromatin. Both were demonstrated for the silencing of eba-140. Insertion of the drug resistance gene hdhfr in the vicinity of the eba-140 locus and subsequent drug pressure resulted in the in situ activation of this gene in some subclones. This suggests that activation of hdhfr disrupted a compact, “closed” conformation of the chromatin around this locus and forced the transition to a more relaxed, transcriptionally active conformation that spread into the neighbour eba-140. Furthermore, silencing of eba-140 in 3D7-A was somehow coordinated with silencing of another gene located in the same subtelomeric region, pfg27/25. This locus was silenced in some subclones of 3D7-A but expressed in others, but in all cases where the more telomere proximal eba-140 gene was active, pfg27/25 was also active. This suggests a model in which silenced chromatin would spread from a telomeric position. The extent of the silenced area would vary stochastically between individual parasites, but once established it would be clonally inherited, which would explain the variegated expression of the two genes in the different subclones. In some subclones, the silenced area would spread as far into the chromosome as the pfg27/25 locus, while in others it would only reach eba-140 but not pfg27/25, and in others it would not even reach eba-140. However, the observation that other genes located in the same region were only silenced to a lower extent or not silenced at all does not support this model, and suggests that either the silenced chromatin structure is only formed late in the life cycle of the parasite, or a mechanism other than heterochromatin spreading is responsible for the coordinated silencing of these genes. Further experiments will be needed to distinguish between these two possibilities. It will also be important to determine whether transcripts from PF13_0076 in schizonts, which occurred at similar levels in 3D7-A and 3D7-B, represent active transcription at this stage or carry over of mRNA transcribed in previous stages.
In contrast to the apparently region-specific, sequence-independent variegated silencing of the eba-140 locus, silencing of clag3.1 or clag3.2 was promoter specific, because the two genes lie adjacent to each other in the genome and expression was mutually exclusive. This situation is more similar to that observed for var genes, where one gene can be activated while its neighbours remain silenced [24]. The chromosomal organization of clag3.1 and clag3.2 resembles that of two other invasion-related genes with a high level of identity, pfRh2a and pfRh2b, which lie adjacent to each other near the centromeric region of Chromosome 13. However, in that case, expression is coordinated rather than mutually exclusive, with all parasite lines analysed so far expressing either none or both of the genes (with the exception of parasite lines in which one of the genes is missing) [14–16,25]. The contiguous genes PfRh4 and eba-165 also seem to be co-regulated [4,20].
This complex picture reveals the existence of multiple different ways of regulation of the expression of genes encoding P. falciparum erythrocyte invasion proteins. Although the pattern of silenced and expressed genes may be very different in parasites living in the context of a host with acquired immune responses, our results on culture-adapted parasite lines demonstrate the existence in P. falciparum of the molecular machinery for the silencing of these genes in addition to the life cycle–dependent silencing common to most Plasmodium genes [18]. Epigenetically transmitted transcriptional silencing of invasion-related genes together with a certain level of mosaicism in a parasite population provides an enormous flexibility and capacity to adapt rapidly to changing host environments by simple means of natural selection and stochastic, low-frequency switching on and off of the expression of these genes.
The biological role of variant expression of invasion-related genes is unknown, and we can only speculate about its possible functions. While targeted disruption of the invasion-related genes eba-175 and pfRh2b resulted in changes in the invasion pathways used by the parasites where they were disrupted [15,26], the active or silent state of the variantly expressed genes described in this study was not associated with detectable differences in growth rates or in the invasion pathways used. Regardless of the combination of silenced and expressed invasion-related genes, all 3D7-A subclones were able to invade using the neuraminidase- and trypsin-resistant receptor A [9], indicating that none of the genes found to vary in expression is responsible for this interaction. The identical invasion phenotype of all 3D7-A subclones suggests that the main force driving the variant expression of invasion-related genes is not the acquisition of the ability to invade different types of erythrocytes, but instead is immune evasion. This is a reasonable hypothesis, considering that in human populations diversity of the erythrocyte receptors that the parasite uses for invasion is somehow limited, whereas merozoites are exposed to protective immune responses that are extremely diverse between different hosts. However, when silencing affects other invasion-related genes like eba-175 or occurs in other genetic backgrounds, it will clearly affect the types of erythrocytes susceptible to invasion. Epigenetic silencing of invasion-related genes is likely to be the mechanism behind switching between invasion pathways.
The number of genes in the invasion-related multigene families is small for a role in immune evasion, especially when compared with the size of well-characterised variantly expressed gene families such as the var genes in P. falciparum or the vsg genes in Trypanosoma brucei. However, all members of the clag, eba, and pfRh families are polymorphic to some extent, and in some cases polymorphism is the consequence of positive selection [27,28]. Polymorphism, which is well known to help the parasite escape immune responses, and variant expression based on a limited number of alternatives, may play additive or even synergistic roles in immune evasion. Variant expression has the potential to enhance the capacity of polymorphism to avoid immune responses by permitting selection of parasites that keep in a silenced state the multigene family members with polymorphic allelic forms that are better recognised by protective immune responses in a particular host.
In summary, we describe an additional layer of complexity of the process of erythrocyte invasion, showing that the parasite has multiple modes of controlling the expression of genes involved in this process. This may provide an advantage to the parasite in its constant race to escape the immune system of its human host.
Parasite cultures were maintained under standard conditions in medium containing Albumax II. The parasite lines 3D7-A and 3D7-B are the same cloned parasite line 3D7 maintained in different laboratories for several years. Genotyping was used to confirm their 3D7 identity [9]. The subclones 4D, 6D, 10E, 10G, 1.2B, and 1.2F were originated by subcloning 3D7-A by limiting dilution [12]. The subclones W4–1 to W4–5 were obtained by subcloning by limiting dilution 3D7-A parasites that had been transfected with the plasmid E140–0 and went through two cycles on/off drug to promote integration of the plasmid (see Figure 8 and Results). Genotyping of the highly polymorphic gene msp2 by HinfI restriction fragment length polymorphism (RFLP) [29] was used to confirm the 3D7 identity of all subclones and rule out contamination from other parasite lines (unpublished data). Details of the methods used for transfection and subcloning are provided in Text S1.
To determine the growth rate of different subclones of the parasite line 3D7-A, synchronised cultures of the different subclones were diluted with fresh erythrocytes to an approximate parasitaemia of 0.7% immediately after sorbitol lysis. After culturing for 15 to 20 h, when most of the parasites were at the late trophozoite or early schizont stage, parasitaemia was determined by FACS as described [30] (time 0) using a FACScalibur cytometer (Becton Dickinson, http://www.bd.com/). Parasitaemia was determined again by FACS 48 h later, and the growth rate determined as the ratio between the parasitaemia at the 48- and 0-h time points. Erythrocyte digestions and invasion assays were performed as described [9], with small modifications explained in Text S1.
To obtain RNA for microarray or RT-PCR analysis, parasites were synchronised to a 5-h window by purifying schizonts from a culture with abundant late forms on 70% Percoll, and removing late forms by sorbitol lysis 5 h later. Parasites were then left undisturbed for 39 h and 30 min (3D7-A and subclones) or 41 h (3D7-B), and harvested in 20-erythrocyte pellet volumes of Trizol (Invitrogen, http://www.invitrogen.com/). These times were determined in preliminary experiments for each parasite line as the times at which 20% of the schizonts had burst (estimated from the ratio of rings to schizonts). RNA in Trizol was frozen at −70 °C and later purified according to the manufacturer's instructions. RNA was then treated with RNAse-free DNAse I (Qiagen, http://www.qiagen.com/) and cleaned with the RNeasy MinElute cleanup kit (Qiagen). cDNA was obtained by reverse transcribing 0.5 μg of total RNA using the AMV reverse transcription kit (Promega, http://www.promega.com/) with oligo dT primers. To rule out the possibility of gDNA contamination, parallel reactions were performed for all samples in the absence of reverse transcriptase and tested by PCR with at least two primer pairs. To achieve semi-quantitative conditions, PCR was performed for only 25 cycles, and the amount of starting cDNA was adjusted for each primer pair to obtain bands that were clearly visible but not saturating. Single-copy genes with similar timing of expression [18,19] to the genes under analysis were used to control the amount of cDNA specific of each stage. All of the primers used in this study are described in Dataset S2.
The Affymetrix PFSANGER array (http://www.affymetrix.com/) was used for these experiments. Details of the array, experimental procedure, RMA normalization, and data analysis are available in Text S1.
The procedure used for the construction of the plasmids E140–0, E140–800, and E-140-1300 (Figure 5B) is explained in Text S1. The antibodies used in this study and their sources are also described in Text S1.
Schizont extracts for western blot were prepared by resuspending pellets of Percoll-purified schizonts into 20-pellet volumes of PBS, adding 40-pellet volumes of 2x SDS protein loading buffer and heating for 5 min at 95 °C before storing at −70 °C until use. NP-40 extracts of schizonts for immunoprecipitation were prepared approximately as described [14]. To obtain culture supernatants, tightly synchronised parasite cultures with abundant segmented schizonts were enriched for schizont-infected erythrocytes by gelatin flotation. Schizont-enriched fractions (typically 80% parasitaemia) were placed back in culture at a haematocrit of approximately 0.3% and supernatants harvested by centrifugation 13 to 20 h later. For the preparation of supernatants in Albumax-free medium, it was critical that the original culture only contained very mature forms, because otherwise it resulted in death of the parasites before rupture and under-representation of some of the proteins usually released into the culture supernatant (unpublished data). Metabolic labelling of parasites was achieved approximately as described [14].
Concentrated culture supernatants (from 5 ml of original supernatant) prepared in the absence of Albumax were loaded into a Q-Sepharose column in 0.5x PBS buffer. After washing with 0.5x PBS + 0.1 M NaCl, elution was performed with 0.5x PBS + 0.25 M NaCl. This fractionation resulted in a significant enrichment in the proteins of interest and elimination of haemoglobin. For mass spectrometry, fractions containing the proteins of interest were concentrated and resolved in 20 cm 5.5% polyacrylamide gels. The bands of interest were excised, reduced, alkylated, and trypsin-digested, then the released peptides were processed for mass spectrometry and analysed in a Reflex III MALDI-ToF mass spectrometer (Bruker Daltonics, http://www.bdal.de/). Data were analysed using Mascot software (Matrix Science, http://www.matrixscience.com/). Analysis of an excision in the 3D7-B lane corresponding to the position of Clag3.2 yielded no signal.
Details of the procedures used for these experiments are available in Text S1.
Microarray data have been deposited with ArrayExpress under accession number E-SGRP-9. The PlasmoDB (http://www.plasmodb.org/plasmo/home.jsp) accession numbers (systematic gene names/IDs) for the genes mentioned in this article are described in Dataset S2. |
10.1371/journal.pntd.0004835 | Pediatric Cutaneous Leishmaniasis in an Endemic Region in Turkey: A Retrospective Analysis of 8786 Cases during 1998-2014 | Cutaneous leishmaniasis (CL) is a major public health concern in Turkey and Sanliurfa represents the most endemic city in Turkey. Although children are most commonly affected by CL, detailed studies of pediatric CL in Turkey are lacking.
In this report we retrospectively evaluated clinical and epidemiological data of 8786 pediatric CL cases, and how children respond to antimonial therapy. CL was observed most frequently in children between 6–10 years old. Interestingly this group showed shorter duration of disease and smaller lesions compared to 0–5 year and 11–15 year old groups. Females were more affected in all groups. Lesion localization and types varied among groups, with 0–5 year old presenting head/neck and mucosal lesions, and more often suffered from recidivans type, this could be associated to the longest duration of the disease in this group. Eleven-15 year old group showed fewer lesions in the head/neck but more generalized lesions. Evaluation of treatment response revealed that intra-lesional treatment was preferred over intramuscular treatment. However, 0–5 year old received intramuscular treatment more often than the other groups. Furthermore, the majority of 0–5 year old group which received intra-lesional treatment did not received subsequent intra-lesional cycles, as did children in the range of 6–15 years old.
We report an increase in pediatric CL patients within the last four years. Analysis of pediatric CL patients by age revealed significant differences in CL progression. The data suggest that children between 0–5 years old responded better than other groups to intralesional treatment, since they received more often a single cycle of IL treatment, although follow up observation is required since they were more prone to develop recidivans. Eleven-15 year old patients comprise the largest percentage of patients receiving two or three cycles of intralesional treatment, suggesting that this group did not respond efficiently to intralesional treatment and highlighting the need for more effective therapeutic strategies against CL.
| In this study, 8786 pediatric (0-15years old) cutaneous leishmaniasis (PCL) patients were retrospectively evaluated for the epidemiological and clinical characteristics. From the records of PCL patients, we focused on the patients' age, gender, lesion type and location, diameter, number and duration of the lesions, and the treatment given. Of the patients, 4050 (46·1%) were males and 4736 (53·90%) were females, with a mean age of 7·52 years (range 0–15 years). Mean lesion diameter was 12·77±0·11 mm (range 1-100mm), while the average lesion duration and the number of lesions was 8·58±0·21 weeks and 1·76±0·01, respectively. The area most involved in the lesions of PCL patients was the head and neck region at a rate of 32·55% (n = 2230). The most (78·79%) of the PCL patients who received intralesional pentavalent antimonial therapy responded to the first cycle of treatment. It was observed that the PCL patients who applied systemic pentavalent antimonial therapy, resistant to intralesional pentavalent antimonial therapy, were completely recovered. We believe that this study will contribute to clinicians to evaluate the clinical features and therapy choice of pediatric CL patients.
| Cutaneous leishmaniasis (CL) is a skin infection caused by various species of the parasite Leishmania, and is spread by the bite of an infected female Phlebotomine sand flies [1,2]. Worldwide, there are 1·5–2 million new CL cases annually and it is estimated that 350 million people are at risk of infection [2]. The number of CL cases has significantly increased in Turkey in recent years representing a major public health problem [3]. In the Old World, CL is often caused by L. tropica and L. major [3–8]. Skin lesions caused by CL usually heal spontaneously or become chronic [9]. The vast majority of CL cases are seen in childhood. In epidemiological studies, it has been shown that children are at greater risk than adults[10–14].
Turkey is an endemic country for CL and the main causative agent is L. tropica. However, CL caused by L. infantum and L. major has also been reported especially from East part of the Mediterranean Region [15,16]. Within the last two years of this study (2012–2014), an outbreak of leishmaniasis occurred in Turkey, which was associated with the Syrian civil war and has severely affected local efforts at controlling the spread of disease [16,17]. CL affects mainly the Southeastern Anatolia region of Turkey, which is close to the border with Syria with about 61% of CL cases, and Sanliurfa represents the most endemic city in this region [8]. Previous studies have shown that 0–19 year old patients represent 60–70% of the total population in Turkey infected with CL, emphasizing the importance of evaluating this population in more detail. However, within this population, studies characterizing age and gender based incidence, as well as age-bias correlations in response to various forms and cycles of treatment have not been fully examined. These data are vital in the reevaluation of prevention and treatment strategies against CL, owing to its ever growing incidence in this region. In this study we examined the epidemiological and clinical characteristics of pediatric (0–15 years old) cutaneous leishmaniasis patients (PCL) over a 16-year period in the province of Sanliurfa, Turkey.
This study retrospectively evaluated 8786 PCL patients aged 0–15 years, who were registered at Harran University Medical Faculty Dermatology Department and the Public Health Oriental Boil Centre between 1998 and May 2014. Diagnosis of PCL was determined based on the patient’s clinical appearance, laboratory tests (positive demonstration of Leishmania parasites (amastigotes) in skin smears and evaluation of histopathology results). From the records of PCL patients at these centers we evaluated patient age, gender, lesion type (ulcer, papule, nodule, recidivans), lesion location (head-neck, trunk, upper extremity, lower extremity, mucosa, generalized), lesion diameter (mm), number of lesions, duration of lesions (weeks), and the treatment given (intralesional and systemic pentavalent antimonial treatment).
The study was approved by the local Clinical Research Ethics Committee (document date: May 29, 2014 and document number: 74059997.050.01.04/75). Patient medical data were anonymized in this study.
Treatment with intralesional pentavalent antimonial (Glucantime) therapy was administered twice a week for 4 weeks (a total of 8 injections) to all patients diagnosed with PCL disease at the Dermatology Clinic of Harran University Medical Faculty and the Oriental Boil Centre of the Sanliurfa Public Health Directorate.
At the 20-day assessment period after treatment with intralesional pentavalent antimonial therapy, the PCL patients who had not recovered were administered a second and third treatment cycle with a similar regimen as in the first treatment cycle (8 injections twice weekly for 4 weeks). PCL patients with more than 5 lesions who did not recover after intralesional pentavalent antimonial injection, those with lesions larger than 5 cm, those with genital and oral mucosa involvement and those with lesions in cartilage tissue [2,18] were administered systemic pentavalent antimonial therapy for 20 days at 10-20mg/kg/day.
Statistical evaluation was performed using SPSS software Version 21 (IBM Inc., Chicago, Il, USA) and p-values ≤ 0.05 were considered to be significant. Median and interquartile range was used to report clinical characteristics and frequency tables were used to express epidemiological findings. Lesion diameter, duration of disease, and number of lesions data was not normally distributed (Kolmogorov Smirnov test). Therefore, the Kruskal-Wallis test (corrected for ties) for nonparametric data was used to analyze overall differences with respect to age, gender, lesion localization, lesion type, and treatment. When significant differences were found, Dunn-Bonferroni test was used to make pairwise comparisons [19]. Kendall’s tau b correlation was used to discriminate significant associations between sequential treatment with IL and lesion size or number of lesions. Independent t-test was used to obtain the differences between the groups according to duration, number of lesion and lesion size.
The study comprised a total of 8786 pediatric patients of which 4050 (46·1%) were males and 4736 (53·90%) were females, with a mean age of 7·52 years (range 0–15 years). The majority of the PCL patients consulted a doctor within 0–6 weeks of the onset of disease. Mean lesion diameter was determined as 12·77±0·11 mm (range 1–100 mm), while the average lesion duration was 8·58±0·21 weeks (range, 1–815 weeks), although 66·03% of the pediatric patients resolved the infection within 6 weeks. The number of lesions was 1·76±0·01 mm (range 1–37 mm), with 91·20% of the patients showing 1–3 lesions, while 0.83% presented more than 8 lesions (Table 1).
To better understand the epidemiological and clinical characteristics of PCL patients in this region, we divided the patients into three age groups: 0–5, 6–10 and 11–15 years old. We found that 6–10 year old patients were the most infected group (Table 1). We investigated whether there were differences in the frequencies of infected males and females among groups. Interestingly the frequency of infected females was higher than males in all age groups. We further observed that 6–10 year-old patients resolved their infection faster and displayed smaller lesion sizes compared with the other groups. The 0–5 year old group showed the longest duration of the disease. No difference in lesion size was found between 0–5 and 11–15 year old groups. The mean of number of lesions was also similar among these 2 age groups (Table 2).
The area most involved in the lesions of PCL patients was the head and neck region at a rate of 47·8% (n = 4196). The second most involved area was body at a rate of 30·1% (n = 2648). Other areas of involvement were the lower extremity at 0·4% (n = 38), upper extremity at 7·3% (n = 642), mucosa at 3·6% (n = 315) and generalized at 9·9% (n = 874) (involvement of 2 or more areas) (Fig 1A). Also lesion size was different accordingly to the location being upper and lower extremities the sties with larger lesions (Fig 1B). Interestingly, the 0–5 year old group showed more lesions in the head/neck and mucosal regions compared with other groups. Body lesions were most common in the 6–10 year old group. Finally, 11–15 year old group presented less head/neck and mucosal lesions compared with other groups, although this group presented more lesions in the upper extremity and generalized locations (Fig 1C). The lesion size was different between groups in head and neck, lower extremities and generalized (Fig 1D–1F). The 0–5 year old and 11–15 year old groups presented larger lesions in head and neck. In lower extremities 0-5y children presented larger lesions and generalized lesions were larger in both 0-5y and 11-15y groups. The remaining sites of infection were not different in size among groups.
The most common clinical form of lesions determined in PCL patients was ulcer at 57·14% (n = 4983), followed by nodules at 36·62% (n = 3193), papules at 3·93% (n = 343) and recidivans at 2·29% (n = 200) (Fig 2A). The larger lesions where ulcers and recidivans whereas the smallest lesions were papules (Fig 2B). As expected recidivans were the lesions having the longest duration. However, no significant difference in duration was observed between ulcer nodule and papule (Fig 2C). Ulcers were most common in the 11-15-year-old group, although they presented less nodular lesions. The frequency of ulcers and nodules was similar between 0-5-year-old and 6-10-year-old groups. However, the 0-5-year-old group presented less papules. Recidivans were more common in 0-5-year-old and 11-15-year-old groups. (Fig 2D).
During this study there were two types of treatment for cutaneous leishmaniasis. Systemic pentavalent antimonial therapy (intramuscular) was applied to 856 PCL patients (9·74% of the total PCL patients), while intralesional pentavalent antimonial therapy was applied to 6848 PCL patients (77·94% of the total). One thousand and eighty-two patients (12·32% of the total) refused any form of treatment.
Patients that did not heal with the intralesional treatment in the first cycle often received a second or even third cycle of intralesional treatment. Of the PCL patients who received intralesional pentavalent antimonial therapy, 78·79% (n = 5393) received only one cycle (total of 8 injections) of treatment. This left 21·2% of the patients (n = 1455) who required a second cycle of intralesional pentavalent antimonial injection. From the second cycle, only 3·32% of patients (n = 227) received a third cycle (Fig 2A). Patients that received intramuscular treatment presented larger lesions compared to patients that received intralesional treatment or those that were untreated (Fig 3B). Interestingly, when analyzed by groups, the 0-5-year-old group was the least frequent group in which the first intralesional treatment cycle was administered. However, this group also received intramuscular treatment most frequently. Also the 0-5-year-old group received less frequently second and third intralesional treatment cycle. Conversely, while intralesional treatment was more common (93·53%) in the 11-15-year-old group, this group presented the highest frequencies of first cycle (32%) and second cycle (8%) intralesional treatment non-responders. Finally, intralesional treatment was administered in 77% of the patients in the 6-10-year-old group. Interestingly, in this age group, most of the first cycle intralesional treatment non-responders resolved their lesions after the second treatment cycle, while the remaining 1% still required a third treatment cycle (Fig 2C and 2D).
The last few years have reported increasing cases of CL in Turkey [7]. We determined whether pediatric patients also showed increased numbers of CL cases (Fig 4). Of 8786 pediatric patients, 3098 (35.26%), 3464 (39.43%), and 224 (25.31%) were 0–5, 6-10-11-15 age groups, respectively. We found that total CL cases in pediatric patients peaked in 1998, 2002–2005 and 2010–2013. In all the years considered in this report, the frequency of 6–10 years old patients was higher than the other groups.
Old World CL is seen widespread throughout several countries of the Mediterranean region, including Turkey. The vast majority of CL patients in Turkey have been determined to be in the south-east of the country, particularly in the province of Şanlıurfa (13). In this study we show that CL disease has increased in pediatric patients in the last four years. However, as shown by our current study, the highest incidence was seen in the 6–10 year old age group, and when compared to children, the incidence of CL is very low in the elderly [4,6]. The most likely reason for the greater incidence in children is that children are exposed to the parasite at an early age, and unlike adults, children have not been exposed to leishmaniasis and hence lack immunity to CL.
Diagnosis is generally more difficult in PCL patients than in adults. As PCL patients are often diagnosed with impetigo, prurigo or folliculitis and may receive unnecessary treatment, the clinical form of the disease could become completely changed [20]. Particularly in endemic regions, PCL can be easily diagnosed with clinical examination and a simple smear test. Physicians working in endemic regions in particular must be alert to this disease, especially due to the rising number of CL cases in the last three years in Sanliurfa.
Attempts at different paramedical treatments before the patient consults a doctor may affect the severity and duration of the disease, diameter of the lesions and changes in the characteristics of the lesions. In the current study, the diameter of the lesions in the PCL patients was determined as 12.77±0.11 mm and the number of lesions as 1.76±0.01. The mean diameter of the lesions in the current study was in parallel with the mean lesion diameter and duration of disease reported in PCL patients in other countries [1,10,21–23].
Incorrect diagnosis and treatment of PCL disease could also result in an atypical appearance of lesions, affecting disease duration and resulting in unwanted permanent scars during the healing process. In a study of 160 PCL patients in Tunisia, mean duration of the disease was reported as 12·71 weeks [1]. In the current study, the mean duration of the disease was 8·58 weeks. However, when classified according to age groups, 6–10 year old patients resolved the infection faster, even though this group was more often infected with CL. In contrast 0–5 year old patients took the longest time to resolve the infection. This could be associated with the fact that this group presented recidivans form more often. Also, younger children lack a more mature immune system relative to older children. In CL, lesions are more often seen on uncovered parts of the body as these are the areas that are accessible to the sand-fly. In our current study, a higher rate of PCL lesions was determined on open areas of the body such as the head and neck and the body. These results on areas of the body involved in PCL disease are supported by previous clinical studies [1,10,21–24]. Interestingly, the localization of the lesions varied among groups. In the 0–5 year old group, lesions were more common in the head/neck and mucosal regions. Generalized and upper extremity lesions were more common in 11–15 year old group.
Lesions caused by CL generally start as a small papule or may occur in a papulo-nodular form. When treatment is not applied, lesions grow in diameter and become ulcerated. Sometimes yellowish-red papulo-nodules may be observed on a healed CL scar or around it, which is known as the recidivans form. Patients who do not reside in an endemic region for CL disease may find it extremely difficult to diagnose the recidivans form. In the current study, the most common lesion forms were ulcers and nodules (57·14% and 36·62% respectively). Recidivans form was only determined in 2·29% of the PCL patients. In a study by Talari et. al., the ulcerated form of lesions was determined in 60·7% of PCL patients and the nodular form in 27·4% [4]. Interestingly, our study showed that recidivans was more common in 0–5 year old group (4.34%) followed by 11–15 year old group (2.31%). These data suggest that a follow up would be necessary for CL patients from 0–5 years old especially for the occurrence of recidivans.
Pentavalent antimonials are used in Turkey for both pediatric and adult CL patients as an effective and safe treatment method and this is often applied in the intralesional form [25]. Interestingly intramuscular injections of pentavalent antimonial were associated with larger lesion size. On the other hand, 78·79% of the PCL patients received only one cycle of intralesional treatment, suggesting that most of the patients responded to the treatment. This is in line with the report by Zaraa et. al. where they found 86·25% efficacy of intralesional antimonial treatment in PCL patients [1]. We cannot discard however that some patients may have not returned to receive a second cycle or third cycle. In other studies, intralesional antimonial therapy has been reported to be effective in PCL patients [10,22,24–26]. When intralesional antimonial therapy is applied correctly and regularly, it remains the gold standard of treatment for PCL patients. In our study we found that in the 0-5-year-old group, intralesional therapy was applied only to 66% of the patients, while 16% received intramuscular therapy. This could be explained by the fact that IL lesions require repeated administration of the drug, which is painful and may result in poor patient compliance. For pediatric patients, the parents and/or doctors may prefer a single intramuscular dose (16·72%), while others may refuse treatment (16%). However, systemic administration of antimonials is known to be toxic, [26] and it is not known if it could affect childhood development. Interestingly, the 0–5 year old group most frequently received one intralesional cycle, probably because they responded more efficiently than the other groups to intralesional treatment, with no need for a second cycle. Therefore, intralesional treatment of 0–5 year old patients may be considered as a preferential form of treatment for CL, so as to avoid the toxicity associated with systemic administration of antimonials. Most of the children in the 11–15 year old group received intralesional treatment, although there were more non-responders in this group who needed a second cycle (34·4%) and a third cycle (8·9%) of treatment. Some pediatric patients between 6–10 years old needed a second cycle (26·79%) but only very few needed a third cycle (1·37%) of treatment. These data should help in the determination and prognosis of the more appropriate form of treatment for CL depending on the age of the children.
In conclusion, CL affects mainly children and is more common in the 6–10 year old range. Interestingly, the categorization of pediatric patients revealed different clinico-epidemiological characteristics of CL including duration of disease, size of lesion, and lesion localization depending on the age of the children. Importantly response to antimonial therapy appears to be dependent on age. The results from this study can help to improve clinical care and treatment of pediatric CL patients in Turkey.
|
10.1371/journal.pntd.0002976 | Identification of sVSG117 as an Immunodiagnostic Antigen and Evaluation of a Dual-Antigen Lateral Flow Test for the Diagnosis of Human African Trypanosomiasis | The diagnosis of human African trypanosomiasis (HAT) caused by Trypanosoma brucei gambiense relies mainly on the Card Agglutination Test for Trypanosomiasis (CATT). There is no immunodiagnostic for HAT caused by T. b. rhodesiense. Our principle aim was to develop a prototype lateral flow test that might be an improvement on CATT.
Pools of infection and control sera were screened against four different soluble form variant surface glycoproteins (sVSGs) by ELISA and one, sVSG117, showed particularly strong immunoreactivity to pooled infection sera. Using individual sera, sVSG117 was shown to be able to discriminate between T. b. gambiense infection and control sera by both ELISA and lateral flow test. The sVSG117 antigen was subsequently used with a previously described recombinant diagnostic antigen, rISG65, to create a dual-antigen lateral flow test prototype. The latter was used blind in a virtual field trial of 431 randomized infection and control sera from the WHO HAT Specimen Biobank.
In the virtual field trial, using two positive antigen bands as the criterion for infection, the sVSG117 and rISG65 dual-antigen lateral flow test prototype showed a sensitivity of 97.3% (95% CI: 93.3 to 99.2) and a specificity of 83.3% (95% CI: 76.4 to 88.9) for the detection of T. b. gambiense infections. The device was not as good for detecting T. b. rhodesiense infections using two positive antigen bands as the criterion for infection, with a sensitivity of 58.9% (95% CI: 44.9 to 71.9) and specificity of 97.3% (95% CI: 90.7 to 99.7). However, using one or both positive antigen band(s) as the criterion for T. b. rhodesiense infection improved the sensitivity to 83.9% (95% CI: 71.7 to 92.4) with a specificity of 85.3% (95% CI: 75.3 to 92.4). These results encourage further development of the dual-antigen device for clinical use.
| Human African Trypanosomiasis (HAT) is caused by infection with Trypanosoma brucei gambiense or T. b. rhodesiense. The diagnosis of T. b. gambiense infections currently relies primarily on a Card Agglutination Test for Trypanosomiasis (CATT), which has acknowledged limitations, and there is no simple test for T. b. rhodesiense infection. Our overall aim is to produce a simple lateral flow test device with a similar or better sensitivity and specificity than CATT but with better stability and ease of use at point of care. In this study, we identified a particular variant surface glycoprotein, sVSG117, with good diagnostic potential and combined it with a previously identified recombinant diagnostic antigen, rISG65, to produce a prototype dual-antigen lateral flow test. We performed a virtual field trial by testing the device blind with 431 randomized serum samples provided by the WHO HAT Specimen Biobank. The results show that, although the prototype lateral flow test is un-optimized, it was able to diagnose T. b. gambiense HAT with a sensitivity and specificity of 97.3% and 83.3% and T. b. rhodesiense HAT with a sensitivity and specificity of 83.9% and 85.3%.
| Human African Trypanosomiasis (HAT), or African Sleeping Sickness, is caused by two sub-species of Trypanosoma brucei. T. b. gambiense accounts for approximately 95% of HAT infections and occurs across East and Central sub-Saharan Africa. The remaining infections are caused by T. b. rhodesiense in West and Southern Africa. The disease has two stages: Stage 1, where the parasites are limited to the bloodstream, interstitial fluids and lymph of the patient, and stage 2, where parasites are also found in the central nervous system.
In recent years, the official number of recorded HAT cases has fallen below 10,000 per year, although possible under-reporting suggests that this is likely a minimum figure [1]–[3]. Nevertheless, with new therapeutic regimes [4]–[6] and with a repurposed drug (fexinidazole) [7] and a new chemical entity (an oxaborole) [8] in clinical trials, the potential to eliminate HAT from many regions of sub-Saharan Africa at last exists. However, disease elimination requires excellent and convenient field diagnostics. Currently, the diagnosis of infected individuals relies principally on screening teams that visit at-risk communities and from patients seeking medical help [9], [10]. Some patients with T. b gambiense infections remain asymptomatic for years, so early diagnosis of infected individuals benefits not only the patient but also the community where these individuals can act as parasite reservoirs [11]. The most widely used diagnostic for suspected T. b. gambiense infections is the Card Agglutination Test for Trypanosomiasis (CATT). This serological test detects host antibodies to a suspension of fixed and stained T. b. gambiense trypanosomes expressing variant surface glycoprotein (VSG) variant LiTaT1.3 [12]. Over the years, the CATT screening tool has been optimised to improve stability, sensitivity (ranging from 87% to 98%) and specificity, (95%) and thermostability [1], [13]–[16]. A positive CATT is followed up by microscopic examination of blood buffy coat smears. Until recently, stage 1 and stage 2 treatment regimes were different, and the latter much more toxic, such that positive diagnosis of infection was then staged by microscopic examination of Cerebral Spinal Fluid (CSF) for the presence of trypanosomes and/or lymphocytes. However, the use of nifurtimox and eflornithine combination therapy (NECT) [4]–[6] in recent years has largely removed the need for staging diagnosis in T. b gambiese infections.
Despite its usefulness, the CATT screening tool has several widely acknowledged limitations [17]–[20]. It requires cultivation of infectious parasites for its manufacture, trained personnel for use and the read out is subjective, causing variability in reported sensitivity and specificity [16], [21]. Significantly, some T. b. gambiense strains do not express the LiTat1.3 VSG gene and, therefore, patients infected with these strains do not generate detectable antibodies [22]. For the same reason, the CATT test cannot detect T. b. rhodesiense infections [23].
There are challenges to developing improved diagnostic assays and devices for HAT. Due to the very low parasite levels in patients infected with T. b. gambiense, a test that detects host antibodies (rather than parasite antigens) is considered more likely to have the necessary sensitivity. The WHO recommends that point of care tests (POCT) should follow the ‘ASSURED’ criteria; which states that a POCT device should be affordable, sensitive, specific, user-friendly, rapid, equipment-free and deliverable to the people at need. Lateral flow tests (LFTs) are inexpensive and simple devices that can rapidly detect nanogram amounts of antibodies in finger-prick blood samples without the need for any ancillary equipment [24]. A first-generation LFT for T. b. gambiense infections, that uses two different purified native VSG antigen bands (LiTat1.3 and LiTat1.5) to detect anti-VSG antibodies, has recently entered clinical use as CATT replacement [25]. We have also produced a promising prototype LFT using a recombinant invariant surface glycoprotein (rISG65) antigen [26]. In this paper, we identify another soluble form VSG (sVSG117 also known as sVSG MITat1.4) with excellent diagnostic properties that we have used together with rISG65 to create a prototype dual-antigen LFT that detects T. b. gambiense infections and, to some extent, T. b. rhodesiense infections.
The serum samples used in this study were from the WHO HAT Specimen Biobank, archived at the Pasteur Institute, Paris. Patients were recruited by WHO to provide serum samples as described in [27] for the development of new diagnostic tests for HAT and patient consent was collected by WHO at the time of sample collection. Further local ethical approval for this study was granted by the Tayside Ethics Review Board. Rodents were used to propagate T. brucei parasites for the purification of soluble form variant surface glycoproteins (sVSGs). The animal procedures were carried out according the United Kingdom Animals (Scientific Procedures) Act 1986 and according to specific protocols approved by The University of Dundee Ethics Committee and as defined and approved in the UK Home Office Project License PPL 60/3836 held by Michael A.J. Ferguson.
All patients were tested with the CATT test (which was followed by parasitological analysis) and examined for clinical symptoms of HAT [27]. Serum samples were stored in the WHO HAT Specimen Biobank at −80°C and shipped to Dundee on dry ice where they were thawed, divided into aliquots and stored at −20°C.
Bloodstream form T. b. brucei Lister strain 427 clones expressing four different VSG variants (117, 118, 121 and 221) were cultivated in rodents as described in [28] and sVSGs were purified by a simplified version of the method of Cross [28], as described in [29]. The sVSGs were further purified by gel-filtration using a Sephacryl-S200 column (4×90 cm) equilibrated and eluted with 0.1 M NH4HCO3. The gel-filtration purified sVSGs were lyophilised to remove NH4HCO3 and stored as dry powders at 4°C before use. Samples were run on an SDS-PAGE gel to check for purity and were considered >95% pure (data not shown).
The ELISA plate preparation details and protocols were as described in [26]. ELISAs were carried out on both pooled and individual serum samples. The pooled sera were from stage 1 T. b. gambiense patients (n = 10), stage 2 T. b. gambiense patients (n = 40) and matched uninfected patient sera (n = 50). Pooled sera were diluted 1∶1000 in in phosphate buffered saline containing 0.1% w/v bovine serum albumin (PBS/BSA) and plated in triplicate in serial (doubling) dilutions in PBS/BSA to 1∶32000. Individual sera were diluted to 1∶1000 in PBS/BSA and applied to ELISA plates in triplicate.
For the sVSG117 single antigen lateral flow test pilot study, forty T. b. gambiense infection sera and forty matched uninfected control sera were randomised and coded by a member of the University of Dundee Tissue Bank. For the dual-antigen lateral flow test virtual field trial, 431 serum samples, representing a mixture of T. b. gambiense (n = 150) and T. b. rhodesiense (n = 56) infection sera and matched uninfected control sera (n = 150 for T. b. gambiense and n = 75 for T. b. rhodesiense) were randomised and coded by the WHO HAT specimen Biobank.
We supplied BBI Solutions with 5 mg sVSG117 to make single antigen sVSG117 LFT prototype devices for preliminary studies and with a further 7 mg of sVSG117 and 7 mg of rISG65 [26] to make dual-antigen LFT prototypes. BBI Solutions is an inmmunoassay development and manufacturing company that has completed more than 250 lateral flow projects over the last 25 years, with manufacturing sites in Europe, USA and South Africa. Both serum- and blood-accepting pad devices were made. For LFTs without blood pads, aliquots of 5 µl of patient sera diluted with 15 µl of PBS were added to the LFTs followed by an 80 µl of chase-buffer (PBS containing 0.05% Tween 20). For LFTs with blood pads, aliquots of 5 µl of patient serum were mixed with 5 µl PBS and 10 µl of freshly reconstituted human type-O blood erythrocytes. These mixtures were added to the LFTs, followed by 80 µl of chase-buffer (PBS containing 0.05% Tween 20). Tests were discarded if upper control line was not clearly visible. After 30 min, scoring of the test bands was performed by visual comparison of freshly completed tests with a scoring card. For the virtual field trial, two people scored all of the LFT devices independently. If there was disagreement about the infection-status of a given serum sample, a third individual provided adjudication.
Line graphs were generated by Microsoft Excel. Receiver Operator Characteristic (ROC) curves, antigen scatter plots and tables of sensitivity and specificity scores were generated by SigmaPlot 12.
Our original rationale for testing HAT sera against a panel of different sVSGs was to look for the presence of anti-Cross Reacting Determinant (CRD) IgG antibodies. The CRD is a peptide-independent epitope common to all sVSGs that is created upon the cleavage of VSG glycosylphosphatidylinositol (GPI) membrane anchors by GPI-specific phospholipase C (GPI-PLC) during cell lysis [30]. However, the ELISA data showed that while there was anti-peptide and/or anti-CRD IgG antibody titre to all four sVSGs, the immunoreactivity of both stage 1 and stage 2 T. b. gambiense HAT patient sera to sVSG117 was far higher than to the other three (Figure 1). From this result, we decided to pursue sVSG117 as a potential diagnostic antigen in its own right. We therefore proceeded to screen randomised and coded sera from 40 T. b. gambiense infected patients and 40 matched uninfected control patients against sVSG117 coated ELISA plates (Figure 2A). These data strongly suggested that immunoreactivity to sVSG117 might be used to reliably discriminate infection from control sera. Consequently, sVSG117 was developed into an un-optimised single-antigen prototype lateral flow test (Figure 3A), which was used with the same set of 80 randomised and coded serum samples. The visual test scores of the decoded data are shown in (Figure 2B). The bands were also assessed by quantitative laser densitometry, as described in [26], (Figure 2C) which showed an excellent correlation between visual- and densitometer-based scoring, with an r2 correlation value of 0.957. These data enabled us to set a cut-off threshold of ≥1 visual units for discriminating infected from uninfected sera on this LFT device. Using this threshold, the test appeared to have 100% sensitivity and 100% sensitivity, albeit based on a relatively small sample set.
An un-optimised dual-antigen lateral flow test prototype, containing one band of recombinant antigen rISG65-1, previously described in [26], and one band of the native sVSG117 antigen, described here, was manufactured by BBI Solutions (Figure 3B). The dual-antigen LFTs were manufactured using the same antigen coupling conditions as the individual rISG65 [26] and sVSG117 (this study) single-antigen LFTs. Thus, visual score cut-offs of ≥2 for the rISG65 band [26] and of ≥1 visual units for the sVSG117 band were expected to define positive immunoreactivity to each antigen, respectively. However, to establish visual cut off values directly for this new LFT, the same 80 randomised serum samples described above were tested blind with the dual-antigen LFT and scored. After decoding, cut-offs were confirmed as being ≥2 and ≥1 for the rISG65 and sVSG117 test lines, respectively. Using these values, and the criterion of two positive test lines to define an infection, the device gave 100% sensitivity and 97.5% specificity in this pilot study with a limited number of serum samples (n = 80).
A virtual field study was performed to assess the diagnostic potential of the dual-antigen LFT. First, aliquots of 431 randomized and coded serum samples, provided by the WHO HAT Specimen Biobank, were mixed with an aliquot of human type-O erythrocytes, provided by the Tayside blood-bank, to produce pseudo blood samples containing red blood cells as well as serum antibodies. These samples were added to the LFTs fitted with blood pads, followed by chase buffer, and read independently by two individuals after 30 min. The LFT was deemed to be positive if the rISG65 band and sVSG117 had mean visual scores of ≥2 and ≥1, respectively. After decoding by colleagues at the WHO HAT Specimen Biobank, we were able to plot ROC curves (Figure 4) and separately assess the sensitivity and specificity of the LFT to detect T. b. gambiense and T. b. rhodesiense infections using the following criteria: (i) two positive antigen bands = infection, (ii) a positive sVSG117 band = infection, (iii) a positive rISG65 band = infection and (iv) any one positive antigen band = infection. The results, in terms of sensitivity, specificity and the respective 95% confidence intervals (CI) are summarised in (Table 1).
Although we selected sVSG117 as a potential diagnostic antigen from empirical data in this study, our results are also consistent with population genetics studies that show that the gene encoding this VSG variant (the same as VSG AnTat 1.8) is ubiquitous in T. b. gambiense isolates [31], [32], whereas those for VSGs 121 and 221 are not [33].
In the virtual field trial, using two positive antigen bands as the criterion for infection, the sVSG117 and rISG65 dual-antigen lateral flow test prototype showed a sensitivity of 97.3% (95% CI: 93.3 to 99.2) and a specificity of 83.3% (95% CI: 76.4 to 88.9) for the detection of T. b. gambiense infections. The sensitivity is comparable to those reported (87–100%) for the currently deployed CATT test and for the latex agglutination test which uses diluted blood and latex beads coated with three different VSG variants (LiTat1.3, 1.5 and 1.6) [17], but is poorer with respect to specificity, which have been reported as 85–97% for CATT and 96–99% for the latex test [13]–[17]. Nevertheless, the dual-antigen LFT described here is only a prototype that needs to undergo extensive optimization with respect to antigen-gold coupling, antigen loading of the test lines and composition of the chase buffer. We therefore suggest that one or both of the sVSG117 and rISG65 antigens be seriously considered for use in the next generation of clinical LFT devices for the diagnosis of T. b. gambiense HAT.
We note that for our prototype dual-antigen LFT the specificity performance of each individual antigen is relatively poor for detecting T. b. gambiense infections (Table 1). For example, the rISG65 test line shows a sensitivity of 98.0% (95% CI: 94.3 to 99.6) but a specificity of only 65.3% (95% CI: 57.1 to 72.9). We have previously reported the sensitivity and specificity performance of a visually-read single-antigen LFT using rISG65 as 88% (95% CI: 73 to 96) and 93% (95% CI: 80 to 98), respectively [26]. While there is good overlap between the 95% confidence intervals for these two assessments with respect to sensitivity, we note that there is a discrepancy with respect to specificity. However, the previous assessment [26] only used 80 randomized infection and control serum samples and we suggest that the figures reported in (Table 1) are likely to be more accurate given the significantly greater sample size (and wider geographic sampling) used in the virtual field trial. Another possibility is that some of the ‘false-positive’ results, which drive down the specificity figures for the dual-antigen LFT, might be due to asymptomatic true positives that had been missed by the CATT test in the virtual field trial cohort. As previously noted, this is entirely possible as not all T. b. gambiense strains express the LiTat1.3 VSG upon which the CATT test is based [22].
The dual-antigen LFT did not perform as well for detecting T. b. rhodesiense infections using two positive antigen bands as the criterion for infection, with a sensitivity of only 58.9% (95% CI: 44.9 to 71.9) and specificity of 97.3% (95% CI: 90.7 to 99.7). A potentially confounding issue for T. b. rhodesiense immunodiagnosis is the typically acute nature of these infections compared to typically chronic T. b gambiense infections, with the latter more likely to produce robust antibody responses to parasite antigens. However, using any one (or both) positive antigen band(s) as the criterion for T. b. rhodesiense infection improved the sensitivity to 83.9% (95% CI: 71.7 to 92.4) with a specificity of 85.3% (95% CI: 75.3 to 92.4). As of yet there have been no confirmed cases of co-existing T. b. rhodesiense and T. b. gambiense infections [34], and given the current lack of immunodiagnostics for T. b. rhodesiense infections [35], an optimized version of the dual-band LFT using the relaxed criteria of one or two positive band(s) to diagnose HAT might be clinically useful in T. b rhodesiense endemic regions.
Taken together, the results described in this paper encourage further development of the dual-antigen LFT device described here (or one or both of its antigens, i.e., recombinant rISG65-1 and native sVSG117) for clinical use for the detection of T b. gambiense infections and, possibly, for T. b rhodesiense infections. LFT technology offers advantages over CATT with respect to the “affordable, user-friendly, rapid, equipment-free and deliverable to the people at need” components of the WHO ‘ASSURED’ criteria of “affordable, sensitive, specific, user-friendly, rapid, equipment-free and deliverable to the people at need”. However, the “sensitive” and “specific” components of the criteria are clearly also key to success and, while data on the currently deployed first-generation LFT [25] (that uses native sVSGs LiTat1.3 and LiTata1.5) are yet to be published, the FIND web site suggests that its performance is comparable to CATT. Like the CATT test, and the currently deployed LFT [25], our dual-antigen LFT requires the cultivation of parasites to make the native sVSG117 component, although sVSG117 can at least be prepared from non-human infectious T. b. brucei. Nevertheless, the ideal second-generation LFT is likely to use two recombinant, rather than native, antigens and recombinant ISG65 [26] and/or VSG domains could be the answer.
|
10.1371/journal.pgen.1002815 | RsfA (YbeB) Proteins Are Conserved Ribosomal Silencing Factors | The YbeB (DUF143) family of uncharacterized proteins is encoded by almost all bacterial and eukaryotic genomes but not archaea. While they have been shown to be associated with ribosomes, their molecular function remains unclear. Here we show that YbeB is a ribosomal silencing factor (RsfA) in the stationary growth phase and during the transition from rich to poor media. A knock-out of the rsfA gene shows two strong phenotypes: (i) the viability of the mutant cells are sharply impaired during stationary phase (as shown by viability competition assays), and (ii) during transition from rich to poor media the mutant cells adapt slowly and show a growth block of more than 10 hours (as shown by growth competition assays). RsfA silences translation by binding to the L14 protein of the large ribosomal subunit and, as a consequence, impairs subunit joining (as shown by molecular modeling, reporter gene analysis, in vitro translation assays, and sucrose gradient analysis). This particular interaction is conserved in all species tested, including Escherichia coli, Treponema pallidum, Streptococcus pneumoniae, Synechocystis PCC 6803, as well as human mitochondria and maize chloroplasts (as demonstrated by yeast two-hybrid tests, pull-downs, and mutagenesis). RsfA is unrelated to the eukaryotic ribosomal anti-association/60S-assembly factor eIF6, which also binds to L14, and is the first such factor in bacteria and organelles. RsfA helps cells to adapt to slow-growth/stationary phase conditions by down-regulating protein synthesis, one of the most energy-consuming processes in both bacterial and eukaryotic cells.
| The YbeB/DUF143 family of proteins is one of the most widely conserved proteins with homologues present in almost all bacteria and eukaryotic organelles such as mitochondria and chloroplasts (but not archaea). While it has been shown that these proteins associate with ribosomes, their molecular function remained mysterious. Here we show that a knock-out of the ybeB gene causes a dramatic adaptation block during a shift from rich to poor media and seriously deteriorates the viability during stationary phase. YbeB of six different species binds to ribosomal protein L14. This interaction blocks the association of the two ribosomal subunits and, as a consequence, translation. YbeB is thus renamed “RsfA” (ribosomal silencing factor A). RsfA inhibits translation when nutrients are depleted (or when cells are in stationary phase), which helps the cell to save energy and nutrients, a critical function for all cells that are regularly struggling with limited resources.
| Escherichia coli harbors a core set of about 190 genes that are conserved in more than 90% of all completely sequenced genomes [1]. Most of them encode well-understood proteins involved in metabolism, transcription, translation, or replication. However, a few of these highly conserved proteins remain functionally uncharacterized and thus enigmatic. One of these mysterious proteins is YbeB. In 2004 it was proposed by Galperin and Koonin as one of 10 top targets of conserved hypothetical proteins for experimental characterization [2]. In recent interactome studies, we and others found this protein to interact with various proteins, including several ribosomal components [3], [4], [5], [6]. Moreover, YbeB was shown to co-sediment with the large ribosomal subunit (LRS) [7], suggesting that it functions in protein translation. Recently it has been suggested that its mitochondrial homologue, C7orf30, is involved in ribosome biogenesis and/or translation [5], [8] although these studies were not able to explain their observations mechanistically. In this work we characterize YbeB's molecular function by identifying its binding site on the LRS and reveal a molecular mechanism of YbeB action: it is down-regulating protein synthesis under nutrient shortage by binding to protein L14 of the LRS, acting as a ribosomal silencing factor (“RsfA”) by blocking ribosome subunit joining. Thus, we will use the term “RsfA” below.
In the Pfam database (V26.0) RsfA sequence homologues are known for at least 2,928 species, including nearly all bacteria as well as almost all eukaryotic species (Pfam entry PF02410, Interpro IPR004394). However, the RsfA protein family is conspicuously absent in archaea (Figure 1A). In the STRING 9.0 database [9] RsfA is clustered with the orthologous protein group “COG0799”, consisting of 932 RsfA homologues in 920 different species, indicating that there is usually one rsfA gene per genome. A multiple sequence alignment of ten representative RsfA orthologues, however, exhibits only limited conservation when compared to ribosomal protein L14 (Figure S1).
Interestingly, more than 80% of all eukaryotic RsfA orthologues are predicted to localize to mitochondria or chloroplasts according to the WoLF PSort program [10]. For the yeast orthologue ATP25, the mitochondrial localization has been experimentally confirmed [11] and the Zea mays homologue, Iojap, was found in chloroplast fractions [12]. This strongly suggests that RsfA functions in a strictly conserved process of bacterial origin. Previously, Butland and colleagues reported L14, L19, L4, L7/L12 and others as interaction partners of RsfA based on protein complex data [3]. Similarly, we found that several interactors of RsfA's Treponema pallidum orthologue TP0738 were involved in protein synthesis [6]. Although these observations provided the first experimental hint that RsfA might function in translation, this has never been functionally demonstrated. Since previous studies have revealed RsfA's association with the large ribosomal subunit (LRS) which offers multiple binding sites, we re-tested all previously detected interactions of T. pallidum RsfA that are involved in protein translation. As expected, several proteins indeed tested positive (Figure 1B). However, the interaction of RsfA with L14 was by far the strongest as determined by using increasing concentrations of 3-amino-triazole (3-AT), a competitive inhibitor of the yeast two-hybrid reporter gene HIS3. In fact, only the interaction with L14 was detectable at more than 1 mM 3-AT. Furthermore, the L14-RsfA interaction was the only one that was detectable in a reciprocal screen, i.e. with RsfA used as both bait and prey.
Given the conservation of RsfA, we wanted to establish to which extent the interactions of RsfA of T. pallidum are conserved in other species. To this end, we first retested whether the interactions of T. pallidum RsfA are conserved in E. coli. We also included eight putative interaction partners that have been identified in a protein complex together with E. coli RsfA and L14 [3] and four interologous pairs detected by Y2H in Campylobacter jejuni [13]. Surprisingly, only the interaction with L14 was conserved in E. coli as a strong (up to 50 mM 3-AT) and reciprocal interaction (Figure 1C, all tested interactions and reference sets are listed in Table S2 and the complete Y2H assays are shown in Figure S2). Moreover, we confirm the interaction of RsfA with L14 from E. coli independently in a pull-down experiment (Figure S3A).
Thus, we conclude that L14 is the primary and specific binding target of RsfA on the LRS and that all other interactions are species specific or even artifacts.
Next we tested whether this particular interaction is conserved in other bacteria. Notably, we could verify the interaction in all tested species, including gram-positive Streptococcus pneumoniae and the cyanobacterium Synechocystis PCC 6803 (Figure S3B and S3C). In addition, we confirmed the interaction between the corresponding orthologues of RsfA/L14 of both human (C7orf30/mitochondrial L14) and Zea mays (Iojap/chloroplastic RPL14) as shown in Figure 1D and 1E, respectively.
In HeLa cells human C7orf30 co-localized with L14mt exclusively to mitochondria (Figure 1F). This supports the hypothesis that eukaryotic RsfA orthologues are functionally active only in organelles. Finally, we verified the human protein interaction in vivo by a bimolecular fluorescence complementation assay using C-terminally tagged Split-Venus constructs (Figure 1G). In summary, these results strongly suggest that the interaction of RsfA and L14 is universally conserved in all species that encode RsfA homologues and that in fact their specific binding site at the LRS is in the ribosomal protein L14.
In order to map the exact binding site of RsfA we used the LRS 3D structure (PDB id: 2AWB [14]: first, we identified amino acids of L14 that (i) are highly conserved (Figure 2A(a) and 2A(b)) and that (ii) are located on the surface exposed towards the 30S small subunit interface. These criteria identified T97, R98, K114, and S117. (Figure 2A(b,c)). In fact, docking a homology model of RsfA and a crystal structure of L14 predicted these residues to be at their interaction interface (Figure 2B). In order to test whether the identified residues of L14 are indeed essential for the L14-RsfA interaction, we substituted T97, R98, K114, and S117 with a single alanine each and tested these L14 constructs if they still bound RsfA by another Y2H experiment (Figure 2C): the K114A and T97A mutants lost the interaction with RsfA already in the presence of 0 to 1 mM 3-AT, while in R98A the interaction was lost at 10 mM and higher concentrations. S117A did not appear to affect the interaction. Several control mutations including moderately conserved amino acids (D80A, F100A, E121A) and none-conserved ones (R49A, K51A) did not show any difference in the Y2H assay compared to the assayed wild type L14 (Figure 2A, 2C).
In summary, the interaction epitope assay confirms that the docking model (Figure 2B) is largely correct. The RsfA-interaction epitope of L14 involves the highly conserved residues K114, T97, and R98 (but not S117) while K114 and T97 are the most critical ones. Notably, T97 and R98 are involved in bridge B8 (Figure 2A(d)) that contacts the small ribosomal subunit [15]. The docking model predicts that binding of RsfA to these residues, as a consequence, would sterically interfere with ribosome subunit joining (Figure 2B(b)) and thus might block translation.
Although RsfA is phylogenetically highly conserved, its gene deletion has been reported not to result in any obvious growth disadvantage in E. coli [7], [16]. We designed a sensitive growth experiment, which compares the WT and the rsfA deletion strain under competitive growth conditions: we mixed equal amounts of both cell types and monitored the populations at constant time intervals under log-phase conditions. Figure 3A demonstrates that the amounts of mutant cells decreased continuously. In other words, WT cells in rich medium steadily overgrew the mutant cells leaving only about 10 to 25% of mutant cells after 35 generations. This modest effect reveals that RsfA mutant cells suffer from a disadvantage when competing with WT cells. Strikingly, a much stronger difference was observed, when cells grown in rich medium were diluted in minimal medium: the WT strain overgrew the mutant ΔrsfA strain within only five generations. The opposite growth transition (poor→rich media) is better tolerated by the mutant strain. The addition of amino acids to the minimal medium completely rescues this striking growth defect of the rsfA mutant in the rich→poor media transition (see Discussion).
These strong defects seen with the ΔrsfA strain in minimal medium rather than in rich medium should be evident also in a direct determination of the doubling times of wild type versus mutant in separate cultures. In rich medium the generation times of WT and mutant strains were not significantly different (30 and 32 min, respectively; Figure 3B). However, a change from rich to poor medium revealed a dramatic difference: initially the ΔrsfA mutant strain showed a growth like the WT strain for about 7 h, but then growth was abrogated for about 14 h before it resumes almost with the same doubling time as the WT strain (130 versus 120 min). The growth block for many hours demonstrates that the lack of the rsfA gene poses a serious adaptation problem on the cells after a transition from rich to poor medium.
It has been reported that the rsfA (formerly ybeB) knock-out can cause a defect in cell separation in a distinct genetic background, and this defect can be complemented with genes of the rsfA operon downstream of the rsfA gene indicating a polarity effect of the rsfA deletion [17]. Therefore, we tested whether we can complement the strong mutant phenotype observed in Figure 3A and 3B by introducing a plasmid carrying the rsfA gene. If so, it would prove that the mutant phenotype is caused by the absence of the RsfA factor. To this end, we removed the kanamycin cassette in place of the chromosomal rsfA gene and introduced a plasmid with the rsfA gene under the native promoter; the expressed RsfA carried a His-tag at the C-terminus to monitor the expression by anti-His antibodies. Figure 3C demonstrates that the mutant phenotype could not be cured probably due to the fact that after the shift to the poor medium RsfA was not sufficiently expressed, whereas taking up growth after 30 h was accompanied by a strong RsfA expression (see red bars in Figure 3C). Therefore, we performed the same experiment but now with the rsfA gene under a tac promoter. The forced RsfA expression could heal the mutant phenotype (Figure 3D; red closed circles). We conclude that (i) the RsfA expression is regulated in a way we do not yet understand, and (ii) that the lack of RsfA is responsible for the mutant phenotype.
Figure 3A and 3B demonstrate that mutant and WT strains showed almost the same growth behavior under log-phase conditions in rich medium (LB). But what happens in a batch culture, when a mixture of both strains reaches the stationary phase in rich medium and protein synthesis has to be down regulated? This was tested in the next experiment. The stationary phase is reached after about 7 h (red line in Figure 3E). At various time points aliquots were taken and the fraction of ΔrsfA mutant strains were determined (blue bars). Until reaching the stationary phase the fraction of mutant cells remains constant at about 35%, but thereafter the fraction of mutant cells sharply declined to less than 10%. This viability competition assay indicates that the mutant cells have serious problems to form stable stationary-phase cells.
The experiments shown in Figure 3A–3E disclose two strong phenotypes caused by the lack of RsfA: (i) The cells adapt poorly after the transition from rich to poor media, and (ii) the viability of cells is dramatically impaired during the stationary phase, eventually causing cell death.
Given RsfA's physical association with the large ribosomal subunit/L14, we wondered whether RsfA has an effect on protein synthesis. To this end we expressed β-galactosidase (as an L-arabinose inducible reporter) in an E. coli gene deletion strain (ΔrsfA) and wild type (WT) cells. At stationary phase the β-galactosidase expression was strongly repressed in wild type cells as expected (Figure 3F). In striking contrast, the ΔrsfA mutant exhibited a significant accumulation of β-galactosidase in the stationary phase. These results demonstrate that RsfA acts as a negative modulator of protein translation in vivo in the stationary phase. Together with the viability assay (Figure 3E) these results suggest that silencing protein synthesis plays an important role for reorganization of the metabolic conversion on the way to the stationary phase.
Next we tested whether RsfA interferes with ribosomal elongation in vitro using a highly resolved E. coli system just containing purified elongation factors EF-Tu, EF-Ts, EF-G, purified precharged [14C]Phe-tRNA, poly(U) programmed ribosomes and GTP as energy source. We added 30S subunits to an excess of 50S subunits in order to facilitate association to 70S ribosomes. Purified RsfA suppressed the translational activity dramatically down to about 20%, when RsfA was added to the 50S subunits before the oligo(Phe) synthesis (Figure 4A, left panel). To test whether RsfA blocks ribosomal activities via interfering with association of the subunits as suggested by our protein docking model (Figure 2B), we subjected an aliquot to a sucrose-gradient analysis before incubating for oligo(Phe) synthesis (Figure 4B). The gradients demonstrate that in the absence of RsfA clearly more 70S ribosomes are formed on the cost of ribosomal subunits. However, when RsfA was added to programed 70S ribosomes carrying an AcPhe-tRNA at the ribosomal P site, no inhibition was observed indicating that RsfA does not interfere with ribosomal functions during the elongation phase (Figure 4A, right panel). We conclude that RsfA blocks association of the ribosomal subunits to functional 70S ribosomes.
Corresponding experiments with the translational elements of mitochondrial ribosomes from mammalian cells (pig liver) confirmed these results. In the presence of purified mitochondrial factors mtEF-Tu, mtEF-Ts, mtEF-G1, poly(U) and [14C]Phe-tRNA oligo(Phe) synthesis was severely reduced upon addition of the mitochondrial RsfA orthologue C7orf30 (mtRsfA; Figure 4C). The results suggest that the function of RsfA is conserved from bacteria to eukaryotic mitochondria.
The cellular synthesis machinery runs at high speed in the exponential (logarithmic) phase of bacterial growth. The growth rate slows in semi-log phase and finally comes to a halt at higher cell density in the stationary phase, usually caused by nutrient depletion. Several bacterial factors bind to ribosomes and thus support the dormant state of the ribosomes in the stationary phase, such as the ribosome modulation factor (RMF), hibernation promoting factor (HPF) or stationary-phase-induced ribosome-associated protein (SRA) [18], [19], [20], [21]. RMF (homologues exist only in the γ-proteobacteria) alone or together with the more broadly distributed HPF are essential for the formation of 70S dimers in the stationary phase, so called 100S particles; an inactivation of the RMF gene causes a viability defect at prolonged periods in stationary phase [22], [23]. Phenotypical effects of knock-out strains concerning the other factors have not been reported.
A first analysis of RsfA-binding partners identified a group of proteins including a number of ribosomal proteins [6]. Similarly, other groups suggested various ribosomal proteins as binding partners [3], [4], [5], the common denominator being that all proteins were derived from the large subunit. Thorough analyses presented here identified the ribosomal protein L14 as the docking station (Figure 1B–1G, Figure 2), and mutation of conserved amino acid residues of L14 at the surface of this protein abolished RsfA binding, clearly demonstrating L14 as the binding protein (Figure 2). Interestingly, the three most conserved residues of RsfA as shown by the multiple sequence alignment (Figure S1A) are located at the interface with L14 predicted by docking. The three residues are W120, D124 and R140 (alignment numbers), corresponding to residue numbers W77, D81 and R95 in E. coli RsfA. D81 is predicted to be in direct contact with R98 of L14 that was shown to disrupt the interaction when mutated. Another such critical residue, K114 of L14, is predicted to be in contact with a fairly conserved residue with RsfA L103 (position 148 in the alignment).
The only other known protein that like RsfA also docks to the ribosomal protein L14 of eukaryotic ribosomes is the so-called initiation factor eIF6, which is not a homologue to RsfA and is thought to block ribosome association in archaea and in eukaryotes from yeast to man [24], [25], [26], [27], [28], [29]. However, in eukaryotes eIF6 is rather a 60S assembly factor and plays an essential role in the late pre-25S rRNA processing and the export of the 60S subunit from the nucleolus to the cytoplasm [30]. Depletion of eIF6 is eventually lethal, in contrast to RsfA. Interestingly, eIF6 is restricted to the eukaryotic nucleus/cytoplasm and to archaea [27], while RsfA is present in almost all bacteria and their descendent eukaryotic organelles (Figure 1A).
Studies with the human mitochondrial homologue of RsfA, C7orf30, have recently suggested that this protein is involved in ribosomal assembly and/or translation [5], [8]. Our results do not indicate any assembly defects as deletion strains of rsfA appear to have perfectly assembled ribosomes (sucrose gradients not shown) and actually translate as well as wild type strains at logarithmic phase (Figure 3F). In addition, we could show that C7orf30 inhibits translation by mitochondrial ribosomes (Figure 4C). It remains possible that C7orf30 has multiple roles in mitochondria or that its role in ribosome assembly is indirect.
In rich medium bacterial cells produce proteins at maximum rates to sustain cell division. Furthermore, bacterial cells take up many metabolic precursors such as amino acids and thus block corresponding synthesis pathways. In contrast, in poor/minimal medium protein synthesis must be down-regulated in a concerted fashion in order to save energy and resources, and at the same time many synthesis pathways such as those for the synthesis of amino acids have to be switched on [31], [32]. The results presented here suggest that RsfA plays a prominent role in this down-regulation by silencing ribosome activities. We observe two strong phenotypes with the ΔrsfA strain: (i) the viability is strongly impaired in the stationary phase (Figure 3E) and (ii) after a transition from rich to poor media the adaptation phase lasts more than 10 hours before resuming growth again in striking contrast to WT cells (Figure 3B), which overgrow the mutant strain in a few generations. Just adding casamino acids to the minimal medium relieves the strong growth defects of the ΔrsfA strain (Figure 3A). Adding amino acids will switch off most of the amino-acid synthesis pathways similar to the situation during the logarithmic phase in the presence of rich medium, when the silencing effect of RsfA is not strictly required. In contrast, during starvation and in the absence of ribosomal silencing (ΔrsfA), energy would be wasted affecting the conversion of the metabolic network, eventually causing deleterious growth defects. Accordingly, protein synthesis is seriously attenuated in the stationary phase, when RsfA is present (i.e. wild type cells) in contrast to protein synthesis in the ΔrsfA strain (Figure 3F). Attenuation of protein synthesis by RsfA seems to be of utmost importance for reorganization the metabolic state on the way to the stationary phase, since the absence of this factor threatens seriously the viability in the stationary phase (Figure 3E), and it explains the well-known effect that ribosomes are much less active, when derived from the stationary rather than from log-phase cells [33].
When RsfA is added to ribosomal subunits it blocks 70S formation and thus protein synthesis (Figure 4A and 4B), whereas the factor does not interfere with the elongation phase of protein synthesis when added to ribosomes that have passed the initiation phase (Figure 4A, right panel). We conclude that RsfA, as a ribosomal silencing factor, is damping the translational activity under restricted energy (stationary phase) or nutrient conditions (growth in poor medium) thus harmonizing translation with the general metabolic state, i.e. RsfA works in line with the stringent response [34] and thus plays a key role in the physiology of the stationary phase and the translational adaptation during the transition from rich to poor medium.
Our experiments suggest a direct silencing effect of RsfA sketched in Figure 5: when the ribosomal activity should be silenced, RsfA binds to the ribosomal protein L14 at the interface of the large subunit and by impairing association of the ribosomal subunits translation is hampered. We demonstrated that RsfA damps the ribosomal elongation in bacterial and mammalian mitochondrial systems (Figure 4A and 4C). The importance of RsfA in eukaryotic organelles is indicated by the fact that a mutation in the gene of the RsfA orthologue Iojap in Zea mays leads to irregular albino patterns on maize leafs and germless seeds due to failure of proplastids to differentiate into chloroplasts [35], [36], [37], [38]. Photosynthesis and respiration can vary enormously in plastids and mitochondria, respectively, and as suggested by the experiment shown in Figure 4C, the RsfA orthologue might accordingly regulate protein synthesis in these organelles using the mechanism suggested here.
ORFs were cloned into pDONR207 by using the Gateway Technology (Invitrogen). Zea mays cDNA was kindly provided by F. Hochholdinger (Tübingen, Germany), HeLa cDNA by O. Kassel (Karlsruhe, Germany), S. pneumoniae TIGR4 DNA by D. Nelson (UMBI, MD, USA), T. pallidum DNA by T. Palzkill (Houston, USA), and Synechocystis PCC 6803 DNA by T. Lamparter (Karlsruhe, Germany). All ORFs were cloned with a stop codon at the 3′-ends. Entry plasmids were sequenced, shuttled into expression vectors (see below), and finally verified by PCR reactions. For the interologous tests E. coli ORFs were kindly provided as pENTR/Zeo clones by S.V. Rajagopala [39] except for RsfA and L14 which have been cloned in this study.
E. coli L14 (b3310) alanine substitutions were directionally introduced by performing standard fusion PCR reactions using mutagenic primers. For cloning PrimeStar HS DNA Polymerase was used (Takara Bio Inc.).
Entry plasmids were recombined with the bait and prey vector pGBKT7g and pGADT7g (Clontech) [40]. These were individually transformed into the haploid yeast strains AH109 and Y187 [41], [42]. After mating the haploids and enrichment of diploids, yeast growth was observed on solid starvation medium lacking Leucine, Tryptophan, and Histidine. The medium contained various concentrations of 3-AT (0 to 100 mM). Detailed procedures were done as described elsewhere [43].
In case of the L14-interaction epitope mapping experiment bait and prey plasmids were sequentially cotransformed into haploid yeast strain CG-1945 (Clontech) and then assayed as described above.
ORFs were shuttled from entry plasmids into pNusA (Santhera, Liestal, Switzerland), pETG-40A, or pETG-30A (EMBL, Heidelberg, Germany) and transformed or co-transformed into E. coli BL21(DE3) (combinations, see main text, Figure 1D and 1E and Figure S3A). Proteins were expressed following standard protocols. Cell pellets were lysed in 500 µl buffer (50 mM Tris-HCL pH 8.0, 100 mM NaCl, 50 µg/ml chicken egg white lysozyme, 50 µM PMSF, Sarcosyl/Triton-X 100 0.1%, each) and then sonicated and centrifuged. The supernatants were used for pull-down experiments: for E. coli RsfA and L14 corresponding volumes of 50 µg soluble protein fractions of co-expressed proteins were applied to beads and aliquots saved as input controls. For human and Zea mays proteins 25 µg soluble fractions were mixed and then applied to the beads. MBP fusions were co-purified with their GST baits on 20 µl glutathione beads and NusA-tagged preys with their MBP fusions on 20 µl amylose beads under buffer conditions indicated above but w/o lysozyme. Binding occurred at room temperature for 30 min. Then, the beads were washed and finally boiled in 50 µl Laemmli buffer. 10 µl of output (∧ = 10 µg protein input) and 10 µg input samples were separated by SDS PAGE using 12% gels. Proteins were transferred onto a polyvinylidene fluoride membrane by semi-dry Western blotting. The recombinant bait and prey proteins were labeled by standard immunodetection procedure and then analyzed by enhanced chemiluminescence.
Human C7orf30 (mtRsfA) and L14mt full-length ORFs were cloned into pcDNA3.1-HA-mCherry [44], pcDNA3.1(+)-HA-VN, and pcDNA3.1(+)-HA-VC [45] (Note: an N-terminal HA tag from the vector backbones was removed under consideration that the native mitochondrial localization peptides of mtRsfA ( = C7orf30) and L14mt are N-terminally exposed).
For localization studies, Hela cells were transfected (100 ng, each plasmid) with mCherry-tagged C7orf30 or L14mt using Promofectin (Promokine, Germany). 100 ng pECFP-Mem (Clontech) was co-transfected to stain cell membranes. 24 h later, MitoTracker Green FM (100 nM f.c., Invitrogen) was added. After washing, DRAQ5 (1∶2,000, Biostatus) was added fur nuclear staining.
For BiFC assays [46], Hela cells were prepared correspondingly. Exceptions: Mitotracker staining was not done and instead of localization constructs, cells were co-transfected with BiFC plasmid constructs (50 ng, each) in combinations as given in Figure 1G.
30 min post DRAQ5 administration cells were analyzed by fluorescence microscopy using a Zeiss LSM 510 Meta confocal laser scanning microscope.
Multiple alignments were generated using ClustalW [47] with the L14 amino acid sequences from E. coli, T. pallidum, S. pneumoniae, Synechocystis PCC 6803, C. jejuni, H. sapiens, Zea mays, Chromobacterium violaceum, Bacillus halodurans, and S. cerevisiae using default parameters. Based on that alignment the conservation scores were calculated with the ConSurf Server [48]. 3D images (Figure 2A) were presented using PyMol 1.5 (http://pymol.org).
Structures of unbound proteins: the E. coli L14 structure was taken from 2AWB PDB entry, chain K [14]. Because the crystal structure of E.coli RsfA is not available, we used I-TASSER server [49] to build a model of that protein. The server built a single model using as templates 2ID1_A and 2O5A_A. The server has estimated the accuracy of the model as 0.90±0.06 (TM-score) and 1.6±1.4 Å (RMSD).
An unconstrained rigid body docking was performed of individual L14 and RsfA structures with GRAMM-X [50]. We then used the coordinates of L14 to superimpose 100 top scored docking models onto the entire 70S unit (2AWB and 2AW7 PDB IDs). Then, each model was evaluated for the backbone clashes between the predicted RsfA position and the rest of the 50S subunit. We defined a clash as having less than 2 Å distance between backbone atoms in order to tolerate some degree of unknown conformational re-arrangement of the 50S components that were not used in docking. Model #17 was the first one in order of the docking score where RsfA had no clashes with other parts of 50S (parts not seen by the docking procedure). Model #17 contained certain surface exposed amino acid residues of L14 that are highly conserved (Figure 2B). To test whether these are involved in mediating the interaction with RsfA they were subjected to alanine substitution constructs (see above and Figure 2A) and analyzed in Y2H experiments (Figure 2C). The interface contacts were defined as having less than 4.6 Å distance between any heavy atoms of the docking subunits. We used PyMol 1.5 (http://pymol.org) for the post-docking analysis and graphics.
ΔrsfA (b0637) [16] and wild type (BW25113) were transformed with a β-galactosidase reporter plasmid, pBAD24-lacZ-HA (based on pBAD24HA) [51], [52] and selected on LB agar containing 50 µg/ml ampicillin. Both were grown overnight in LB in the presence of 50 µg/ml ampicillin and 0.4% glucose as inhibitor of leaky expression. For stationary phase expression cultures were centrifuged at 5,000 rpm (15 min) and pellets were resuspended in the cell-free supernatant of an LB overnight culture (BW25113/ΔrsfA, no plasmid) lacking glucose. β-galactosidase expression was induced with 2% arabinose; the resuspension was adjusted to the same cell density as the previous stationary-phase culture. For logarithmic phase expression overnight cultures were centrifuged at 5,000 rpm for 15 min and pellets were resuspended in fresh LB medium (no glucose) with 50 µg/ml ampicillin for both strains. Cultures were then diluted to OD600 = 0.05 and grown for 2 h. β-galactosidase expression was induced by adding 2% arabinose to the medium. The cultures were shaken at 37°C. Every hour 300 µl suspension was withdrawn, 100 µl from it was loaded into a well of a 96-well plate (flat bottom) and the growth was followed by monitoring the extinction at 600 nm (ELISA spectrophotometer). The rest of aliquots were centrifuged at 12,000 rpm for 5 min and pellets were resuspended in 20 µl loading buffer (2×) Tris-glycine SDS and incubated at 95°C for 5 min to denature proteins. Samples were loaded on SDS-polyacrylamide gel (10%) and the β-galactosidase amount was quantified as relative protein-band intensity using ImageJ 1.45.
For growth competition assays (Figure 3A) the same amount of cells from overnight cultures of wild type and ΔrsfA strains were mixed, yielding a final OD600 of 0.01 in a volume of 5 ml, and incubated with mild shaking either in LB (rich) or M9 medium with 0.4% glucose (poor). Aliquots were withdrawn every 3 h or 6 h or 24 h (depending on the growth rate) and OD600 was measured. Simultaneously, dilutions to approximately 5,000 cells/ml (according to the assumption that 1 OD600 corresponds roughly to 109 cells) were made and 100 µl of each was plated in duplicates on either LB plates or LB plates containing 25 µg/ml kanamycin. The number of colonies (ΔrsfA contained a kanR-cassette, WT not) was counted after incubation at 37°C for overnight. For viability competition experiment in stationary phase (LB medium; Figure 3F) ΔrsfA mutant and wild type strain were separately grown overnight. Subsequently two cultures were diluted to OD600 = 0.005 and incubated with shaking till 0.5 OD600. Then two cultures were mixed and the fitness of ΔrsfA was monitored as numbers of colonies on LB plates (mutant and wild type colonies) and LB plates containing kanamycin (only mutant colonies) after 2, 6, 9, 21, 32, 52, 78 hours of incubation at 37°C.
The kanamycin resistance gene that substituted the rsfA was removed by introducing a flippase-encoding plasmid pCP20 as described elsewhere [53]. The successful flip-out was verified by a genotyping PCR.
For the media shift (Figure 3B) wild type and ΔrsfA strains were grown overnight in LB medium (rich) and then diluted in either LB (rich) or M9 medium (poor) yielding a start OD600 = 0.005. Cultures were incubated at 37°C with shaking (200 rpm) and growth was monitored measuring the OD600 over a time of up to 40 hours.
For curing the phenotype of the ΔrsfA strain during the transition from rich to poor (Figure 3C and 3D) ΔrsfA cells lacking the kanamycin resistance gene and wild type cells were transformed with a plasmid harbouring the gene coding for RsfA fused with a C-terminal His-tag under control of either the native promoter or the IPTG inducible tac-promoter and with the corresponding empty plasmid.
The transformed strains were grown overnight in rich (LB) medium at 37°C and then diluted in poor M9 medium yielding a start OD600 = 0.005 and incubated like described above. At several time points samples were withdrawn and the expression of RsfA was analysed after SDS-PAGE and Western-blot using an antibody directed against the His-tag. The intensity of the RsfA-His bands was quantified using ImageQuant 5.2 and normalized for correction of the input to a non-altered protein band of the Coomassie stained gel.
The gene coding for E. coli RsfA (b0637) was expressed as an N-terminal His6 tag fusion in E. coli BL21(DE3). Expression was induced at OD600 = 0.4 with 0.1 mM IPTG and carried out for 2 h at 30°C to decrease the formation of inclusion bodies. The soluble protein was purified via nickel-nitrilotriacetic-acid-agarose (Qiagen, according to the manufacturer's manual) and anion exchange chromatography (Source 15Q, GE Healthcare). The purified protein was dialyzed against 20 mM Hepes, 6 mM Mg-acetate, 150 mM K-acetate, 4 mM β-mercaptoethanol, pH 7.6 at 0°C.
The gene coding for the mature human mitochondrial RsfA (C7orf30; amino acids 23–234) was expressed and the protein purified like the E. coli RsfA orthologue.
Both proteins were expressed using the Gateway System-compatible plasmid pHGWA [54].
Ribosomes and ribosomal subunits were prepared from E. coli strains CAN20-12E [55] as described [56]. Preparation of mammalian mitochondrial ribosomes and ribosomal subunits (pig liver) followed [57] with minor modifications. Hepes-buffer and TCEP were utilized instead of Tris-buffer and 2-mercaptoethanol, respectively. Isolation of mitochondrial factors are described in [58].
18 pmol 50S ribosomes were incubated with 180 µg poly(U) with or without 360 pmol RsfA in 90 µl for 10 min at 37°C in binding buffer (20 mM Hepes, pH 7.6 at 0° C, 4.5 mM Mg-acetate, 150 mM K-acetate, 4 mM β-mercaptoethanol, 2 mM spermidine, 0.05 mM spermine, H20M4.5K150SH4Spd2Spm0.05). Reaction was further incubated with 10 pmol 30S ribosomes for 10 min at 37°C and then analyzed in poly(U) dependent oligo(Phe) synthesis and sucrose gradient centrifugation.
15 µl of the reaction was used for oligo(Phe) synthesis. 2.4 pmol EF-G together with the ternary complex mix were added yielding 30 µl in binding buffer H20M4.5K150SH4Spd2Spm0.05. The ternary complex mix contained in 15 µl 30 pmol [14C]Phe-tRNAPhe, 45 pmol EF-Tu, 45 pmol EF-Ts, 3 mM GTP and was preincubated 5 min at 37°C. Incubation was at 30°C for 2 min and 12.5 µl aliquots were precipitated with TCA, incubated at 90°C in the presence of 2 drops of 1% (w/v) BSA and filtered through glass filters and counted.
60 µl of the reaction was mixed with 40 µl H20M4.5K150SH4Spd2Spm0.05 and loaded onto a 10–30% sucrose gradient prepared in the same buffer. Centrifugation was carried out at 42,000 rpm for 4 h in an SW60 rotor. The gradient was pumped out from bottom to top and the A260 was measured to obtain the ribosome profile.
The corresponding assay with mitochondrial components from pig liver was performed in H20M4.5K150SH4Spd2Sp0.05 pH7.5 (at 0°C). mtRsfA was pre-incubated with 2.5 pmol large subunit 39S in 80 molar excess over ribosomes, before the same amount of 28S subunits were added; likewise 2.5 pmol 55S ribosomes were incubated with the same amount of RsfA. EF-G1 was added in a 0.8-fold excess over ribosomes. 37.5 pmol of [14C]Phe-tRNA were present and the mitochondrial factors mtEF-Tu and mtEF-Ts, were added both in an excess of 1.5 over Phe-tRNA. The total volume was 100 µl, the main incubation 20 min at 30°C. The following processing was as described above.
The oligo(Phe) synthesis with reassociated 70S ribosomes (Figure 4A, right panel) was performed in the following way: 3 pmol 70 S ribosomes were incubated with 30 µg poly(U) and 6 pmol Ac-Phe-tRNA for 10 min at 37°C. When indicated 60 pmol RsfA was added and the oligo(Phe) synthesis performed as described above. The total volume was 20 µl, the mixture was incubated for 5 min at 37°C.
|
10.1371/journal.pmed.1002280 | Silk garments plus standard care compared with standard care for treating eczema in children: A randomised, controlled, observer-blind, pragmatic trial (CLOTHES Trial) | The role of clothing in the management of eczema (also called atopic dermatitis or atopic eczema) is poorly understood. This trial evaluated the effectiveness and cost-effectiveness of silk garments (in addition to standard care) for the management of eczema in children with moderate to severe disease.
This was a parallel-group, randomised, controlled, observer-blind trial. Children aged 1 to 15 y with moderate to severe eczema were recruited from secondary care and the community at five UK medical centres. Participants were allocated using online randomisation (1:1) to standard care or to standard care plus silk garments, stratified by age and recruiting centre. Silk garments were worn for 6 mo. Primary outcome (eczema severity) was assessed at baseline, 2, 4, and 6 mo, by nurses blinded to treatment allocation, using the Eczema Area and Severity Index (EASI), which was log-transformed for analysis (intention-to-treat analysis). A safety outcome was number of skin infections.
Three hundred children were randomised (26 November 2013 to 5 May 2015): 42% girls, 79% white, mean age 5 y. Primary analysis included 282/300 (94%) children (n = 141 in each group). The garments were worn more often at night than in the day (median of 81% of nights [25th to 75th centile 57% to 96%] and 34% of days [25th to 75th centile 10% to 76%]). Geometric mean EASI scores at baseline, 2, 4, and 6 mo were, respectively, 9.2, 6.4, 5.8, and 5.4 for silk clothing and 8.4, 6.6, 6.0, and 5.4 for standard care. There was no evidence of any difference between the groups in EASI score averaged over all follow-up visits adjusted for baseline EASI score, age, and centre: adjusted ratio of geometric means 0.95, 95% CI 0.85 to 1.07, (p = 0.43). This confidence interval is equivalent to a difference of −1.5 to 0.5 in the original EASI units, which is not clinically important. Skin infections occurred in 36/142 (25%) and 39/141 (28%) of children in the silk clothing and standard care groups, respectively. Even if the small observed treatment effect was genuine, the incremental cost per quality-adjusted life year was £56,811 in the base case analysis from a National Health Service perspective, suggesting that silk garments are unlikely to be cost-effective using currently accepted thresholds. The main limitation of the study is that use of an objective primary outcome, whilst minimising detection bias, may have underestimated treatment effects.
Silk clothing is unlikely to provide additional benefit over standard care in children with moderate to severe eczema.
Current Controlled Trials ISRCTN77261365
| Prior to this trial, evidence on the use of silk garments for the management of eczema was limited.
Three randomised controlled trials (RCTs) had been conducted, but these were small (74 participants in total) and at risk of bias.
The existing evidence was insufficient to guide clinical practice on the use of silk clothing in the management of eczema, and no cost-effectiveness analyses had been undertaken.
We conducted a pragmatic, observer-blind RCT that recruited 300 children with moderate to severe eczema and followed them for six months.
Participants were randomised to receive standard eczema care plus silk clothing (100% sericin-free silk garments; DermaSilk or DreamSkin) or standard care alone.
After six months, there was no evidence of a difference between the groups in eczema severity (Eczema Area and Severity Index score) assessed by research nurses; the 95% confidence interval ranged from 1.5 points favouring silk clothing to 0.5 points favouring standard care, which is not a clinically important difference.
Even if the potential small benefit of silk garments was genuine, our analysis suggests that they are unlikely to be cost-effective within currently accepted thresholds, with an incremental cost per quality-adjusted life year of £56,811.
The CLOTHES Trial is the first large, independent RCT to have evaluated silk garments for the management of eczema.
The results of this trial suggest that silk garments are unlikely to provide additional clinical or economic benefits over standard care for children with moderate to severe eczema.
These results provide robust evidence for health commissioners and prescribers to make informed clinical decisions.
| Eczema (also called atopic dermatitis or atopic eczema) is a chronic, itchy inflammatory skin condition that is common throughout the world [1]. Childhood eczema has a substantial impact on the quality of life of children and their families [2]. Many families are keen to identify new ways of managing the symptoms of eczema using non-pharmacological approaches [3].
Clothing may play a role in either soothing or exacerbating eczema symptoms, and patients are commonly advised to avoid wool because of its tendency to worsen itch, and to use cotton or fine weave materials next to the skin [4]. Specialist clothing is now available on prescription in a variety of forms including sericin-free silk, viscose, and silver-impregnated fabrics. These garments are claimed to be beneficial for the management of eczema as they can help to regulate the humidity and temperature of the surface of the skin, are smooth in texture, and may reduce skin damage from scratching. Some products have anti-microbial properties that could help to reduce the bacterial load on the skin, which may be important in eczema [5].
To date, there have been just three small randomised controlled trials (RCTs) of silk clothing for the management of eczema [6–8]. These trials involved very few participants (n = 22, 30, and 22 participants, respectively), were of generally short duration, did not incorporate an economic evaluation, and were at risk of bias [9].
In view of the limited evidence for the use of silk clothing for eczema management, the UK National Institute for Health Research Health Technology Assessment programme commissioned the CLOTHing for the relief of Eczema Symptoms (CLOTHES) Trial. The trial had two main objectives: (1) to assess whether use of silk garments plus standard eczema treatment reduces eczema severity in children with moderate to severe eczema compared with standard treatment alone, and (2) if so, to establish the likely cost-effectiveness of silk garments.
The protocol for this study has been published [10], and the protocol (S1 Protocol) and statistical analysis plan are available (http://www.nottingham.ac.uk/CLOTHES). The study was approved by the Health Research Authority East Midlands–Nottingham 1 Research Ethics Committee (13/EM/0255), and parents/guardians gave written informed consent (children gave assent as appropriate). The trial was registered on Current Controlled Trials prior to start of recruitment (ISRCTN77261365; 11 October 2013). This study is reported as per CONSORT guidelines (S1 Checklist). A full trial report is available [11].
The CLOTHES Trial was a multi-centre, parallel-group, observer-blind, pragmatic RCT with 6 mo of follow-up. Children aged 1 to 15 y were randomised (1:1) to receive silk garments plus standard eczema care or standard eczema care alone. The primary outcome was assessed by research nurses blinded to the treatment allocation at baseline, 2, 4, and 6 mo. The trial included a nested qualitative evaluation and health economic analysis. Changes to the protocol after start of participant recruitment included amendment of the number of FLG mutations to be included in the genetic analysis and addition of details of the nested qualitative evaluation.
Recruitment took place at five UK medical centres: Nottingham University Hospitals NHS Trust, Royal Free London NHS Foundation Trust, Cambridge University Hospitals NHS Foundation Trust, Portsmouth Hospitals NHS Trust, and Isle of Wight NHS Trust. Participants were identified through secondary care, through primary care, or in response to local media advertising.
Children aged 1 to 15 y were enrolled. All had a diagnosis of eczema according to the UK Working Party’s Diagnostic Criteria for Atopic Dermatitis [12] and a score of nine or more on the Nottingham Eczema Severity Score, denoting moderate to severe eczema over the last 12 mo [13]. All participants had at least one area of active eczema on part of the body that would be covered by the garments.
Children were excluded if they had taken systemic medication (e.g., ciclosporin or oral corticosteroids) or had received light therapy for eczema in the preceding 3 mo, had used wet/dry wraps ≥5 times in the last month, had started a new medication or treatment regimen that may affect eczema in the last month, were currently using silk clothing for their eczema and were unwilling to stop during the trial, or were currently taking part in another clinical trial. Only one child was enrolled per family.
The silk garments used in the trial (DermaSilk or DreamSkin) are licensed as a medical device with a CE mark for use in eczema, denoting that they comply with EU legislation and safety requirements. Two brands were included to improve the generalisability of the trial findings, to avoid commercial advantage to any one company, and to limit the financial commitment for the companies that donated the garments.
The garments are made with antimicrobially protected, knitted, sericin-free silk (100%). Sericin is removed from the silk fibres during manufacture because it is a protein that coats the outside of silk fibres and has the potential to cause allergic reactions. Participants received three sets of garments (long-sleeved undershirts and leggings or bodysuits and leggings, depending on the age of the child) and were instructed to wear the clothing as often as possible during the day and at night.
Standardised usage instructions were provided, and participants were advised to allow topical medications to absorb into the skin prior to wearing the garments. Replacement garments were provided if they were worn out, lost, or no longer fitted during the 6-mo period of the trial.
Participants in both the intervention and control group continued with their standard eczema care in line with National Institute for Health and Care Excellence (NICE) guidance [14], including regular emollient use and topical corticosteroids (or calcineurin inhibitors) for controlling inflammation. Participants were asked not to change their standard eczema treatment for the duration of the trial unless medically warranted. If a skin infection was suspected, participants were advised to contact their normal medical team for confirmation of diagnosis and subsequent treatment.
Core outcomes as defined by the Harmonising Outcomes Measures for Eczema (HOME) initiative [15,16] were included.
Eczema severity captured using the Eczema Area and Severity Index (EASI) [17] was assessed by trained research nurses at baseline, 2, 4, and 6 mo. Baseline EASI score was used as a covariate in the analysis model. EASI is a validated scale recommended as the core outcome instrument for eczema signs [18]. EASI scoring involves an evaluation of four eczema signs (erythema [redness], excoriation [scratching], oedema/papulation [swelling and fluid in the skin], and lichenification [thickening of the skin]) and an assessment of percentage area affected by eczema in four body regions (head and neck, upper limbs, trunk, and lower limbs). Higher scores represent more severe disease.
Secondary outcomes were the following:
Safety outcomes were skin infections requiring antibiotic or antiviral treatment and serious adverse events (SAEs) related to eczema.
Three hundred participants provided 90% power at the 5% significance level (two-tailed) to detect a difference of three points between the groups in mean EASI score. Although this between-group difference is approximately half the published minimum clinically important difference for EASI (suggested from one study in adults receiving systemic therapy) [26], we wanted to be sure that a clinically important difference was not missed. Sample size was based on repeated measures analysis of covariance, a standard deviation (SD) of 13, a correlation between EASI scores at different time points of 0.6, and a loss to follow-up of 10%.
Randomisation was stratified by recruiting centre and by participant age: <2 y, 2 to 5 y, and >5 y. A computer-generated pseudo-random code with random permuted blocks of randomly varying size was created by the Nottingham Clinical Trials Unit. Research nurses accessed the randomisation website via unique user logins. The sequence of treatment allocations remained concealed until the database was locked at the end of the study, when it was revealed to data analysts.
Staff at the coordinating centre sent confirmation of treatment allocation to participants (along with the silk clothing as necessary). Whilst it was not possible to blind participants to their treatment allocation, efforts were made to minimise expectation bias by emphasising in the trial documents that the evidence supporting the use of silk garments for eczema was limited and that it was not yet known if such clothing offered any benefit over standard care. Participant-facing study documents also avoided the use of value-laden terms such as “specialist” or “therapeutic” clothing.
In order to preserve blinding of the research nurses, participants were reminded in the study literature and in their clinic appointment letters/texts not to wear the clothing when they attended clinic or to mention the clothing when talking to the research nurses. All questions relating to the acceptability and use of the clothing were completed using either postal or online questionnaires, and telephone and email contact with participants was made by staff from the coordinating centre whenever possible. If the research nurses became unblinded, this was recorded.
Saliva samples were collected for DNA extraction and FLG genotyping. Only participants of white European ethnicity were included in this analysis, because FLG mutations are ethnically specific. Results for the four most prevalent loss-of-function mutations in the white European population (R501X, 2282del4, R2447X, and S3247X) were obtained for 217 individuals and were used to define genotype categories: FLG wild type (no mutations identified), FLG heterozygote (one FLG null mutation), and FLG homozygote or compound heterozygote (two FLG null mutations).
Analyses were carried out by L. E. B. (trial statistician) using Stata/SE 13.1. The main approach to analysis was modified intention to treat, i.e., analysis according to randomised group regardless of adherence to allocation and including participants who provided data for at least one follow-up time point. Estimates of the intervention effect are presented with 95% confidence intervals and p-values. All regression models included the randomisation stratification variables (recruiting centre and age) as covariates, and baseline scores, if measured. Adjusted differences in means for the intervention group compared to the standard care group are presented for continuous outcomes, and adjusted risk differences and relative risks for binary outcomes. For outcomes collected at the 2-, 4-, and 6-mo visits, we explored whether the effect of the trial garments on the outcome changed over the study period by including an interaction term between treatment group and time point in the model. As there was no evidence of a differential effect over time for any outcomes, we report a single estimate per outcome that averages the treatment effect over all time points.
The primary analysis used a multilevel model with observations at 2, 4, and 6 mo nested within participants. The model used a random intercept and slope at the participant level with an unstructured covariance matrix for these random effects. The model assumed that missing EASI scores were missing at random given the observed data. EASI scores were right skewed at all time points. Diagnostic plots indicated that the assumptions for the multilevel model in the original EASI units were not met. The data were log-transformed for analysis and the treatment effect presented as a ratio of geometric means [27,28]. This ratio was back-transformed to the original EASI units to facilitate interpretation of findings.
Sensitivity analyses for the primary outcome adjusted for variables that had an observed imbalance between the groups at baseline, used multiple imputation for missing outcome data, and explored the impact of adherence in wearing the clothing by estimating the complier average causal effect (CACE) at 6 mo using instrumental variable regression.
A planned subgroup analysis based on presence or absence of loss-of-function mutations in FLG (which are associated with impaired skin barrier function and more severe disease) was conducted for the primary outcome by adding an interaction term between allocated treatment and FLG genotype (none, one, or two FLG null mutations) to the primary analysis model.
The global assessment scores (IGA and PGA) were dichotomized into clear, almost clear, or mild eczema versus moderate, severe, or very severe eczema, and analysed using generalised estimating equations. The mean weekly POEM scores, percentage of days that topical steroids were used, and quality of life outcomes were analysed using linear models (weighted according to the number of questionnaires completed for the weekly POEM and topical steroid use). The TIS score was analysed using the multilevel model framework as outlined above for the primary outcome (not transformed). Changes to treatment regimen were based on whether a participant had reported treatment escalation over the 6-mo RCT period and were analysed using a generalised linear model. Skin infections were analysed using negative binomial regression. SAEs and durability and acceptability of use of the garments were summarised descriptively.
Adherence in wearing the trial clothing was summarised using the percentage of days and nights that the study clothing was worn. Participants were classified as being broadly adherent if they wore the trial clothing for at least 50% of the days or 50% of the nights. This classification was done for participants for whom at least half (12/24) of the weekly questionnaires were completed, and sensitivity analysis explored the impact of different assumptions for those participants who completed less than 50% of the weekly questionnaires. Adherence with the trial clothing was explored descriptively according to age and baseline eczema severity using correlation coefficients.
Full details of the analysis are documented in the statistical analysis plan, which was finalised prior to database lock and release of treatment allocation codes for analysis.
Following concerns that the baseline EASI scores appeared lower than might be expected for children with moderate to severe eczema, an additional post hoc analysis was conducted to explore the interaction between baseline severity and treatment group by adding an interaction term between allocated group and baseline EASI score (log-transformed and continuous) to the primary analysis model.
Public and patient involvement (PPI) was embedded throughout the CLOTHES Trial. Various PPI methods such as online surveys, discussion groups, and patient panels were used to inform multiple aspects of the trial design including choice of comparator, eligibility criteria, potential barriers to participation, and outcome measures. PPI members of the trial team also contributed to the development of patient-facing study materials and took part in media interviews to enhance recruitment. A PPI representative was a co-applicant on the grant and was involved in all stages from trial design through to data interpretation and write up, and another PPI representative was a member of the trial steering committee.
The study results will be published on the CLOTHES Trial website, and a written summary and child-friendly animated film will be sent to trial participants.
The within-trial economic analysis (conducted by T. H. S. using Stata/SE 14.1) compared the costs and QALYs in the standard care and intervention groups from the perspective of the NHS. We attached published unit costs (2014–2015 UK pounds sterling) [29–31] to individual-level quantities of resource use (S1 Table) and estimated the mean cost per participant incorporating the cost of the intervention and wider healthcare resource use (primary care, secondary care, and medications).
QALYs were estimated using linear interpolation and area under the curve analysis, adjusting for baseline values, age, and recruitment centre. A regression-based approach (seemingly unrelated regression equations) [32] was used for the statistical analysis. The level of uncertainty associated with the decision over which option was most cost-effective was explored using non-parametric bootstrapping [33] to construct the cost-effectiveness acceptability curve [34]. Neither costs nor QALYs were discounted reflecting the time frame.
To test the impact of taking an alternative approach to costing the silk garments, sensitivity analysis included an estimate of the amount pharmacists are reimbursed for each item of clothing they prescribe. This analysis was based on the NHS Business Services Authority formula to estimate the actual cost to the NHS. The analysis was rerun using the March 2015 tariff data [35], where the average discount was 7.43% and the pharmacist’s professional fee £0.90 per prescription item.
Three hundred children were randomised between 26 November 2013 and 5 May 2015 (last study visit 21 October 2015). The primary analysis included 141 participants in each group who had at least one primary outcome assessment after baseline (Fig 1). For all but four participants, outcome assessments were performed by the same nurse at all study visits.
For the weekly online questionnaires (24 questionnaires over 6 mo), 126/149 (85%) participants in the intervention group and 127/151 (84%) participants in the standard care group completed 12 questionnaires or more. The median number completed was 22 (25th to 75th centile 17 to 24) in both groups.
Participants had a mean age of 5 y, 42% were girls, and 79% were white. At recruitment, 72% had moderate or severe eczema, as judged by the IGA (Table 1). Demographic and clinical characteristics were well balanced at baseline apart from gender and parental reported history of asthma and food allergy (Table 1). The mean baseline EASI score was slightly higher in the intervention group as more children had a baseline EASI score of over 30 points (14 participants in the intervention group, four participants in the standard care group). However, the median and interquartile range for EASI score were similar between the groups (Table 2).
Adherence in wearing the garments was good. The garments were worn more often at night than in the day (median of 81% of nights [25th to 75th centile 57% to 96%] and 34% of days [25th to 75th centile 10% to 76%]) (Fig 2; S2 Table). Adherence in wearing the garments was not associated with age or eczema severity at baseline (S2 Table). Contamination of the standard care group was low; six participants reported wearing silk clothing during the trial (including one participant who was allocated to the standard care group but was sent the silk clothing in error; this participant was included in the analysis according to randomised allocation).
Acceptability of the garments as assessed at 6 mo suggested that 85/121 (70%) participants were satisfied or very satisfied with the clothing (95% CI 61% to 78%), and 89/121 (74%) participants were either happy or very happy to wear the garments (95% CI 64% to 81%). Some participants raised concerns about the garments, including poor durability and fit.
Research nurses remained blinded to treatment allocation for 289/300 (96%) of participants. Unblinding occurred for three participants in the standard care group and eight in the intervention group.
For the primary outcome of eczema severity, there was no difference between the groups in the nurse-assessed EASI scores. For EASI scores averaged over the 2-, 4-, and 6-mo follow-up visits, the adjusted ratio of geometric means was 0.95, with 95% CI 0.85 to 1.07 (p = 0.43) (Table 2; Fig 3). This confidence interval equates to a difference of approximately −1.5 to 0.5 points in the original EASI units.
All sensitivity analyses for the primary outcome (adjusting for additional baseline factors, imputing missing values, and exploring the impact of adherence [CACE analysis]) were supportive of the primary analysis (S3 Table). There was no differential effect of the clothing on EASI score (eczema severity) according to FLG subgroup (S4 Table) or severity of eczema at baseline (S5 Table).
For the secondary outcomes, there were no between-group differences in nurse-assessed eczema severity (IGA, TIS), quality of life (DFI, EQ-5D-3L, CHU-9D), or medication use (percentage of days eczema medications used, escalation of eczema treatment) (Tables 2 and 3). However, small differences were observed for two of the participant-reported secondary outcomes of eczema severity (PGA, POEM) (Tables 2 and 3; Fig 4). Safety outcomes (number of skin infections and hospitalizations due to eczema) were similar in the two groups (Table 4).
The economic evaluation included all participants with complete resource use and ADQoL data at baseline and 6 mo (n = 273). The cost of a single set of tops and leggings ranged from £66.02 to £155.49, depending on the size of the child. The mean cost of silk garments for 6 mo, including initial and replacement garments, was £318.52 (SD £136.60) per participant in the base case (Table 5). The mean number of sets of garments (tops and leggings) per participant in the base case was 4.15 (SD 1.56). Sixty-one (45.54%) intervention participants received replacement garments over the 6 mo.
Combined with wider health resource use, the adjusted mean difference in cost per participant was £364.94 (95% CI £217.47 to £512.42, p < 0.001) for those who received silk garments compared to those who did not in the base case (Table 5). The difference in total costs between groups reflects the cost of the intervention; wider NHS costs were not significantly different between groups (£48.57 higher per participant on average in the intervention group, 95% CI −£105.92 to £203.05, p = 0.537). For resource use and costs for all resource items, see S6 and S7 Tables.
The adjusted mean difference in QALY per participant was 0.0064 (95% CI −0.0004 to 0.0133, p = 0.07) (Table 5). The adjusted incremental cost per QALY was £56,811, suggesting that silk garments for AE are not cost-effective using currently accepted thresholds. At a willingness to pay of £30,000 per QALY, the probability of silk garments being cost-effective was 12.13%. This conclusion did not change in sensitivity analysis testing an alternative approach to costing the silk garments. Although the cost of silk garments was reduced with the alternative approach, at £53,989 per QALY, the estimated incremental cost per QALY was still over the accepted NICE threshold value (see S8 Table).
This trial found little evidence of clinical or economic benefit of using silk garments in addition to standard care, compared with standard care alone, in children with moderate to severe eczema. There were no differences between the treatment groups for any of the outcomes that were assessed by research nurses, who were unaware of participants’ treatment allocation, and the percentage of days on which topical corticosteroids or calcineurin inhibitors were used did not differ between the groups. The 95% confidence intervals around the primary efficacy estimates were narrow, suggesting that a clinically important treatment effect is unlikely to have been missed, and sensitivity analyses (imputing missing values, adjusting for baseline imbalances, and exploring the impact of adherence in wearing the garments) supported the primary analysis.
Subgroup analysis based on FLG genotype showed no evidence of differential treatment response in children with an inherited impairment in skin barrier function, and a post hoc analysis exploring the impact of baseline eczema severity on the primary outcome showed no effect, suggesting that children with more severe disease were no more likely to benefit from silk clothing than those with milder disease.
The trial garments are marketed as possessing antimicrobial properties, but this study found no evidence to suggest a reduction in the number of skin infections in those using the clothing compared to those randomised to standard care alone.
Of the seven unblinded secondary outcomes, two (POEM and PGA) showed small differences in favour of the silk garments, most noticeably in the first 3 mo of the trial. Whilst these small differences could have been genuine, they are most likely due to an expectation bias that declined with time. Our nested qualitative study (to be reported separately) highlighted the hopes that both children and parents placed on the silk clothing [11]. A previous eczema trial reported differences between blinded and unblinded outcomes when expectation regarding the benefits of the trial intervention was high [36].
To our knowledge, there have been no further RCTs on the effectiveness of silk garments for eczema since the CLOTHES Trial began (search updated 14 March 2016), and meta-analysis of the available silk clothing trials is not possible due to heterogeneity of designs. Additional brands of silk garments have since become available for use in eczema (e.g., Skinnies), but these have not been formally evaluated in RCTs. At the time of commissioning this research (2011), £840,272 was spent on prescriptions for silk garments per annum in the UK (for all indications). By 2014, this amount had risen to £2,082,810 per annum [37–40].
The CLOTHES Trial was an adequately powered RCT, with high follow-up rates and good adherence. The pragmatic study design meant that use of silk garments was evaluated as they might be used in normal practice, with mixed patterns of adherence. The trial placed special emphasis on objective outcome measures in order to minimise response bias.
It is possible that our emphasis on objective eczema severity outcomes meant that some important potential benefits were not captured in the primary analysis. Other factors, such as improvements in quality of life or a reduction in symptoms (especially itch and sleep loss, as measured by POEM), may be important drivers in determining whether or not patients feel that the garments are helpful. Nevertheless, we found no evidence of improved quality of life amongst trial participants using a range of validated scales.
Eczema severity scores improved for both groups during the trial, probably due to a combination of regression to the mean and regular monitoring of the eczema resulting in enhanced adherence to standard care. It is possible that treatment effects were masked by these general trial effects.
The study has strong external validity as it was pragmatic in design to reflect normal clinical practice, and participants were recruited from five UK medical centres covering a range of urban and rural settings. We recruited children with a range of eczema severities, but the majority had moderate to severe disease; 32% had at least one mutation in the FLG gene, a proportion typical of eczema patents with moderate or severe disease [36]. Overall, 49% had self-reported food allergy, which is high for children with moderate to severe disease, and 15% reported a history of anaphylaxis. However, these data were collected by self-report and so may include food intolerance as well as food allergy.
We are unable to comment on the effectiveness of the silk garments if used continuously day and night, although sensitivity analysis found no evidence of improved outcomes in those who adhered more fully in wearing the garments. It is also possible that the beneficial effects of silk garments are best realised during a period of eczema flare, and daily use of the garments in the CLOTHES trial could have led to more rapid deterioration of the clothing than might have been seen if the garments were worn occasionally when the eczema was at its worst.
This is the first large, independent trial to have evaluated silk garments for the management of eczema. The nested economic evaluation suggests that use of these garments is unlikely to be cost-effective for health providers, even if the small observed benefits were genuine. These trial results provide health commissioners with a better evidence base on which to make informed decisions about silk garments for eczema. Whether or not parents feel that the small benefits identified in some of the secondary outcomes are sufficient to justify purchasing these garments is something for individuals to consider on a case-by-case basis.
|
10.1371/journal.ppat.1002380 | Indirect DNA Readout by an H-NS Related Protein: Structure of the DNA Complex of the C-Terminal Domain of Ler | Ler, a member of the H-NS protein family, is the master regulator of the LEE pathogenicity island in virulent Escherichia coli strains. Here, we determined the structure of a complex between the DNA-binding domain of Ler (CT-Ler) and a 15-mer DNA duplex. CT-Ler recognizes a preexisting structural pattern in the DNA minor groove formed by two consecutive regions which are narrower and wider, respectively, compared with standard B-DNA. The compressed region, associated with an AT-tract, is sensed by the side chain of Arg90, whose mutation abolishes the capacity of Ler to bind DNA. The expanded groove allows the approach of the loop in which Arg90 is located. This is the first report of an experimental structure of a DNA complex that includes a protein belonging to the H-NS family. The indirect readout mechanism not only explains the capacity of H-NS and other H-NS family members to modulate the expression of a large number of genes but also the origin of the specificity displayed by Ler. Our results point to a general mechanism by which horizontally acquired genes may be specifically recognized by members of the H-NS family.
| Pathogenic Escherichia coli strains and other enterobacteria carry genes acquired from other bacteria by a process known as horizontal gene transfer. Proper regulation of the genes that are expressed in a given moment is crucial for the success of the bacteria. The protein H-NS is a global regulator that binds DNA and maintains a large number of genes silent until they are required, for example, to sustain the bacteria's colonization of a new host. Ler is a member of the H-NS family that competes with H-NS to activate the expression of a group of horizontally acquired genes that encode for a molecular machine used by E. coli to infect human cells. Ler and H-NS share a similar DNA-binding domain and can bind to different DNA sequences. Here, we present the structure of a complex between the DNA-binding domain of Ler and a natural DNA fragment. This structure reveals that Ler recognizes specific DNA shapes, explaining its capacity to regulate genes with different sequences. A single arginine residue is key for the recognition of a DNA narrow minor groove, which is one of, though not the only, hallmarks of the DNA shapes that are recognized by H-NS and Ler.
| Enteropathogenic Escherichia coli (EPEC) and enterohaemorrhagic E. coli (EHEC) are causal agents of infectious diarrhea. While the former is responsible mainly for infantile diarrhea, EHEC infections are associated with hemorrhagic colitis and may produce a life-threatening complication known as hemolytic uremic syndrome. EPEC and EHEC are non-invasive pathogens that produce characteristic attaching and effacing (A/E) intestinal lesions [1]. The genes required for the formation of A/E lesions are clustered on a pathogenicity island known as the locus of enterocyte effacement (LEE). LEE genes are organized in five major operons (LEE1 to LEE5) and several smaller transcriptional units and they encode the components of a type III secretion system (TTSS), an adhesin (intimin) and its receptor (Tir), effector proteins secreted by the TTSS, chaperones, and several transcription regulators [2]. The first gene of the LEE1 operon encodes the LEE-encoded regulator Ler, which is essential for the formation of A/E lesions in infected cells [3], [4] and for the in vivo virulence of A/E pathogenic E. coli strains [5].
Ler (123 amino acids, 14.3 kDa) is the master regulator of LEE expression and is required to activate LEE genes that are otherwise repressed by the histone-like nucleoid structuring protein H-NS [2].
The H-NS protein, best characterized in E. coli and Salmonella, is a member of a family of transcriptional regulators with affinity for AT-rich DNA sequences that mediate the adaptive response of bacterial cells to changes in multiple environmental factors associated with colonization of different ecological niches, including human hosts. H-NS is usually an environmentally-dependent transcriptional repressor. H-NS-mediated repression (usually termed silencing) is alleviated either by alterations in physicochemical parameters (i.e., a transition from low (25°C) to high (37°C) temperature), by the activity of proteins that displace H-NS from its target DNA sequences, such as Ler, or by a combination of both. H-NS regulation is strongly associated with pathogenicity, thus understanding the basis of the selective regulation of virulence genes could lead to sustainable antimicrobial strategies that are less susceptible to acquiring resistance.
In addition to the LEE genes, Ler is also involved in the regulation of other horizontally acquired virulence genes located outside the LEE loci and scattered throughout the chromosome of A/E pathogenic strains [3], [6], [7]. However, Ler does not regulate other H-NS-silenced operons such as bgl [8] and proU [3]. This observation shows that Ler is not a general antagonist of H-NS, but a specific activator of virulence operons acquired by horizontal transfer (HT). Selective regulation of HT genes has been demonstrated in the plasmid R27 encoded H-NS paralogue (H-NSR27) and in chromosomal H-NS in the presence of a co-regulator of the Hha/YmoA family [9].
The mechanism of Ler-mediated activation has been extensively studied in operons located both within the LEE loci, such as LEE2/LEE3 [10], grlRA [11], [12] and LEE5 [8], and outside, including nleA (for non-LEE-encoded effector A) [13] and the lpf1 fimbrial operon [6], [14]. These studies suggest that Ler counteracts the silencing activity of H-NS by directly binding to DNA and displacing H-NS from specific promoter regions. Ler does not exert dominant negative effects on H-NS function and there is no evidence of a direct interaction between Ler and H-NS [8]. Despite the wealth of biochemical/biophysical data, including the proposal of a DNA sequence consensus motif for H-NS [15], the lack of structural data on the complexes formed between H-NS or H-NS family members and DNA has until now prevented a detailed understanding of the mechanism of DNA recognition and the basis of the selectivity within H-NS family proteins.
All H-NS-related proteins identified to date are predicted to be organized in two structurally different domains. While the oligomerization domains of Ler and H-NS differ greatly, their DNA binding domains are very similar, thereby suggesting that they account for the similar recognition properties of both proteins, and possibly also for their distinct selectivity. While a possible interplay between protein oligomerization and DNA binding cannot be ruled out, a detailed understanding of the recognition mechanism by individual DNA-binding domains is a prerequisite for further studies.
The C-terminal domain of Ler (CT-Ler), exhibits significant amino acid homology with the C-terminal H-NS DNA-binding domain (CT-H-NS; 36.0% identity, 63.8% similarity) and its deletion abolishes DNA binding [16]. CT-Ler contains a sequence (TWSGVGRQP) similar to the consensus core DNA-binding motif found in H-NS-like proteins (TWTGXGRXP) [17]. Here we present the solution structure of a complex formed by CT-Ler bound to a natural occurring DNA sequence of the LEE2/LEE3 regulatory region. This is the first report of a DNA complex that includes a member of the H-NS family characterized at atomic detail. Our results reveal that CT-Ler does not participate in base-specific contacts but recognizes specific structural features in the DNA minor groove. The indirect readout mechanism can be extended to H-NS and other H-NS family members and explains their capacity to modulate the expression of a large number of genes. The CT-Ler/DNA structure provides clues for the mechanism by which HT genes may be specifically recognized by members of the H-NS family and illustrates the general features of DNA minor groove readout.
We used a CT-Ler construct encompassing residues 70–116 (Figure 1A). This construct gave rise to a folded and functional domain (Figure S1) with excellent solubility and long-term stability. Residues 117–123 are part of an extension that is dispensable to counteract H-NS repression [18]. NMR spectra of a construct including these residues showed that they are disordered and have no effect on the structure of the folded domain, as seen by the exact coincidence of the cross-peak position of most residues in HSQC NMR spectra of different constructs (Figure S2).
The sequence of the short DNA fragment used to form the complex was based on the regulatory region of the LEE2/LEE3 operons spanning positions -221 to -101. This region was protected by Ler in footprinting experiments [10]. Seven 30 bp long dsDNA, LeeA-LeeG, with a 15 bp overlap between consecutive fragments (Figure 1B, Table S1) were tested for binding to CT-Ler using fluorescence anisotropy. As positive and negative controls, we used two 30-mer duplexes: an adenine tract that was previously employed to study the DNA-binding properties of CT-H-NS, (GGCAAAAAAC)3 [19] and (GTG)10 (Figure S3). CT-Ler showed the highest affinities for LeeF and LeeG (Figure 1B) and we further analyzed its binding to the 15 bp overlapping region of theses two fragments, namely LeeFG (AAATAATTGATAATA). Fluorescence anisotropy titrations showed small but systematic deviations from the 1∶1 model, suggesting simultaneous multiple binding to this DNA sequence (Figure 1C). Since the consensus binding motif proposed for H-NS is only 10 bp long [15] we designed a new 15 bp DNA, LeeH (GCGATAATTGATAGG), containing the central 10 bp of LeeFG flanked by GC base pairs for thermal stability. LeeH partially matches the proposed H-NS consensus sequence (tCG(t/a)T(a/t)AATT) [15]. A good fit to a 1∶1 model with apparent Kd 1.10±0.05 µM was observed for this duplex (Figure 1C).
The complex of CT-Ler with LeeH was solved by a combination of NMR and small-angle X-ray scattering (SAXS). The structure determination protocol consisted of the independent calculation of the structure of bound CT-Ler and DNA, followed by intermolecular NOE (iNOEs) driven docking and a final scoring including SAXS data. CT-Ler structures were calculated based on 1302 NOE distance restraints, together with torsion angle and experimentally determined hydrogen bonds. The restraint and structural statistics of the 20 lowest energy structures are shown in Table S2. None of the structures contained distance or dihedral angle violations >0.5 Å or 5°, respectively.
The pattern and intensities of bound DNA NOEs were typical of a B-form. The DNA structure was optimized in explicit solvent using experimental restrains determined in the bound form, starting from canonical B-DNA as described in the Materials and Methods section. The absence of major distortions in the DNA structure caused by CT-Ler binding was confirmed by the good agreement between the experimental SAXS curve of free LeeH and the prediction based on the DNA model extracted from the final complex (Figure S4).
The DNA region most affected by CT-Ler binding, identified by the combined chemical shift perturbations of nucleotide protons, is centered in the symmetrical 4 bp AT-tract, AATT (Figure 2A). The largest chemical shift perturbations of CT-Ler (Figure 2B) were observed for residues Val88 to Arg93. The 30 assigned iNOEs involve protein residues located in the region where the chemical shift perturbations were observed. On the basis of these iNOE restraints and the mapped interfaces, 400 CT-Ler/LeeH complex structures were generated as described in Materials and Methods and ranked by energy and NMR intermolecular restraint (irestraint) violations. The quality of the structures was confirmed by comparing the predicted and experimentally determined SAXS curves of the complex. The SAXS profile predicted for the best NMR-derived complex structure is in good agreement with the experimental curve (Figure 3A). The scatter plot in Figure 3B shows that, in general, the best NMR structures also fit SAXS data well. The final ensemble of 20 structures was selected using a scoring function that combined docking energy and measures of the agreement with experimental NMR and SAXS data (red circles). The ensemble is well defined (Figure 3C), with a pairwise RMSD (heavy atoms) of 1.30±0.38 Å and all conformers exhibited good geometry, no violations of iNOE distance restraints >0.5 Å and correctly explained the SAXS data. Most of the protein residues are in the core region of the Ramachandran plot. The small irestraint deviations illustrate that the protein-DNA interface is well defined, allowing us to elucidate a molecular basis for CT-Ler/LeeH recognition.
The structure of DNA-bound CT-Ler contains a central helix (residues 93–101) and a triple-stranded antiparallel β-sheet (β1:76–78, β2:84–85, β3:109–110). The β1-β2-hairpin is connected to the α-helix by a loop (Loop2:86–92). A turn and a short 310-helix (105–108) link the helix to the β3 strand. The similarity between the Cα and Cβ secondary chemical shifts of the free and bound forms indicate that the secondary structure is retained upon binding (Figure S5). The overall protein fold is analogous to that previously described for CT-H-NS in the absence of DNA [19].
CT-Ler binds as a monomer inserting Loop2 and the N-terminal end of the α-helix into the DNA minor groove and contacting the central 6 bp region (A6A7T8T9G10A11) (Figure 4). The complex buries 953±55.64 Å2 of surface area and is stabilized by non-specific hydrophobic and polar contacts, involving mainly the sugar-phosphates backbone and residues of the consensus DNA-binding motif found in H-NS-like proteins. Residues Trp85, Gly89, Arg90 and Pro92 (Figure 1A), highly conserved among H-NS-like proteins, are located in the complex interface (Figure 4B), and all gave rise to iNOE restraints with DNA. A summary of the observed intermolecular contacts is shown in Figure 4D.
The interaction surface of CT-Ler is positively charged and the Arg90 side chain is deeply inserted inside a narrow minor groove (Figure 4B and C). In addition, Arg93 at the N-terminus of the α-helix and the helix-dipole moment itself create a positively charged region that points into the negatively charged minor groove.
The width of the LeeH minor groove varies along the sequence and deviates significantly from the average value of canonical B-DNA (Figure 5). The groove progressively narrows towards the A7pT8 base step, and widens at the T9pG10 base step. The DNA electrostatic potential is modulated by the width of the minor groove. The guanidinium group of Arg90 interacts with the narrowest region of the groove where the electrostatic potential is most negative (Figure 5A and B). The approach of Loop2, where Arg90 is located, is enabled by the adjacent widening of the minor groove.
Sequence-dependent variations of DNA structure can be described in terms of helical parameters, such as roll and helix twist (Figure 5C and D). The roll angle is most negative (−4.64°±1.38) at the A7pT8 base step and is small or negative for most of the steps in LeeH except for the pyrimidine-purine base steps, which show large positive values. A series of consecutive small/negative roll angles leads to the narrowing of the minor groove [20]. The groove widening at T9pG10 can be traced to a combination of positive roll and a small helix twist of 33.8°±0.8, indicating that the segment is slightly unwound with respect to the standard B-form. The region including the A6A7T8T9 stretch is slightly overwound, with an average helix twist of 37.4°±1.6.
To verify the relevance of Arg90 in the interaction, we replaced this residue by glycine (R90G), glutamine (R90Q) or lysine (R90K) and tested their effects on the affinity of CT-Ler to LeeH. All CT-Ler variants were properly folded, as determined from NMR, and their interaction with LeeH was measured by fluorescence anisotropy (Figure 6A). The mutated domains showed no affinity to LeeH or highly reduced affinity (R90K), thereby confirming that Arg90 is an essential residue.
The effect of these mutations on the binding of Ler(3–116), including the oligomerization domain, to the LEE2 regulatory region (positions −225 to +121) was determined using electrophoretic mobility shift assays (EMSA) (Figure 6B). In agreement with the results obtained with the isolated CT-domain, DNA binding by Ler is abolished by R90Q and R90G mutations and strongly reduced in the case of the R90K variant. These experiments confirm the essential role of Arg90 in the context of the oligomeric Ler protein and for the range of binding sequences present in one of its natural targets.
The structure of the CT-Ler/LeeH complex does not show base specific contacts. On the contrary, the structure of the complex suggests that CT-Ler recognizes local structural features of the minor groove that may be associated with distinct DNA sequences. In order to gain some insight into the range of DNA sequences that can be recognized by CT-Ler, we measured the dissociation constants of complexes formed by two series of short DNA duplexes related to the LeeH sequence. In the first series we introduced a single base pair replacement in each of the ten central positions of LeeH. Adenines and thymines were replaced by guanines and cytosines, respectively, and guanine in position 10 was mutated to adenine, to preserve the purine-pyrimidine sequence. In the second series, we compared the binding of CT-Ler to several 10-mer duplexes. One of these contained the AT-tract (AATT) that interacts with CT-Ler in the LeeH complex flanked by GC base pairs to ensure thermal stability. Variants were designed to test the effect of interrupting the AT-tract by TpA steps at a number of positions.
Affinity to CT-Ler was measured by fluorescence anisotropy. The results are shown in Figure 7 and the DNA sequences and dissociation constants are listed in Table S3.
Figure 7A shows the relative Kd values of the single base-pair replacements of LeeH. The largest effects were observed when the base pairs of A6 or A7 were replaced. The base pair of G10 resulted to be similarly relevant. A smaller effect was observed at the position of T8. Small non-specific effects were observed in all the remaining sites except that of A4. The most affected base pairs were at the sites where the minor groove width in LeeH is more different from the standard B-DNA and define the features that we hypothesize to be recognized by CT-Ler: the narrow groove where the Arg90 side chain is inserted and the wide adjacent region that enables the approach of Loop2.
Figure 7B show the relative dissociation constants of the complexes formed by the 10-mer duplexes. The presence of TpA steps in CGCAATAGCG, CGCTATAGCG and CGCTTAAGCG results in a decrease in the stability of the complexes. The remaining three sequences (CGCAATTGCG, CGCAAATGCG, and CGCAAAAGCG) show AT-tracts of the same length but their affinity for CT-Ler differs. The complex with the A4 stretch is 2-fold less stable than that containing the AATT motif.
The AT-tract in LeeH is terminated by a TpG pyrimidine-purine step. Replacing it by a TpC pyrimidine-pyrimidine step in a 10 bp duplex had only a minor effect on the affinity for CT-Ler (cf. AATT and AATTC in Table S3). Interestingly, replacement of the T9pG10 step in LeeH by the alternative pyrimidine-purine step, TpA, resulted in a major loss of stability of the complex.
The DNA binding domains of Ler and H-NS share a high degree of similarity both in sequence and in structure. We carried out experiments to specifically test two key points that are apparent from the analysis of the Ler/LeeH complex, namely the role of the conserved arginine residue (Arg90 in Ler, Arg114 in H-NS) in Loop2 and the requirement for an AT-tract and the effect of interrupting TpA steps.
H-NS Arg114, corresponding to Arg90 in Ler, was mutated to glycine and the affinity towards the −225 to +121 LEE2 region was compared with that of the wild type form by EMSA. As in the case of Ler, replacing the arginine residue in Loop2 results in a substantial loss of affinity (Figure 8A). However, H-NS retains some residual activity even when arginine was replaced by glycine while this drastic mutation caused a complete loss of activity in the case of Ler.
The requirement for a narrow minor groove in the case of Ler can be assessed by the relative affinities towards the AATT and TATA 10-mer duplexes. Titrations of CT-H-NS with both oligonucleotides (Figure 8) provided dissociation constants of circa 41 μM for the AATT complex and 102 μM, 2–3-fold larger, for the TATA complex. CT-Ler showed similar relative affinities for the same oligonucleotides (Table S3), thereby suggesting that these two domains have similar requirements for a narrow minor groove.
As many H-NS and Ler target sequences may overlap, the relative affinity of the DNA-binding domains of these two proteins is relevant. As the CT-Ler complex studied included only the structured domain, we compared CT-Ler with the CT-domain of H-NS including only residues 95 to 137, excluding linker residues. This H-NS construct is properly folded as shown by the observation of well resolved NMR spectra (Figure 8). The same natural DNA fragment (LEE2 positions −225 to +121) used in EMSA assays with Ler (Figure 6B) and H-NS (Figure 8A) was selected to compare the affinities of the CT-domains of these two proteins. The large number of binding sites for Ler and H-NS in this extended DNA fragment, as shown by footprinting experiments, allows the assessment of the relative overall affinities of the two domains for the whole range of sequences present in one of their common natural targets. The affinity of CT-Ler is larger than that of CT-H-NS, which under the conditions of the experiment hardly caused any retardation (Figure 8C). This observation contrasts with the similar affinity towards the same DNA fragment shown by longer constructs of Ler and H-NS that include the oligomerization and linker domains (cf. Figure 6B and 8C) and highlights varying relevance of interactions outside the folded CT-domains of these two proteins. The contribution of residues outside of the structured H-NS DNA-binding domain has been previously described [21], [22].
The structure of the complex between CT-Ler and LeeH shows that DNA shape and electrostatics, rather than base specific contacts, form the basis for the recognition of the CT-Ler binding site. This mechanism is referred to as indirect readout. Arg90 is a key residue for the CT-Ler interaction with DNA. Its side chain is inserted deep into a narrow minor groove. The requirement for Arg90 is strict in the case of CT-Ler and the R90G and R90Q mutants of Ler are totally inactive. The R90K mutant shows some residual binding suggesting that a positive charge is required. Arginine interactions with the DNA minor groove have been described in eukaryote nucleosomes [23], [24] and in DNA interactions by a nucleoid-associated protein of Mycobacterium tuberculosis [25]. These observations suggest that this mechanism may be universal for indirect DNA recognition of AT-rich sequences. A correlation between minor groove width and the electrostatic potential has been demonstrated as well as the preference for arginine binding to the narrowest regions where the electrostatic potential is more negative [23].
For CT-Ler, the narrow minor groove may be provided by a relatively short AT-tract as only the Arg90 side chain has to be inserted. The minimum width in the AATT motif is observed at the ApT step, matching the site where the guanidinium group is inserted. Continuous polyA tracts of 4 (Figure 7) and 6 nucleotides (Figure S3) of length give less stable complexes than sequences combining A and T. However, the presence of highly dynamic TpA steps [26] interrupting the AT-tracts decreases the affinity for CT-Ler. The presence of guanine, with its 2-amino group extending into the minor groove and increasing its width is also predicted to destabilize the insertion of the arginine side chain. We explored the effect of introducing TpG or TpA steps in the sequence recognized by CT-Ler. Figure 7 clearly shows that an uninterrupted AT-tract is needed for an efficient interaction with CT-Ler. However, a narrow AT-tract is not the only requirement for CT-Ler interaction. The lower affinity of the G10A variant of LeeH shows that, next to the narrow region, a rigid wide minor groove is also required to enable the access of Loop2 delivering the side chain of Arg90 into the narrowest region of the minor groove. Both sequences, T9pG10 in LeeH and T9pA10 in the mutated duplex, could adopt wide minor grooves. However, while the former is expected to provide a permanently wide groove, the flexible TpA step may switch between expanded and compressed forms, interfering with the approach of Loop2 directly or indirectly through the entropic penalty associated to stiffening of the DNA in the complex.
The structure of the complex as well as the affinity data with DNA sequence variants show that CT-Ler recognizes a pattern in the minor groove of DNA formed by two consecutive regions that are narrower and wider, respectively, with respect to standard B-DNA and show the optimal shape and electrostatic potential distribution for binding.
This structural pattern is present in the free LeeH DNA fragment as shown by the observation of diagnostic inter-strand NOES between AdeH2 and ThyH1' protons of A7/A23 and T25/T9, respectively supporting minor groove narrowing both in the free and bound forms of LeeH. Moreover, the SAXS data of free LeeH is better explained by the structure of LeeH in the complex than the structure of a canonical B-DNA LeeH (Figure S4). Therefore, at least in the case of LeeH, CT-Ler recognizes pre-existing DNA structural features following an indirect readout mechanism.
The molecular basis of the preference that H-NS displays for some promoter regions has been extensively studied. AT-tracts were initially postulated to be high affinity sites for H-NS and related to the presence of a narrow minor groove [27]. More recently, two short high affinity H-NS sites with an identical sequence, 5'-TCGATATATT-3' were identified in the E. coli proU promoter [28]. Lang et al. proposed that a 10 bp long consensus sequence (tCG(t/a)T(a/t)AATT) [15] acts as a nucleation site for cooperative binding to more extensive regions. In a recent study, a shorter segment of 5–6 nucleotides comprising only A/T nucleotides was found to be over-represented in genomic loci bound by H-NS in E. coli [29]. The interaction of the H-NS CT-domain, including a few residues from the linker region, with a short oligonucleotide was studied by NMR [22]. The authors concluded that a structural anomaly in the DNA associated with a TpA step was crucial for H-NS recognition.
Our results suggest that AT-tracts and wide TpA steps may be simultaneously required by H-NS family proteins. The correct positioning of a compressed and widened minor groove is the specific recognition signal for CT-Ler. Pyrimidine-purine steps tend to widen the minor groove and TpA steps may contribute to its widening, which is required after the AT-tract. However, in the case of Ler, a TpG step was preferred to the TpA step, suggesting that a wide narrow groove after the AT-tract is the true structural requirement.
CT-Ler and CT-H-NS showed similar structural requirements: mutation of Arg114 reduced the affinity of the complex, and introduction of TpA steps in the AT-tract caused a similar decrease in stability. This result is consistent with the fact that Ler targets can also be occupied by H-NS. Ler and H-NS bind to multiple sites. An indirect readout mechanism allows recognition of multiple sequences, if they adopt similar minor groove patterns.
The absence of structural changes between the free and bound forms of CT-Ler (Figure S5) supports a lock and key model for interactions involving the structured CT-domain and may account for the relatively high specificity of Ler, as compared with H-NS where additional interactions outside the CT-domain are comparatively more important. Comparison of constructs containing exclusively the structured region of the CT-domains of Ler and H-NS show that the former has higher affinity for the range of sequences present in a natural segment where both proteins bind. Several features, not present in CT-H-NS, may contribute to the higher stability of the CT-Ler complex. An additional arginine residue (Arg93) combined with the helix dipole provides additional electrostatic interactions, thus stabilizing the CT-Ler complex. While both Ler and H-NS have a conserved tryptophan residue that, in the case of Ler, forms hydrophobic interactions with DNA, CT-Ler presents an additional tryptophan residue in close contact with DNA. The dipoles of both indole rings are oriented with their positive end towards the negatively charged DNA backbone and the side chain NH of Trp94 forms a hydrogen bond with the DNA backbone.
We have determined for the first time the structure of a complex formed by the DNA-binding domain of a member of the H-NS family. Our results highlight the similarities in the DNA recognition mechanisms used by CT-Ler and CT-H-NS but also evidence some differences that may contribute to the differential recognition of some genes by Ler and H-NS.
DNA fragments containing the coding sequence of Ler residues 65–123, 70–116 (CT-Ler) and 3–116 fused to an N-terminal His6-tag were amplified by PCR from EHEC strain 0157:H7 and subcloned into the pHAT2 vector. To overexpress CT-H-NS, DNA encoding this fragment (amino acids 95–137) with six histidine residues tagged at its N terminus was amplified by PCR using the full length H-NS construction [30] as template and then subcloned into the pHAT2 vector. Point mutations were generated using the QuikChange site-directed mutagenesis kit (Stratagene).
Ler fragments 65–123, 70–116 and 3–116 and CT-H-NS were overexpressed in BL21(DE3) cells with overnight incubation at 15°C by induction with 0.5 mM IPTG when an O.D.600 of 0.7 was reached. For 15N and/or 13C isotopic labeling, cells were grown in M9 minimal media containing 15NH4Cl and/or 13C-glucose. For 10% 13C enrichment we used a carbon source consisting of a 1∶10 mixture of 12C-glucose/13C-glucose [31], [32]. Cells were harvested by centrifugation, frozen and resuspended in 20 mM HEPES (pH 8.0), 1 M NaCl, 5 mM imidazol, 5% (v/v) glycerol, treated for 30 min with lysozyme and DNAse and sonicated (6×10 s on ice). After centrifugation, the His-tagged fusion proteins were isolated with Ni-NTA beads (Qiagen) and further purified by size exclusion chromatography on a Superdex 75 column in 20 mM sodium phosphate, 150 mM NaCl, 0.2 mM EDTA, 0.01% (w/v) NaN3 pH 5.7 or 20 mM sodium phosphate, 300 mM NaCl, 0.01% (w/v) NaN3 pH 7.5. The expression and purification procedure for full length H-NS has been previously described [30].
DNA samples were prepared by hybridization of complementary oligonucleotides purchased from Sigma-Aldrich. Quality control was assessed by MALDI-TOF mass spectrometry. Oligonucleotides were mixed in equimolar amounts and annealed by heating to 92°C for 4 min and slowly cooled to room temperature.
Changes in CT-Ler intrinsic fluorescence anisotropy were monitored upon DNA addition. All measurements were recorded on a PTI QuantaMaster spectrophotometer equipped with a peltier cell, using an excitation wavelength of 295 nm to selectively excite CT-Ler tryptophans and emission detection at 344 nm. Fluorescence measurements were performed in 40 mM HEPES (pH 7.5), 60 mM potassium glutamate, 0.01% (w/v) NaN3 at 20°C. More details on data acquisition and equipment settings were previously described [33]. For the initial screening of the -221 to -101 regulatory region of LEE2, the apparent fraction saturation of CT-Ler was used to infer about DNA binding preferences. To measure the affinity of CT-Ler for 15 bp and 10 bp DNA fragments, titrations were performed at least in duplicate. The fitting was performed assuming a 1∶1 binding using the following equations [34]:(1)(2)where A is the observed anisotropy, Af and Ab are the anisotropies of free CT-Ler and the complex respectively, fb is the fraction of bound CT-Ler and Q is the ratio of quantum yields of bound and free forms. Equations 1 and 2 were solved iteratively until the theoretical binding isotherm matched the experimental data. Kd and Ab were considered to be adjustable parameters.
All spectra were acquired at 25°C on 600, 700, 800 or 900 MHz Bruker spectrometers. Data processing and analysis were carried out with NMRPipe [35], NMRViewJ [36], and CARA [37].
NMR spectra for structure determination were recorded on a ∼1 mM sample containing a 1∶1 complex of uniformly 13C- and 15N-labeled CT-Ler and unlabeled DNA in 20 mM sodium phosphate (pH 5.7), 150 mM NaCl, 0.2 mM EDTA and 0.01% (w/v) NaN3. Backbone and aliphatic assignments of free and DNA-bound CT-Ler were obtained by standard methods. Aromatic resonances were assigned using 2D 1H-13C-edited-NOESY optimized for aromatic resonances. Stereospecific assignments of Val and Leu methyl groups were obtained from a constant time 1H-13C-HSQC on a 10% 13C-labeled protein sample [31]. Non-exchangeable protons of the LeeH duplex bound to CT-Ler were assigned using 2D F1,F2-13C-filtered TOCSY and NOESY spectra in D2O [38]. Exchangeable protons and H2 protons were assigned from 2D F1,F2-15N/13C-filtered NOESY spectrum in H2O [39]. Free DNA resonances were assigned using 2D DQF-COSY, TOCSY and 2D NOESY spectra. Proton chemical shifts were referenced using 4,4-dimethyl-4-silapentane-1-sulfonic acid (DSS) as an internal standard, whereas 15N and 13C chemical shifts were indirectly referenced. Chemical shift assignments have been deposited in the BioMagResBank database under BMRB accession number 17729.
Protein distance restraints were obtained from 2D 1H-13C-edited NOESY (aromatic optimized in D2O), 3D 1H-15N-edited NOESY-HSQC and two 3D 1H-13C-edited NOESY-HSQC (in H2O and in D2O) experiments with a mixing time of 120 ms. Data were automatically assigned and the NOE distance restraints were obtained iteratively using the Unio'08/CYANA 2.1 suite program [40], [41] and manually inspected. The distance restraints for the DNA in complex with CT-Ler were obtained measuring initial NOE build-up rates from 2D F1,F2-15N/13C-filtered NOESY spectra recorded with mixing time of 50, 75, 100 and 150 ms. Intermolecular NOEs were detected using a combination of 2D NOESY, 2D F1,F2-13C-filtered NOESY and 2D F2-13C-filtered NOESY experiments, together with 3D F1-13C,15N-filtered, [F2] 13C-edited 3D NOESY spectrum [42]. Additional intermolecular NOEs were obtained by analyzing the 3D 15N-edited and 13C-edited NOESY spectra.
Protein backbone dihedral angle restraints were derived using a combination of TALOS [43] and quantitative analysis of 3JHNHα obtained from a 3D HNHA spectrum [44]. Restraints on side chain angle and stereospecific assignments of Hβ proton resonances were based on 3JNHβ couplings, obtained from a 3D HNHB spectrum, in combination with observed intraresidual NOEs using the HABAS routine of the CYANA 2.1 program [45].
1H-15N HSQC spectra for analysis of the interaction of 15N-labeled CT-H-NS (100 µM) with dsDNA were obtained at 25°C in 20 mM sodium phosphate (pH 5.7), 150 mM NaCl, 0.2 mM EDTA and 0.01% (w/v) NaN3.
The structure of CT-Ler was determined by simulated annealing using the torsion angle dynamic simulation program CYANA 2.1 [45] and further water refinement with CNS 1.2.1 [46], [47]. Protein structure calculation was based on Unio'08/CYANA-generated upper distances, 3JHNHα/3JNHHβ couplings, and TALOS-driven dihedral angle restraints. Based on H/D exchange experiments, backbone NOE pattern and 13Cα/13Cβ chemical shifts, hydrogen bond restraints were also used in the structure calculation. An ensemble of 100 protein structures was generated and the 20 lowest energy conformers were docked onto a B-DNA.
The observed overlap and broadening of DNA resonances hampered the complete quantitative analysis of NOESY spectra for bound DNA. Only a set of 282 well resolved cross-peaks were converted into distances using initial build-up rates and reference to the cytosine H5-H6 cross-peaks. Upper and lower limits were defined as ± 20% of the calculated distances. The structure of LeeH was fixed as B-DNA and further energy-refined using miniCarlo [48] followed by a 20 ps molecular dynamics refinement in explicit solvent using the Amber force field [49] and including NOE-derived distance restraints. To preserve the helical conformation of DNA, weak planarity restraints were also introduced. The DNA backbone was constrained to a range typical of B-form and all glycosidic angles were restrained as anti. Hydrogen bond restraints were used for all base pairs in which the imino proton was observed. The complex structure was generated employing 30 iNOEs, supplemented with highly ambiguous intermolecular restraints (AIRs) that were driven from the mapped binding interfaces. A total of 22 intermolecular NOE restraints were simultaneously assigned to the two symmetry-related protons in the AATT central region of the DNA and used as ambiguous restraints. HADDOCK 2.0 [50] was used to generated 2000 structures by rigid docking energy minimization, and 400 structures with the lowest energy were selected for semi-flexible refinement process. These 400 structures were finally refined in explicit water including all experimental restraints. Structures were then ranked using the energy-based HADDOCK scoring function (sum of intermolecular electrostatic, van der Waals, desolvation and AIR energies) and NOE energy term. The quality of these structures was evaluated in terms of the violations to the NOE data and the value defining the agreement to SAXS curve. A final ensemble of 20 structures was obtained by re-scoring the pool of 400 structures using the following scoring function.(3)(4)where and correspond to the root mean squared deviations with respect to the best possible value in and Ei respectively. Coordinates of the final ensemble were deposited in the Brookhaven Protein Data Bank under the accession number 2lev.
Minor groove geometry and helical parameters were analyzed using w3DNA [51]. Electrostatic potentials were obtained at physiological ionic strength using DelPhi [52].
SAXS data for LeeH and the CT-Ler/LeeH complex were collected on a MAR345 image plate detector at the X33 European Molecular Biology Laboratory (DESY, Hamburg, Germany) [53]. The scattering patterns were measured at 25°C for 2 min at sample concentrations of 4.6 and 2.7 mg/ml and 6.6 and 3.3 mg/ml for LeeH and CT-Ler/LeeH, respectively. A momentum transfer range of 0.018< s <0.62 Å−1 was measured. Repetitive measurements indicated that samples did not present radiation damage. Buffer subtraction and the estimation of the radius of gyration, Rg, and the forward scattering, I(0), through Guinier's approach were performed with PRIMUS [54]. The scattering profile of LeeH was obtained from merging curves at both concentrations. For CT-Ler/LeeH, SAXS profiles at both concentrations were virtually equivalent and only data from the highest concentrated sample were used for further analysis. Using Guinier's approach, the radii of gyration of LeeH and CT-Ler/LeeH were estimated to be 15.6±0.1 and 18.2±0.1 Å, respectively. All data manipulations were performed with the program PRIMUS. Using a bovine serum albumin sample (3.3 mg/ml), an estimated molecular weight of 18 kDa was obtained for CT-Ler/LeeH (theoretical MW of 16.3 kDa), thereby indicating the presence of a monomeric particle in solution. The agreement of the SAXS curve to various three-dimensional models was quantified with the program CRYSOL [55] using a momentum transfer range of 0.018< s <0.40 Å−1.
The DNA fragment used in this assay (LEE2 positions −225 to +121) was obtained by PCR amplification from EHEC strain 0157:H7. The indicated concentrations of PCR-generated DNA and H-NS or Ler proteins were mixed in a total volume of 20 μl of 15 mM sodium phosphate, 100 mM NaCl, 0.01% (w/v) NaN3 pH 7.5. 1 mM tris(2-carboxyethyl)-phosphine (TCEP) was included for samples containing full length H-NS. After 20 min of incubation at room temperature, glycerol was added to 10% (w/v) final concentration and the reaction mixtures were electrophoresed on either 1.5% agarose or 7% polyacrylamide gels in 0.5x Tris-borate-EDTA buffer. The DNA bands were stained with ethidium bromide.
|
10.1371/journal.ppat.1005424 | Oncogenic Herpesvirus Utilizes Stress-Induced Cell Cycle Checkpoints for Efficient Lytic Replication | Kaposi’s sarcoma herpesvirus (KSHV) causes Kaposi’s sarcoma and certain lymphoproliferative malignancies. Latent infection is established in the majority of tumor cells, whereas lytic replication is reactivated in a small fraction of cells, which is important for both virus spread and disease progression. A siRNA screen for novel regulators of KSHV reactivation identified the E3 ubiquitin ligase MDM2 as a negative regulator of viral reactivation. Depletion of MDM2, a repressor of p53, favored efficient activation of the viral lytic transcription program and viral reactivation. During lytic replication cells activated a p53 response, accumulated DNA damage and arrested at G2-phase. Depletion of p21, a p53 target gene, restored cell cycle progression and thereby impaired the virus reactivation cascade delaying the onset of virus replication induced cytopathic effect. Herpesviruses are known to reactivate in response to different kinds of stress, and our study now highlights the molecular events in the stressed host cell that KSHV has evolved to utilize to ensure efficient viral lytic replication.
| Herpesviruses are known to wake up and reactivate in response to different kinds of stress. Our study now highlights the key molecular host cell events that KSHV has evolved to utilize for efficient viral lytic replication: the activation of p53 and upregulation of p21, which slows down the cell cycle, but promotes viral replication and transcription of viral lytic genes. Mutations in TP53 gene are rarely found in KSHV-associated malignancies. Therefore, our work now provides a mechanistic explanation as to why the virus has evolved to retain p53.
| Kaposi’s sarcoma-associated herpesvirus (KSHV) is a human tumor virus in the family of gamma2-herpesviruses. KSHV is the etiologic agent of Kaposi’s sarcoma (KS) and other KSHV-associated lymphoproliferative diseases such as primary effusion lymphoma (PEL) [1,2]. KSHV genome consists of linear double-stranded DNA (dsDNA), and like other herpesviruses, the virus displays two modes of infection in the infected cells, the latent and lytic replication phase. Upon entry into the host cell nucleus, the linear dsDNA genome circularizes forming a non-integrated viral episome that persists as multiple copies in the latently infected cells [3]. The latent infection (latency) provides an immunologically silent mode of persistence, whereas the lytic replication phase allows replication and production of new virions to be shed and transmitted to new cells and hosts.
The switch between the latency and lytic replication (virus reactivation) is a critical step in viral pathogenesis. Although the KSHV-associated tumors typically show low level of virus reactivation [4,5], epidemiological studies support the importance of lytic replication in the initiation and progression of KS [6–8]. Despite of active research, the regulation of viral reactivation is not completely understood. However, significant advances have been made in recent years, and the reported mechanisms of KSHV reactivation involve hypoxia [9–11], reactive oxygen species [12], inflammation [13–15], activation of cellular kinases [16–20] and epigenetic mechanisms [21–25]. KSHV reactivation can also be chemically induced e.g. with certain kinase agonists (TPA) and chemical inhibitors affecting histone acetylation (HDAC inhibitors) or DNA methylation (reviewed in [26]).
Recent studies by several groups have demonstrated that the intracellular viral genome has chromatin structures similar to that of the host chromosome (reviewed in [22]). The latent KSHV genome is epigenetically modified with methylation at CpG dinucleotides as well as mutually exclusive activating and repressive histone modifications [27–29]. The bivalent chromatin structure represents a poised state of repression during viral latency, which can be rapidly reversed once the lytic cycle is induced, and enables the virus to fine-tune its gene expression patterns in response to changes in virus infected cells. Further support for the importance of epigenetic regulation in the switch from latency to lytic replication was provided by the demonstration of the cohesin subunits as major repressors of KSHV lytic gene activation suggesting that cohesins could be a direct target of butyrate-mediated lytic induction [30]. Other recently identified epigenetic regulators of KSHV reactivation include the H3K27me3 histone methyltransferase of the Polycomb group proteins, EZH2 [28], HDAC class I and II [25], and the histone demethylase JMJD2A [31].
To discover novel mechanisms regulating KSHV reactivation we designed and performed a small interfering RNA (siRNA) screen using a library of siRNAs specific for human genes involved in epigenetic processes. In this screen we assessed which epigenetic enzymes help the virus to maintain latency. We identify MDM2, an E3 ubiquitin ligase, as a novel modulator whose depletion by siRNA accelerates KSHV reactivation. We also show that MDM2 down-regulation leads to subsequent activation of p53 and p21 as well as induction of a p21-dependent cell cycle arrest, which are required for the induction of efficient viral lytic replication cascade.
To identify novel regulators of KSHV reactivation we performed a siRNA screen using a custom-made siRNA library targeting 615 human genes with Gene Ontology (GO) annotations related to epigenetics, chromatin remodeling/maintenance, and co-regulatory functions, and consisting of two independent siRNAs for each gene [32]. Before reverse transfection, the siRNAs, mixed with transfection reagent and components of the extracellular matrix, were spotted onto a microplate-sized array plate and analyzed by the cell-spot microarray technique (CSMA; [33]. For the screen, we used SLK cells stably infected with a recombinant KSHV (rKSHV.219) which during latency constitutively expresses green fluorescent protein (GFP) under the control of the cellular EF-1α promoter, and upon reactivation the red fluorescent protein (RFP) from the promoter of the viral early lytic gene PAN [34]. The workflow of the primary screen is depicted in S1A Fig (for details see S1 Methods).
Among the siRNAs that enhanced reactivation, one of the strongest and the most reproducible ones were the siRNAs against MDM2 (siMDM2). The depletion of MDM2 led to RFP levels that were three SDs above the mean RFP value for the screen (S1 Fig). In this study, we pursued validation and characterization of the role of MDM2 in KSHV lytic replication.
To validate the results of the screen, we tested two additional MDM2 siRNAs targeting different, non-overlapping sequences of the MDM2 transcript. The siRNAs were pooled and transfected into iSLK.219 cells, a recently developed cell clone derived from SLK cells stably infected with rKSHV.219 [35]. These cells contain an exogenous copy of the viral gene RTA that can be induced by doxycycline. Addition of low doses of doxycycline results in efficient RTA synthesis (S2A Fig) that triggers virus reactivation in approximately 5–15% of cells at 24 hours post induction (hpi) (S2B and S2C Fig). This can be substantially augmented by the combination of doxycycline (Dox) with TPA (TPA/Dox) or NaB (NaB/Dox) which increase the fraction of RFP-positive cells to 80–90% within 24 hours of treatment (S2B and S2C Fig).
In the absence of Dox, treatments with TPA or NaB alone do not result in lytic reactivation (S2B and S2C Fig), thereby providing an invaluable control to monitor possible pleiotropic effects of the two compounds.
To assess the effect of MDM2 depletion in the iSLK.219 cells [35] we incubated the siRNA transfected cells for 72 hours (h), and treated the cells by suboptimal doses of doxycycline that alone produce no or very low level of RFP induction. After fixation, cells were imaged and analyzed for RFP expression using automated high-content fluorescence microscopy. The depletion efficiency of the MDM2 siRNAs was 62% as monitored by quantitative real time PCR (qRT-PCR). As shown in Fig 1A and 1B, and confirming the results of the screen, depletion of MDM2 led to a 3-fold increase in the number of RFP—positive cells compared to the control siRNA transfected cells.
To further validate the results of the screen in cells that are naturally infected by KSHV, we silenced MDM2 expression in BC-3 cells, a patient-derived primary effusion lymphoma (PEL) cell line[36]. BC-3 cells were transduced with lentiviruses expressing shMDM2, or shCtrl, and 72 h later virus reactivation was induced by TPA treatment. We first assessed whether the depletion of MDM2 induced spontaneous reactivation by analyzing the mRNA levels of the lytic genes ORF50 (encoding for RTA; immediate early), ORF57 (encoding for MTA; delayed early), vGPCR (intermediate) and K8.1 (late) in non-induced BC-3 cells. Whole cell extracts were collected at indicated times and viral lytic transcripts were analyzed by qRT-PCR (Fig 1C and 1D). Depletion of MDM2 induced the expression of the viral transcripts by 3.2 to 5.6 fold compared to controls (Fig 1C) indicating spontaneous viral reactivation. Similarly, during TPA-induction MDM2 depletion led to a rapid increase of all tested viral lytic transcripts that ranged from 2.5 (ORF50) up to 15 fold (ORF57) at 4 hpi (Fig 1D). Interestingly, over time, the magnitude of the up-regulation gradually decreased, reaching control levels at 48 hpi. Thus, silencing of MDM2 in PEL cells led to spontaneous reactivation and accelerated the kinetics of lytic gene expression, without altering their maximal levels during the time course studied.
MDM2 is the major regulator of the tumor suppressor p53 and targets p53 to proteasomal degradation through its ubiquitin E3 ligase activity [37]. While p53 is inactivated during KSHV latency [38–45] its potential role during the lytic phase of the virus replication cycle has not been investigated.
The possibility that the transcription factor p53 could be required for efficient viral lytic gene expression prompted us to conduct a comprehensive, unbiased analysis of the p53 chromatin binding sites during the reactivation of BC-3 cells. To this end, we carried out chromatin immunoprecipitation using antibodies against p53, followed by genome-wide deep sequencing (ChIP-seq) analysis of the associated DNA. Non-specific IgG antibody was used as a negative control. BC-3 cells were subjected to ChIP-seq after TPA induction for 0 or 24 h. As a positive control for detection of the canonical p53 targets, non-induced BC-3 cells were treated with Nutlin-3 (here referred to as Nutlin), a small molecule that binds and inhibits MDM2, stabilizing p53 and inducing a potent p53 response.
Compared to the DMSO-treated cells, the TPA treatment induced a global activation of p53 that could be visualized by averaging the sequencing signal over the whole cellular genome (S3A Fig). Specific sequencing signal was found at regions preceding known p53-target genes, such as MDM2, CDKN1A (p21Cip1), P53R2 and PAG608 (Fig 2A). For some genes, the sequencing peak obtained from TPA-treated cells was comparable to that of Nutlin treated cells (Fig 1A, p53R2; S3B Fig, red arrow heads). Pathway analysis indicated that the genes close to the 300 most significant p53 binding sites (p<0.01) were involved in Cell cycle arrest, Apoptosis, DNA repair and p53 regulation (i.e. MDM2) (S3C Fig). 94 of the 99 peaks with the lowest p-values from the TPA sample overlapped with the peaks identified from Nutlin-treated cells (S3D Fig), confirming that in BC-3 cells TPA treatment induced a bona fide p53 response.
Notably, the ChIP-seq analysis of TPA- or Nutlin-treated cells did not produce statistically significant p53 binding events on the viral genome. The sequencing signal from TPA- or Nutlin- treated cells (Fig 2B red and blue lines, respectively) was comparable to the background noise obtained with the nonspecific IgG controls (Fig 2B, light blue and green lines, respectively) despite the high copy number of viral genomes per cell in BC-3 and PEL cells in general [46]. The absence of p53-binding sites on the viral genome prompted us to perform a quality control to confirm that the p53 binding sites identified in the cellular DNA corresponded to the expected consensus sequences [47]. A de novo identification of the p53 DNA-binding sequence returned its known consensus sequence (Fig 2C) [47], for both TPA and Nutlin treated cells.
Thus, if p53 was involved in virus reactivation, this effect occurred through the activation of cellular rather the viral genes.
Of the different pathways induced by p53, we investigated the role of cell cycle arrest, a phenomenon known to occur during the lytic phase of herpesviruses and other DNA viruses [48]. Of particular interest was p21Cip1 (p21), a strong inhibitor of cell cycle progression [49], which was also identified in our ChIP-seq analysis of TPA induced cells (Fig 2A and S3B and S3C Fig).
To validate the results of the ChIPseq experiment, we determined the level of p53 and its target p21 by immunoblotting of cell extracts collected at indicated times after reactivation with TPA or Nutlin treatment (Fig 2D). Compared to the DMSO controls (0 h), induction with TPA caused a small increase in the levels of p53 and a strong induction of p21 (Fig 2D).
The up-regulation of p21 was fast and could be detected already 4 h after TPA treatment in BC-3 cells (S4E Fig). The p21 levels also increased in iSLK.219 cells treated with TPA/Dox or NaB/Dox (Fig 2E). To test if this effect was due to viral lytic gene expression, we compared by WB the levels of p21 in iSLK.219 cells treated with DMSO (non induced), Dox, TPA or TPA/Dox (S3E Fig). Viral lytic gene expression was monitored with antibodies against MTA (ORF57, early lytic gene). In iSLK.219 cells, TPA treatment is not sufficient to trigger lytic reactivation (see also S2B and S2C Fig). However, the levels of p21 in cells treated with TPA alone for 4h were comparable to those in cells treated with TPA/Dox that results in lytic reactivation (S3E Fig). No increase in p21 levels was observed after 4h treatment with Dox that leads to the synthesis of RTA (ORF50) from a plasmid stably maintained in the iSLK.219 cells. Four hours after Dox treatment the synthesis of lytic genes downstream of RTA is still not detectable by WB. Thus, the fast increase in p21 levels was not due to viral lytic gene expression but appeared to be an intrinsic effect of the TPA treatment. To confirm this, we incubated non-infected SLK cells with similar doses of TPA/Dox or NaB/Dox and monitored by WB the levels of p21 over time. Nutlin treatment was used as a positive control. Although with slower kinetics, TPA and NaB induced an increase of p21 levels by 12 h (S3F Fig). These results are consistent with numerous earlier reports obtained in other cell types [50–53].
To test whether the induction of p21 was p53-dependent we depleted p53 in iSLK.219 cells and then monitored the levels of p21 after treatment with DMSO, Nutlin or TPA/Dox. At 48 h after transduction with lentiviruses expressing shRNA against p53 (sh-p53) or nonspecific non-targeting controls (sh-Ctrl), iSLK.219 were treated with indicated compounds for 4 h and the levels of p53 and p21 were monitored by WB of whole cell extracts (Fig 2F). The specificity of the p21 antibody was confirmed by including iSLK.219 cells treated with siRNAs against p21 (si-p21). Compared with DMSO controls, treatment with Nutlin increased the levels of p53 and p21. As in our previous results in BC-3 cells (Fig 2D), the levels of p21 in Nutlin and TPA/Dox treated cells were comparable despite the large difference in the levels of p53 (Fig 2F, Nutlin and TPA/Dox, sh-Ctrl lanes). Depletion of p53 in DMSO- and Nutlin-treated cells was very efficient and abrogated the induction of p21 (Fig 2F, DMSO and Nutlin). Upon TPA/Dox treatment, however, p21 was induced even in the absence of detectable p53 (Fig 2F, TPA/Dox, sh-p53 lanes), indicating that upon TPA/Dox induction the levels of p21 are regulated by both p53-dependent and -independent mechanisms.
To test if an increase in p53/p21 levels would favor virus lytic gene expression we induced iSLK.219 cells with Dox or TPA/Dox in the presence of Nutlin for 24 h, and measured the levels of RFP (virus reactivation) by automated microscopy. In each experiment, we monitored the stabilization of p53 by Nutlin using immunofluorescence (S4A Fig). In parallel experiments using non-induced iSLK.219 cells, the decrease in the number of cells after Nutlin treatment was used to measure the proliferation arrest in response to p53 stabilization (S4B Fig). Consistent with the results obtained by depletion of MDM2, inhibition of MDM2 by Nutlin led to a two-fold increase in virus reactivation (Fig 3A and 3B).
Conversely, depletion of p53 using RNAi decreased RFP expression induced by TPA/Dox in iSLK.219 cells (Fig 3C and 3D). In these experiments the silencing of p53 was monitored by WB analysis (Fig 3D, insert). Similar results were obtained in TPA and NaB treated BC-3 cells, where the levels of lytic gene expression were monitored by qRT-PCR and WB (S4C–S4E Fig).
We then addressed the relevance of p21 in mediating the lytic reactivation. Similarly to the depletion of p53, siRNA-mediated depletion of p21 in TPA/Dox induced iSLK.219 cells decreased fluorescence intensity of RFP (Fig 3E and 3F) and other viral lytic genes (Fig 3G). To address the role of p21 for efficient viral lytic replication in a biologically relevant KSHV-infection system, we chose to use two PEL cell lines, TPA-treated BC-3 and TREx BCBL1-Rta (here referred to as BCBL1RTA) cells that can be induced to lytic replication through doxycycline-inducible expression of RTA [54]. We stably depleted p21 in both cell lines by transducing them with lentiviruses expressing a control unspecific shRNA (sh-Ctrl) or two different sh-p21 constructs (sh1-p21, sh2-p21) followed by puromycin selection. Twelve days after selection, both BC-3 and BCBL1RTA cell lines expressing the sh-p21 constructs had growth kinetics undistinguishable from their respective sh-Ctrl transduced cells (S5A and S5B Fig).
As shown for BC-3 and iSLK.219 cells (Fig 2D and 2E), the levels of p21 mRNA also increased in these new cell lines upon TPA (Fig 4A, BC-3 sh-Ctrl) or doxycycline (Fig 4B, BCBL1RTA shCtrl) treatments.
The silencing efficiencies of p21 were monitored by qRT-PCR (S5C and S5D Fig) and WB (Fig 4C). Stable depletions of p21 had little or no effect on the expression levels of the latent gene ORF73 (Fig 4D) at 24h after TPA or Dox induction in BC-3 or BCBL1RTA cells, respectively (Fig 4D). However, the expression levels of early, intermediate and late lytic genes were significantly decreased at 24 h (Fig 4E) or 48 h (Fig 4F) after induction of reactivation in BC-3 and BCBL1RTA, respectively. Of the two cell lines used, BC-3 cells were more sensitive to the depletion of p21 than the respective BCBL1RTA cells (compare Fig 4E and 4F).
Thus the expression of p53 and p21 favored efficient viral lytic gene expression in both TPA-treated and RTA-overexpressing naturally KSHV-infected cells. Moreover, the decreased lytic gene expression in the p21-depleted cells resulted in a reproducible and significant delay in the onset of cytopathic effect (cell lysis), a known consequence of efficient virus replication (Fig 4G and 4H). Interestingly, despite the significant reduction of viral lytic gene expression, we saw only a modest decrease (up to approximately 30%) in the amount of infectious viruses released in the supernatant of p21-depleted BCBL1RTA cells compared with their respective shCtrl expressing cells after reactivation by doxycycline for 24h (S5E and S5F Fig).
p21 can arrest the cell cycle at G1/S or G2 phase by inhibiting the cyclin/Cdk complexes [49]. To assess how these properties of p21 contribute to the KSHV lytic reactivation, we developed an automated, image-based assay in which we used iSLK.219 cells and monitored viral reactivation and cell cycle progression at single-cell level. The expression of RFP served as a marker of virus reactivation, while cell cycle progression was monitored by immunofluorescence analysis using antibodies against histone H3 phosphorylated on serine 10 (pH3-S10). In dividing cells that enter the G2 phase, this antibody gives a punctate signal that labels the sites of chromatin condensation. As cells progress towards prometaphase, chromatin condensation continues and the levels of pH3-S10 increase proportionally, reaching a maximum at metaphase and anaphase, when the cellular DNA is tightly packed into chromosomes [55]. At the end of mitosis, histone H3 is rapidly dephosphorylated, and is undetectable in G1 and S phases. These stages could be easily distinguished in the iSLK.219 cells with automated fluorescence microscopy (Fig 5A). Using the fluorescence-intensity of the pH3-S10 staining and automated image analysis, cells were assigned to G1/S (no pH3-S10 signal, panel 'a'), G2 (weak and punctate pH3-S10 signal, panels 'b' and 'c', respectively) or M phase (bright pH3-S10 signal; panels 'd-f') (Fig 4A and S6A Fig). To further demonstrate that cells displaying pH3-S10 fluorescence are indeed in G2 or M-phase, we co-stained non-induced iSLK.219 cells with cyclin B1, another cellular marker of G2 and M-phase. As expected, the two markers were perfectly correlated (S6B Fig). Cells in G2 displayed a weak and punctate pH3-S10 signal and cytoplasmic cyclin B1, while cells in M-phase had bright and diffused pH3 S10 fluorescence and nuclear cyclin B1 (S6B Fig).
To test the accuracy of our image analysis method, we used as controls Nutlin, that arrests cells in G1/S, and Etoposide, an inhibitor of cellular topoisomerase-II that causes DNA damage and a G2/M cell cycle arrest (S6C and S6D Fig). Compared to DMSO controls, both Nutlin and Etoposide treatments strongly decreased the number of M phase cells (S6D Fig). Consistent with a robust G1/S arrest, Nutlin also decreased the number of cells in G2, while Etoposide increased the fraction of cells in G2 by more than six fold compared to DMSO treated samples (S6C and S6D Fig). Treatment with TPA/Dox led to an increase in the fraction of cells in G2, suggesting that during reactivation iSLK.219 cells arrest in G2 (S6C and S6D Fig, TPA/Dox). In parallel experiments, we excluded cross-talk between the RFP and Alexa488 fluorescence signal used to detect pH3 S10 (S7A Fig). Using similar drug treatments, the image analysis based cell cycle measurements were confirmed by traditional FACS analysis using propidium iodide to stain DNA (S8 Fig). Also in this case, the induction of iSLK.219 cells with TPA/Dox induced an increase in the fraction of S- and G2/M-phase cells, similar to the effect of etoposide treatments (S8B–S8G Fig). In these experiments the settings of the FACS detection were adjusted such that the RFP fluorescence did not contribute to the detection of PI (S8D Fig). Although very sensitive and in agreement with the results of the image analysis, the PI FACS analysis method could not distinguish between cells in G2- or in M-phase.
Based on the pH3 S10 imaging method, about 85% of non-induced iSLK.219 cells were at G1/S phase, 10% at G2 and 5% at M phase (Fig 5B and 5C, DMSO). At 24 h after TPA/Dox treatment, 62% of cells expressed RFP (compare with S2C and S2D Fig). The number of cells in M phase decreased from 4.6% to 1.5%, and the majority of reactivated cells (77% of all cells and 86% of RPF positive cells) displayed the weak pH3-S10 signal indicating an arrest in G2 phase (Fig 5B and 5C, TPA/Dox). To confirm that reactivated cells arrest in G2 we repeated the experiment using antibodies against cyclin B1. Indeed, compared to DMSO controls (S7B Fig), the vast majority of RFP positive cells also contained cyclin B1 (S7C Fig).
In the absence of Dox, TPA treatment did not induce G2 arrest, but only slightly decreased the number of M phase cells (Fig 5C and 5D, TPA). Thus viral lytic gene expression coincided with the G2 arrest.
To test whether depletion of p21 affected the G2 arrest observed during KSHV reactivation we silenced p21 for 48 h and subsequently monitored the levels of pH3-S10 in iSLK.219 cells treated with TPA/Dox for 24 h (Fig 5E–5G). As in previous experiments, induction with TPA/Dox induced a G2 arrest in cells treated with si-Ctrl (Fig 5E–5G, si-Ctrl). However, depletion of p21 in the TPA/Dox treated cells resulted in a significant decrease in the fraction of cells in G2 and a correspondent increase in the number of mitotic profiles to a level approaching the non-induced cells (Fig 5E–5G).
Upon lytic reactivation, the levels of p21 rise within the first 4 h, and increasing its levels with Nutlin did result in a complete G1/S arrest in latently infected cells. Why didn't reactivated cells then arrest in G1/S? If a mechanism existed to inactivate the G1/S checkpoint during the lytic phase, then reactivated cells would move on to the S-phase and reach the G2. To test this possibility, we pre-treated non-induced iSLK.219 cells with Nutlin for 18 h to induce a complete G1/S arrest, and then triggered virus reactivation by TPA/Dox in the presence of Nutlin. DMSO served as a control (Fig 6A). Cells were then fixed 18 h after reactivation and, the cell cycle progression was monitored by image analysis of pH3-S10 as described before.
As expected, after Nutlin treatment the non-induced cells arrested in G1/S (Fig 6B and 6C, Nutlin/DMSO). Again, TPA/Dox treatment arrested most of the cells in G2 (Fig 6B and 6C, DMSO/TPA/Dox). Strikingly, despite the Nutlin pre-treatment, the addition of TPA/Dox led to an increase in the fraction of cells in G2 while those in G1/S phase decreased (Fig 6B and 6C, Nutlin/TPA/Dox). This was not observed in the non-infected SLK cells used as a control (S9 Fig). In the absence of Dox, TPA treatment alone does not reactivate the virus in iSLK.219 cells and did not allow the cells to overcome the Nutlin-induced G1/S arrest (Fig 6C, Nutlin/TPA). Thus, during virus reactivation the G1/S checkpoint is inactivated while the G2/M checkpoint remains active.
The G2 arrest in lytic cells suggested activation of a DNA damage response (DDR). We and others have previously reported DNA damage induced by KSHV [56]. We therefore next assessed whether TPA/Dox or NaB/Dox activated DDR in iSLK.219 cells. DDR was monitored by immunofluorescence using antibodies against the phosphorylated forms of the Ataxia telangiectasia mutated (pATM, S1981) kinase, histone H2AX (γ-H2AX), and checkpoint kinase 1 (pChk1, S317) proteins (Fig 6D–6F). Non-induced iSLK.219 cells treated with Etoposide (10μM) served as positive controls (Fig 6F, dashed red line). Based on automated quantitative immunofluorescence analyses TPA/Dox-induced RFP-positive cells expressed robustly all DDR markers to a similar degree as the Etoposide-treated cells (Fig 6D–6F). The kinetics of DDR response was slower in the NaB/Dox treated cells but became comparable to those of TPA/Dox treated cells at 48 hpi (Fig 6F). Treatments with TPA or NaB in the absence of Dox, which is not sufficient to induce virus reactivation in this cell line, did not result in detectable DDR at 24 h (Fig 6G). Thus, reactivated cells accumulated DNA damage and activated DDR. Similar to the G2 arrest, the DDR was not due to pleiotropic effects of the TPA of NaB treatments but required viral lytic gene expression.
When cells accumulate DNA damage, cell cycle arrest is necessary to allow DNA repair and prevent the propagation of damaged DNA to dividing cells, which would otherwise induce cell death during the M phase. Given the extent of DDR accumulated during the lytic phase, we hypothesized that in iSLK.219 the G2 arrest could prevent premature cell death.
The G2 arrest is enforced by the effectors of DDR, such as the kinases Chk1 and 2. We therefore tested the ability of different inhibitors of DDR (S10A Fig) to impair the G2 arrest and cause death in cells reactivated by TPA/Dox (lytic reactivation). The assay was optimized using increasing concentrations of Etoposide and caffeine, a broad inhibitor of ATM, ATR and other cellular kinases. The extent of DNA damage and cell cycle progression were monitored by immunofluorescence using antibodies against γH2AX and pH3-S10, respectively. Cells were treated with caffeine 1 h prior to addition of Etoposide, fixed 48 h later and processed for image analysis. Through inhibition of the upstream signals of the DDR (ATM/ATR), caffeine was very effective in inhibiting the Etoposide-induced G2 arrest and induced ‘mitotic catastrophe’ (S10B and S10C Fig; compare insets 1,2 and 3).
Unlike caffeine, a specific inhibitor of ATM (KU-55933) was not effective in overcoming the G2 arrest caused by Etoposide (S10D Fig). A specific inhibitor of ATR, VE-821 (S10D Fig), and inhibitors of Chk1/2 (ADZ-7762) or Chk1 (MK-8776), did restore the number of M phase cells when low concentrations of Etoposide were used (S10D and S10E Fig).
Having established the appropriate experimental conditions and the most effective concentrations for each of the inhibitors, we tested their effect in reactivated iSLK.219 cells. Drugs were added 1 h prior reactivation, and maintained in the culture medium throughout the experiment. Cells were fixed 48h after reactivation (TPA/Dox). The extent of viral lytic gene expression and of cell cycle progression was monitored by image analysis of indicated viral lytic genes and pH3- S10, respectively. As shown in Fig 7A, only caffeine partially inhibited the viral lytic gene expression. Consistently, none of the drugs was able to overcome the G2 arrest and restore progression to M-phase (Fig 7B). Our results are consistent with other reports in which the ability of these compounds to sensitize cancer cells to radiation-induced cell death failed if the cells had active p53 and p21 [57].
Through silencing the E3 ubiquitin ligase MDM2 in a siRNA screen we here identify the p53/p21 axis as an important positive regulator of viral reactivation, and demonstrate that cellular stress, an inducer of herpesvirus reactivation, favors KSHV lytic replication. Our data further show that lytic replication leads to severe and sustained DNA damage response and lead to a prominent G2 arrest, indicative of a virus-activated cellular checkpoint. Depletions of p21 in iSLK.219 cells reactivated by TPA/Dox treatment significantly decreased the kinetics of early (RFP, MTA), intermediate (ORF59) and late (K8.1) lytic gene expression and alleviated the G2 cell cycle arrest. Similar results were obtained in biologically relevant PEL cells, BC-3 and BCBL-1. Interestingly, despite the significant reduction in viral lytic gene expression, we observed only a modest decrease in the amounts of infectious virions produced from BCBL1RTA cells after p21 depletion. The modest decrease, however, could be misleading, as compared to p21-depleted cells, the sh-Ctrl expressing cells lysed significantly faster, which could affect the infectivity of the released viruses.
Other DNA viruses, including human papilloma virus, adeno and parvoviruses, have been shown to induce a G2/M arrest during viral DNA amplification [48]. Why would KSHV then benefit from halting the cell cycle at G2? One possibility is that the virus needs to access cellular factors expressed at the G2 phase which could include factors for recombination, repair etc. [58]. Interestingly, we found that reactivated cells were able to bypass the G1/S arrest induced by treatments with Nutlin prior to reactivation. As the G1/S checkpoint is inactivated, the infected cells progress to the S-phase and eventually arrest in G2. While the detailed mechanism of this inactivation is under investigation, we and other groups have demonstrated that the restoration of p53 activity by Nutlin leads to efficient cell death in latent PEL cells and KSHV-infected endothelial cells [44,56,59,60], but fails to kill reactivated cells [61]. The described inactivation of the G1/S checkpoint in cells undergoing lytic replication now explains why the same treatment was not effective in eliminating the reactivated cells [61].
Another advantage of the G2 arrest could be to avoid factors active in M phase that could impair virus amplification. Similar to cellular DNA, the viral DNA genome in infected cells is complexed with nucleosomes [22]. The viral genome may therefore also undergo some degree of modifications (e.g. condensation) when the cell enters mitosis, which could disturb viral lytic gene expression and DNA replication once the lytic cycle is initiated. While more experiments are required to address this possibility, a recent report showed that KSHV genome can be co-immunoprecipitated using antibodies against pH3-S10 [17], a modification present during chromatin condensation at the onset of mitosis [55,62–65]. Interestingly, depleting the enzymes that are responsible for H3 phosphorylation at S10 led to robust spontaneous reactivation and decreased levels of pH3-S10 on the viral genome [17]. Both p21 and the DNA-damage-activated Chk1, which we found active during reactivation, can inhibit the G2/M transition [66–68]. p21 has been shown to efficiently inhibit CDK1 in response to DNA damage, which was sufficient to cause a permanent G2 arrest [69]. The early induction of p21 described here could therefore inhibit Cdk1, support the G2 arrest and increase the efficiency of virus reactivation. Intriguingly and supporting this hypothesis, Cdk1 inhibition by small molecule inhibitors has been shown induce spontaneous KSHV reactivation in PEL cells [70].
Although the activation of p21 seems to be an early event during KSHV reactivation, the activation of a DNA damage response (e.g. pATM, pChk1) observed at later stages of the lytic cycle could reinforce the inhibition of the cell cycle progression [66]. We therefore attempted to inhibit the DDR by small molecule inhibitors, hoping to restore the cell cycle progression and activate cell death responses during M phase, a process known as ‘mitotic catastrophe’ [71,72]. This strategy has been successfully used to sensitize cancer cells to radiation-induced DNA damage [73]. However, the DDR inhibitors have not been effective in cells with functional p53/p21 axis, because cell death is prevented by the p53-dependent cell cycle arrest [74,75]. Our attempt of using inhibitors of DDR to induce cell death in reactivated cells was also unsuccessful. Similar to other cancer cells, this resistance could be due to p21-induced cell cycle arrest.
The tumor suppressor p53 is often mutated and inactivated in human cancers. However, this is not the case in KSHV-associated malignancies, where p53 mutations are rarely found [39,59,76]. Instead, the virus has co-evolved with wild type p53 and utilizes a large repertoire of mechanisms to bypass its growth restrictive functions during the latency [38–42,44,45,59]. Our work now provides a mechanistic explanation as to why the virus has evolved to retain p53 and p21 activities to support the lytic phase.
Our ChIPSeq cell analyses indicated that p53 does not bind KSHV genome during the reactivation. Work on murine gammaherpesvirus 68 (MHV68) and Epstein Barr virus (EBV), two other gammaherpesviruses, has demonstrated that p53 contributes to the expression of the lytic genes, by stimulating transcription of RTA or RTA and an immediate-early viral Zta gene, respectively [77–79]. Our results now demonstrate a different role for p53 in KSHV reactivation, where this transcription factor is required to enhance viral replication through modulation of the host cell stress responses induced upon viral lytic replication.
Cells, plasmids, antibodies, primers, and standard methods are described in SI Materials and Methods.
Lentiviral expression plasmids pLKO.1-sh-Scr, sh5-MDM2, sh1-p53, sh2-p53, sh1-p21 and sh2-p21 were obtained from Open Biosystems and Biomedicum Functional Genomics Unit (Helsinki, Finland). Lentivirus stocks were produced as described [56]. To establish BC-3 cell lines stably expressing sh-p53, cells were selected with medium containing 3.5 μg/ml puromycin (Sigma) and to establish TREx BCBL1-Rta (here referred as BCBL1RTA) and BC-3 cell lines stably expressing sh-p21, cells were selected with medium containing 2 μg/ml puromycin (Sigma).
The nonspecific siRNAs control (ON-TARGETplus, D-001810-10) and siRNA against p21 (ON-TARGETplus, L-003471-00) were from Dharmacon (SMARTpool siRNA). Cells were reverse-transfected (2000 cells per well) for 48 h in 96-well view plates (Perkinelmer) using RNAiMax (Invitrogen), and reactivated by indicated treatments for 24 h. For more details see supplemental information.
The Primary screen was performed with the cell spot microarray technique [33]. 48 h after reverse transfection, rKSHV-SLK cells were reactivated for 30h, fixed and imaged with an Olympus Scan-R high-content microscope. For more details see S1 Methods.
For immunofluorescence in 96 well imaging plates (Perkin Elmer) cells were fixed in PBS containing 4% paraformaaldehyde, (20 min, room temperature), permeabilized for 10 min in PBS containing 0.1% Triton-X 100 (Sigma). All antibody incubations were performed in PBS containing 2% BSA for 45 min. Alexafluor-conjugated secondary antibodies were from Invitrogen.
For western blot analysis cells were lysed in RIPA buffer containing protease (Thermo scientific; Cat # 88666) and phosphatase inhibitors (Thermo scientific; Cat # 88667) and whole cell extracts were then clarified by centrifugation, mixed with Laemmli buffer and boiled for 5 min. Protein concentration was determined by Bio-Rad protein assay (Bio-Rad) and 20–40 μg of protein were loaded in each lane of a Criterion TGX midi gel 4–15% (Bio-Rad) and transferred to nitrocellulose membranes (Protran nitrocellulose membrane 0,45 um, Perkin Elmer). Membranes were immunostained in TBST buffer containing 5% fat-free milk, and HRP-conjugated secondary antibodies detected with the Enhanced Chemiluminescence kit (Western Bright Sirius, Advansta) on Fuji films. Densitometry analysis was performed using the ImageJ software (National Institutes of Health, USA).
ChIP and sequencing were carried out as described [80] using a monoclonal antibody against p53 (Clone DO-1, GeneSpin) and an Illumina HiSeq 2000 (single 36 bp reads) system. The analysis of the sequencing data was performed as described [81]. For details see S1 Methods.
Caffeine (Sigma) was dissolved in water. All the other drugs were dissolved in DMSO and stored at -20 C. Nutlin-3, Etoposide, KU-55933 and UCN-01 (Sigma); MK-8776 (also called SCH900776, Chemie Tech); VE-821 and AZD-7762 (Selleckchem). All experiments were performed in 96-well imaging plates (PerkinElmer). For details see S1 Methods.
All experiments were performed in 96-well imaging plates (Perkin Elmer) and images acquired using the automated fluorescence microscope Cellinsight (Thermo Scientific). Image analysis was performed with the Cell Profiler 2 software [82]. For the analysis of the pH3 S10 staining we modified a premade imaging pipeline ("Percent positive") available in the Cell Profiler web site (www.cellprofiler.org). After detection of nuclei, the G1/S, G2 and M-phase cells were classified by thresholding the mean fluorescence intensity of pH3 S10 signal in each nucleus.
For FACS analysis cells were fixed/permeabilixed in 70% Ethanol at -20°C for 3 h, washed once with PBS and incubated for 45 min in PBS containing 30 μg/ml RNase and 10 μg/ml propidium iodide. Cells were washed two times with PBS before FACS analysis. For each condition, 10.000 cells were analyzed.
MDM2 (Gene ID: 4193); TP53 (Gene ID: 7157); p21(CDKN1A; Gene ID: 1026); P53R2 (Gene ID: 50484); PAG608 (Gene ID: 64393); Cyclin B1(CCNB1; Gene ID: 891); H2AX (Gene ID: 3014); ATM (Gene ID: 472); ATR (Gene ID: 545); CHK1 (Gene ID: 1111); CHK2 (Gene ID: 11200) KSHV genes: ORF73/LANA (Gene ID: 4961527); ORF50/RTA (Gene ID: 4961526); ORF57/MTA (Gene ID: 4961525); ORF74/vGPCR/K14 (Gene ID: 4961465); ORF25 (Gene ID: 4961452); ORF29 (Gene ID: 4961443); K8.1 (Gene ID: 4961469)
|
10.1371/journal.ppat.1002880 | Elucidation of Hepatitis C Virus Transmission and Early Diversification by Single Genome Sequencing | A precise molecular identification of transmitted hepatitis C virus (HCV) genomes could illuminate key aspects of transmission biology, immunopathogenesis and natural history. We used single genome sequencing of 2,922 half or quarter genomes from plasma viral RNA to identify transmitted/founder (T/F) viruses in 17 subjects with acute community-acquired HCV infection. Sequences from 13 of 17 acute subjects, but none of 14 chronic controls, exhibited one or more discrete low diversity viral lineages. Sequences within each lineage generally revealed a star-like phylogeny of mutations that coalesced to unambiguous T/F viral genomes. Numbers of transmitted viruses leading to productive clinical infection were estimated to range from 1 to 37 or more (median = 4). Four acutely infected subjects showed a distinctly different pattern of virus diversity that deviated from a star-like phylogeny. In these cases, empirical analysis and mathematical modeling suggested high multiplicity virus transmission from individuals who themselves were acutely infected or had experienced a virus population bottleneck due to antiviral drug therapy. These results provide new quantitative and qualitative insights into HCV transmission, revealing for the first time virus-host interactions that successful vaccines or treatment interventions will need to overcome. Our findings further suggest a novel experimental strategy for identifying full-length T/F genomes for proteome-wide analyses of HCV biology and adaptation to antiviral drug or immune pressures.
| Hepatitis C virus infects as many as 170 million people worldwide. Globally, there are seven major genotypes of HCV that differ by approximately 30% in nucleotide sequence. Importantly, the natural history of HCV infection is variable, ranging from spontaneous resolution to persistent viremia and chronic disease. Factors responsible for this variability in clinical outcome are unknown but likely involve a combination of viral and host determinants. To this end, a precise molecular identification of transmitted HCV genomes could illuminate key aspects of transmission biology, immunopathogenesis and natural history. We used single genome sequencing of plasma viral RNA to identify transmitted viral genomes and their progeny in 17 subjects with acute infection. Numbers of transmitted viruses leading to productive clinical infection ranged from 1 to 37 or more (median = 4). Surprisingly, we found evidence of high multiplicity acute-to-acute HCV transmission in 3 of 17 subjects, which suggests that clinical transmission of HCV, like that of HIV-1, may be enhanced in early infection when virus titers are highest and neutralizing antibodies are absent. These results provide novel insight into HCV transmission and early virus diversification key to our understanding of virus natural history and response to drug selection and immune pressure.
| Hepatitis C virus (HCV) infects as many as 170 million people or nearly 3% of the world's population. The virus causes a wide variety of pathologic outcomes, the most significant being chronic liver disease, cirrhosis and hepatocellular carcinoma, which is nearly always fatal. HCV infection is the leading indication for liver transplantation in the United States [1].
HCV is a positive strand, non-segmented, enveloped RNA virus of approximately 9.6 kb in length. The virus is classified in the genus Hepacivirus within the larger family of Flavivirus, which includes the human pathogens West Nile virus, yellow fever virus and dengue fever virus among others [2]. A common feature among the Flaviviridae is their dependence on a virally-encoded RNA-dependent RNA polymerase (RdRp) for replication [3]. RdRp is error-prone, and HCV is notable for its extensive diversity within and among individuals. Globally, there are seven major genotypes of HCV that differ by approximately 30% in nucleotide sequence [1], [4], [5].
The extraordinary diversity of HCV complicates studies of virus biology, pathogenesis and susceptibility to novel therapeutics. Clinically, the different HCV genotypes exhibit variable natural history and responsiveness to interferon, ribavirin and the newer direct acting antiviral (DAA) agents [2], [6], [7]. HCV variation poses similar challenges to the development of effective vaccines and to the elucidation of viral immunopathogenesis [5], [8]–[10]. It is of interest then that the extraordinary diversity of HCV is similar to that of HIV-1 and that a novel experimental strategy to identify transmitted/founder (T/F) HIV-1 genomes has led to new insights into virus transmission and persistence [11]–[18].
Acute HCV infection isconventionally defined as the initial 6 months of infection and sets into motion virus-host interactions that to a large extent dictate the natural history of the disease [9], [10], [19]–[30]. Depending on viral genotype and host immunogenetic factors, most importantly IL28B alleles, a proportion of newly infected individuals spontaneously controls or eliminates virus [9], [10], [31]–[33]. A greater number can be cured if the infection is treated with interferon and ribavirin alone or in combination with DAA drugs [6], [34]–[36]. Mechanistically, how this occurs is incompletely understood. From a vaccine perspective, the acute infection period is critical, since transmitted viruses are the obvious targets of a vaccine and early stages of infection when viral diversity is lowest represent a period when the virus may be most vulnerable to elimination by vaccine-elicited immune responses [26], [37]. For all these reasons, there is considerable interest in the molecular features of the initial virus population ‘bottleneck’ associated with virus transmission and the subsequent pathways of virus evolution that lead to persistence [19]–[25], [28], [29], [38], [39].
Previous reports have described different experimental approaches to the analysis of the HCV transmission bottleneck. These include studies that employed a DNA heteroduplex gel shift method to estimate viral diversity [39], [40], conventional polymerase chain reaction (PCR) methods to bulk amplify, sequence, or clone and sequence fragments of HCV genomes [20], [21], [23], [25], [28], [38], [41], [42], and 454 pyrosequencing to interrogate early viral sequences more deeply but narrowly [19], [43]. These reports documented a restriction in viral diversity associated with virus transmission, but despite the use of increasingly sensitive methods, a precise quantitative, molecular description of HCV transmission and early diversification has remained elusive. In the current study, we hypothesized that T/F HCV genomes could be identified unambiguously and their early pathways of diversification mapped precisely by single genome amplification (SGA) followed by direct amplicon sequencing, otherwise known as single genome sequencing [17], an approach we used previously to gain insight into HIV-1 transmission [15], [18]. This strategy differs from previous methods applied to HCV by providing gene-wide or genome-wide viral sequences that are proportional to their representation in human plasma and are not confounded by template resampling or by Taqpolymerase errors of nucleotide misincorporation or recombination [15], [17], [18], [44]–[47]. We amplified and sequenced HCV core, E1, E2, P7, NS2 and NS3 genes and analyzed them by adapting a model of random HIV-1 evolution [15], [48], [49] to account for differences in the biology of replication between the HIV-1 and HCV. In an accompanying report, Ribeiro and colleagues [50] used these sequences together with plasma viral load kinetic data to develop an agent based stochastic model of acute HCV replication dynamics resulting in new estimates of the HCV mutation rate in humans.
One hundred fifty-four plasma specimens from 17 subjects with acute HCV infection and 14 subjects with chronic HCV infection were analyzed for viral RNA (vRNA) load (Figure 1; Table S1). Acute infection subjects were source plasma donors who had undergone once or twice weekly plasmapheresis for months, and in some cases years, with persistently negative testing for HCV, HBV and HIV antibody or RNA before becoming acutely infected by HCV. Chronically infected subjects were patients at the University of Alabama at Birmingham who were known to have been HCV infected for approximately 3 to >20 years. All subjects were untreated with anti-HCV medications. A median of 8 (range 5–11) sequential specimens per acutely infected subject was analyzed for vRNA load spanning the period of plasma vRNA negativity through exponential increase to an early plateau (Figure 1), and a subset of these was analyzed for sequence diversity. The median peak plasma viral load in acute infection subjects was 2,850,000 IU/ml (range = 527,000–10,300,000 IU/ml). Five of 17 acutely infected subjects developed HCV antibodies by the last sampling time point. Chronic subjects had a median vRNA load of 2,081,138 IU/ml (range = 24,000–7,690,001 IU/ml), all were HCV antibody positive, and all were sampled once for vRNA sequence diversity. A total of 2003 5′ half genomes and 919 5′ quarter genomes were amplified, sequenced, and recorded (Genbank accession numbers JQ801756–JQ804520; JX178293–178443).
Sequences corresponding to core antigen and envelope E1 and E2 genes (2.2 kb) from the initial HCV RNA positive sample from acute and chronic subjects were subjected to maximum-likelihood (ML) phylogenetic analysis (Figure 2). Sequences formed subject-specific clades (bootstraps 93–100%) and represented HCV genotypes 1a (n = 23), 1b (n = 4), 2b (n = 2) or 3a (n = 2). No subject was infected by more than one virus genotype and there was no intermixing of sequences between subjects in the phylogenetic tree. Sequences from acute and chronic subjects revealedwidely varying degrees of maximum within-subject diversity ranging from 0.14% to 6.40% (Tables 1 and S1), which was not different between the two groups (p>0.05, Mann-Whitney test). Acute sequences were, however, distinctly different from chronic sequences in having one or more discrete lineages characterized by extremely low diversity. This was true for sequences from each acutely infected subject but from none of the chronically infected subjects. Maximum diversity of sequences within these discrete viral lineages from acute subjects (mean 0.12%; median 0.12%; range 0.04–0.19%) was significantly lower than the overall diversity observed within chronic subjects (mean 2.27%; median 2.37%; range 0.56–3.83%; p<0.0001, unpaired T-test with Welch's correction) (Figure S1).
Because of the distinctive replication strategy of HCV, which is still not fully understood [3], [5], we could not predict a priori the patterns of early virus diversification that we might observe in acutely infected subjects. Thus, we developed two mathematical models to analyze HCV sequence diversity (Figure 3). The first model [15], [48], [49], employed previously to analyze early HIV-1 sequence diversification, assumes a narrow genetic bottleneck associated with virus transmission, initial rapid exponential virus growth, constant lineage-independent mutation rates at all sites, no recombination between sequences or back mutations, and no differential selection. When diversity is low, most base substitutions occur at distinct loci, pairwise differences between sequences (i.e., Hamming distances, HD) follow a Poisson distribution, and sequences exhibit a star-like phylogeny and coalesce to distinct unambiguous T/F genomes [15], [18], [51]. Early HIV-1 diversification conforms well to this model and it was suggested that other viruses including HCV might also [48]. However, HCV replication differs from HIV-1 in that HCV RNA does not integrate into chromosomal DNA and it does not produce all of its daughter progeny in a large burst of viruses within a couple days after infecting a cell [52]. Instead, it forms as many as 40 cytosolic replication complexes that continue to produce virions throughout the lifetime of the cell [50], [53]. In early infection prior to the onset of HCV specific cellular immune responses, the lifetime of infectedcells is likely to span our sampling period given that the lifetime of uninfected hepatocytes is estimated to be months to years [54], [55]. To account for these differences with HIV-1, we developed an alternative simplified deterministic model of HCV diversification (Figure 3). This model predicts occasional violations in star-like diversification and in the Poisson fit of mutations with increasing probability as time goes on, and it predicts greater numbers of shared stochastic mutations between HCV sequences compared with HIV-1 sequences. The latter model provided us with a mathematical and statistical basis for distinguishing closely related T/F HCV lineages from sequences that evolved from a single genome but shared early stochastic mutations. Importantly, it allowed us to vary key assumptions regarding the contributions of long-lived hepatocytes containing multiple generations of replication complexes and assess the effects on sample-based T/F virus enumeration. From this analysis, we adopted a conservative operational cut-off of >2 shared mutations per quarter genome (∼2500 bp) or >4 shared mutations per half genome (∼5000 bp) to distinguish T/F genomes in most cases of HCV transmission from chronically infected individuals (see Methods). In more complicated transmission scenarios, where the index case was hypothesized to be either acutely infected or to have experienced a viral genetic bottleneck due to antiviral drug therapy, we applied both an empirical approach where a single shared polymorphism could represent a distinct T/F viral genome as well as the more conservative modeling approaches. Importantly, both empirical and model-based analyses predicted that consensus sequences of low diversity lineages before the onset of immune-driven positive selection coalesced to founder viral genomes at or near the moment of transmission.
Figure 4 depicts ML phylogenetic trees and Highlighter plots of viral 5′ half genome sequences from two subjects, one typical of chronic infection (WIMI4025, panel A) and one illustrating the reduced genetic diversity characteristic of a subject productively infected by a single viral genome (10051, panel B). Sequences from the chronic subject showed broad genotypic heterogeneity with a maximum inter-sequence diversity of 3.83% (mean 1.2%; median 1.21%; range 0.12–3.83%) typical of chronic infection (Table S1). Sequences from the acute subject revealed a very different pattern of diversification (Figure 4B). These sequences, which were derived from the last sampled time point 21 days after the beginning of documented viremia (see Figure 1), were extremely homogeneous with a mean diversity of 0.03%, median diversity of 0.02%, and a range in diversity of 0–0.19%. Nucleotide substitutions corresponded to a near star-like phylogeny, although unlike most early HIV-1 sequence sets they deviated from a Poisson distribution (p<5×10−5), a finding consistent with predictions of the HCV adapted model. This deviation resulted from two sequences (2C3 and 2C2) that contained a single shared polymorphismat position 2197 and two other sequences (2A2 and 2B34) that contained a different shared polymorphism at position 4344. The shared polymorphism at position 2197 was a first position GAG to TAG transversion that resulted in the introduction of a stop codon. Surprisingly, the same nonsense mutation was found in the identical position in two additional sequences (2B8 and 2B9) from this subject in a plasma sample taken 12 days earlier (Genbank accession nos. JQ803586, JQ803587, JQ803590 and JQ803591). We could be assured that these persistent but defective genomes were authentic and did not result from cross-contamination of amplicon sequences since each of the four sequences had additional distinguishing nucleotide polymorphisms (e.g., compare sequences 2C3 and 2C2 in Figure 4B) or were processed, PCR amplified and analyzed on different days (e.g., 2C3 and 2C2 versus 2B8 and 2B9). Occasional shared polymorphisms are commonly found in acute HIV-1 infection sequences and can be explained by polymerase errors early in infection being retained in the population [15], [49], but for neither HIV-1 nor HCV would nonsense mutations be expected to be retained unless they were complemented by competent genomes [50], [56], [57]. Regardless, the 60 sequences depicted in Figure 4B coalesced to a single unambiguous consensus, which we inferredto represent a likely T/F virus in this subject. To explore if additional T/F sequence lineages might have been overlooked due to inadequate sampling, we sampled 243 additional 5′ quarter genome sequences from this subject at three earlier time points spanning a 19 day period (Figures 1 and S2). Power calculations indicate that a sample size of 60 sequences provides 95% likelihood of detecting variants present at 5% in the population [15], whereas a sample size of 303 sequences provides 95% likelihood of detecting variants present at 1% prevalence. 43 of 46 (93%) of the sequences at the initial time point were identical, and diversity increased with time (Figure S2). Whether the 303 5′ quarter genome sequences from the different time points were considered separately or together, they conformed to a near star-like phylogeny and coalesced to the same T/F genome. Thus, we can conclude with a high level of confidence that subject 10051 was productively infected by a single virus whose sequence is represented by the consensus in Figure 4B and S2.
Figure 5 extends the analysis of T/F HCV genomes to four acutely infected subjects where sequences from sequential time points revealed variable patterns of early viral diversity. In each subject, sequences from the initial sample were more homogeneous than those from later time points, as expected in a model of random accumulation of mutations. For subject 10021 (panel A), 19 of 24 (79%) sequences from the initial timepoint were identical. For subject 10025 (panel B), 32 of 43 (74%) of initial sequences were identical. The remaining sequences from this first time point in each subject differed from the respective consensus sequences by only 1 or 2 nucleotides. At the second and third sampling time points 1–4 weeks later, an increasing proportion of sequences from each subject differed from the respective consensus sequences by as many as 3 or 4 nucleotides. Interestingly, in both samples rare shared mutations became evident at later time points, again consistent with predictions of the HCV adapted model. For subjects 10021 and 10025, we thus concluded that the respective consensus sequences corresponded to single T/F HCV genomes. Of the 17 acutely infected subjects, four had evidence of productive clinical infection by single viruses (Tables 1 and S2). Panels C and D depict sequences from subjects 10012 and 10062, each of whom had evidence of productive infection by more than one genetically distinct virus based on the presence of multiple discrete low diversity lineages whose consensus sequences differed from each other by far more than the 2 nucleotides per quarter genome cut-off. Importantly, unlike HIV-1 where viral recombination in acute and early infection is extremely common [12], [16], we found no evidence of inter-lineage recombination in HCV sequences from these subjects or from any other subjects in this study.We thus interpreted the consensus sequences of each low diversity lineage in subjects 10012 and 10062 to correspond to a unique T/F HCV genome, three for each subject. This represents a minimum estimate, since deeper sampling could conceivably identify additional T/F sequence lineages, although with 188–230 sequences analyzed there was a 95% likelihood of detecting variants present at 2% in the population (Table S2). Figure 6 depicts sequences from subject 10029 where discrete low diversity sequence lineages indicated clinical acquisition of a minimum of 9 T/F viruses. In 6 additional acutely infected subjects, clearly distinguishable T/F lineages ranged from 3 to 13 per subject (Figures S3, S4, S5, S6, S7, S8).Importantly, for all of the acutely infected subjects described above, the numbers of T/F sequence lineages identified by visual inspection using the cut-off of >2 per quarter genome and >4 per half genome were nearly identical to those inferred from the HCV adapted model of early virus diversification using standard (maximum cut-off) or stringent (average cut-off) assumptions (Table S2).
Virus diversity in three acutely infected subjects (10003, Figure 7; 10020, Figure S9A; 10016, Figure S9B) shared features that distinguished them from the other 14 acute subjects. In each case, maximum sequence diversity was relatively low (0.27%–0.41%), consistent with recent infection. However, unlike in other subjects, samples from 10003, 10020 and 10016exhibited many closely related sets of sequences that shared unique polymorphisms, which appeared as multiple vertical ‘stripes’ in the Highlighter plots and shared nodes in the phylogenetic trees (Figures 7 and S9A,B). In subject 10003 (Figure 7), there were at least 37 distinctsequence sets, and in subjects 10020 (Figure S9A) and 10016 (Figure S9B) there were 10 and 15, respectively. Although we recognized that occasional shared mutations are predicted by the model and were found in samples from subjects infected by single or few viruses with well-defined T/F sequence lineages (e.g., see Figures 4B, 5 and 6), the high frequency of unique shared mutations in subjects 10003 (Figure 7), 10020 (Figure S9A) and 10016 (Figure S9B) was quite unusual and led us to hypothesize that most of the 10–37 distinct virus lineages observed in these individuals resulted from discrete transmitted viruses from individuals who themselves were recently infected by a single virus or by closely related viruses. An alternative hypothesis that we considered was that infection by single viruses had occurred in subjects 10003, 10020 and 10016but was followed by atypical diversification patterns not seen in any of the other acutely infected subjects. Our strategy to distinguish between these scenarios was both empirical and model-based. First, we observed that the maximum Hamming distances, expressed as per cent diversity, between the consensus sequences of the discrete sequence sets (i.e., between potential T/F genomes) from subjects 10016, 10020 and 10003, were 0.21%, 0.28% and 0.35% (mean 0.28%; median 0.28%), respectively. These values exceeded the maximum intra-lineage diversity (mean 0.12%; median 0.12%; range 0.04–0.19%; p = 0.0051, Mann-Whitney test) found in all lineages from all subjects in the initial ∼6 weeks of infection. These findings suggested acquisition of multiple distinct variants from recently infected subjects, not evolution of viruses from single transmitted variants. Secondly, we found that the multiple sequence sets in subjects 10016, 10020 and 10003 were present from the initial sampling time points and did not accumulate over time, again suggesting acquisition not evolution of the variants. Thirdly, we noted that the linear distribution of shared nonsynonymous nucleotide substitutions across the viral genomes in subjects 10016, 10020 and 10003 was not random but instead was concentrated in regions of E1 and E2 (Figure S10) previously associated with immune escape [41], [42], [58]. Inspection of the amino acid substitutions in the E2 region of the putative T/F genomes revealed combinations of different nonsynonymous mutations in identical or neighboring positions in the primary sequence that were unlikely to have occurred by chance (probability estimates <0.001). Since these subjects were sampled very early after infection, well before antibody seroconversion (Figure 1) or the onset of cellular immune responses [9], [10], the data suggest that the patterns of viral diversity observed in the putative T/F viral genomes from subjects 10016, 10020 and 10003 were the consequence of acquisition of multiple variants from transmitting individuals who themselves had become infected in the preceding ∼6 months and whose viral sequences had been subjected to early epitope-focused immune selection.
We next applied our HCV adapted model of early virus diversification to the sequences from the three subjects 10016, 10020 and 10003(Figures 8; S14–15). Under standard model assumptions (maximum cut-off, see methods) for delineating T/F lineages, transmission by at least 6–19 genetically distinct viruses was necessary to explain the viral diversity observed in the three subjects. Even under the most conservative model assumptions (average cut-off), transmission of at least 2–9 distinct viruses was required. Based on these findings, we conclude that the pattern of viral sequence variation in subjects 10003, 10020 and 10016can be most plausibly explained by acute-to-acute virus transmission.
The phylogenetic pattern of sequences from subject 106889 (Figure 9) was still more complicated than that of any of the other 16 acutely infected subjects. This subject exhibited a typical acute infection viral kinetic profile with four sequential plasma samples negative for HCV vRNA and antibody followed by rapid vRNA ramp-up to nearly 4×106 vRNA IU/ml (Figure 9 insert). These plasma samples from subject 106889 were obtained in June and July 2008 and were preceded by over 150 plasma collections from this individual in 2003–2008, all of which were negative for HCV RNA or antibody, proving this individual had incident infection. Eighty-seven 5′ half genome sequences were obtained from the initial plasma vRNA positive time point (Figure 9). Maximum diversity of these sequences was 1.03%, indicating multi-variant transmission. Similar to subjects 10003, 10020 and 10016, numerous discrete, low diversity sequence clades, many of which were closely related to each other, were apparent in the maximum-likelihood tree and Highlighter plot. Unlike sequences from the other subjects, however, the distinct sequence sets in 106889 clustered with high bootstrap values into larger lineages (color-coded in Figure 9). Thoseclades that contained sufficiently large numbers of sequences for analysis (e.g., clades identified by sequences 5.B.F9 and 5.02C22) exhibited a star-like phylogeny and Poisson distribution of mutations (Table 1), indicating that they had evolved very recently from discrete T/F genomes.These findings suggested that subject 106889 had been infected by large numbers of virusesfrom a chronically infected individual whose HCV sequences had been subjected to a stringent genetic bottleneck. A quite unexpected finding suggested a likely explanation: 86 of the 87 sequences from subject 106889 were found to contain two signature mutations in the NS3 protease gene (V36M and R155K) that confer high level drug resistance to the NS3 protease inhibitors Boceprevir and Telaprevir [7]. One of 87 sequences (sequence 02B11 in Figure 9) contained one of these DAA resistance mutations (V36M). Since the combination of V36M and R155K mutations is uncommon in treatment-naïve individuals [59], this result suggested that a transmitting partner to 106889 was chronically infected with HCV, was treated with an investigational NS3 protease inhibitor, experienced a DAA-induced viral population bottleneck followed by the emergence of NS3 protease resistant variants, and as these variants rebounded, transmitted multiple drug resistant variants directly to subject 106889 or indirectly through a second acutely infected individual. Under this scenario, where closely related sequences differing by as few as one nucleotide would be expected to be transmitted, we estimated by visual inspection of the phylogenetic tree that as many as 30 or more T/F virus could have been responsible for productive clinical infection (Table S2). By the standard model analysis, 28 distinct T/F viral genomes were necessary to account for the observed diversity. By the more conservative model analysis, at least 16 distinct T/F genomes were needed (Table S2).
The identification of T/F viral sequencesin 17 subjects provided us with a unique opportunity for analyzing HCV sequence evolution in natural human infection beginning at or near the moment of virus transmission and extending through the establishment of early viral load setpoint, and in some subjects, antibody seroconversion. Among the 17 subjects, we identified a total of 146 T/F genomes. As expected, all 146T/F sequences had intact open reading frames for core, E1, E2, P7, NS2 and NS3. Thirty T/F genomes had 5 or more identifiable progeny (range 5–303; median 19) from which we could analyze molecular features of sequence diversification in vivo using phylogenetic tools and algorithms. A summary of this analysis is presented in Table 1. Maximum intra-lineage diversity for the 17 subjects ranged from 0.04% to 0.19% (mean = 0.12%; median = 0.12%), which was significantly lower than the maximum, mean and median viral diversities observed in chronically infected subjects (3.83%, 2.27% and 2.37%, respectively; p<0.0001, unpaired T-test with Welch's correction) (Tables 1 and S1; Figure S1). Evolved sequences compared to their respective T/F genomes revealed low frequencies of per nucleotide insertions (1×10−6), deletions (3×10−6) and stop codons (1×10−6). Transitions outnumbered transversions by 8.8 to 1, and when corrected for the number of available sites, by 18 to 1. The average dN/dS ratio was low at 0.39. The overall mutation frequencyamong all 17 acute subjects uncorrected for time from transmission or numbers of virus replication cycles was 1.4×10−4. When sequences were analyzed in the context of an agent-based stochastic model of virus diversification that incorporates estimated time from transmission and HCV specific parameters of virusreplication, the mutation rate of HCV in vivo was estimated to be 2.5×10−5 per nucleotide per genome replication [50]. We confirmed this low value by an analysis of the nonsense codon frequency per nonsense mutation target site [50]. Progeny of T/F viruses sampled at the earliest time points generally conformedto a star-like phylogeny and a Poisson distribution of randommutations (Table 1), but at later time points there were occasional deviations. Deviations from the model were of three types: (i) shared polymorphisms resulting from stochastic changes after the transmission event or from the transmission of multiple closely related viruses; (ii) immune selection or reversion in later samples at the time of antibody seroconversion in a single subject 9055 (Figures S16, S17); (iii) rare examples of short perfect inverted repeats, 3–20 nucleotides in length, that resulted from template switching between double-stranded RNA molecules in locations prone to RNA stem-loop secondary structure (Figure S18). This latter finding, which was found in 7 sequences out of 2922 analyzed,was observed in samples from three different study subjects. Four of these strand transfers occurred at the same location in the core gene.
The present study provides new quantitative and qualitative insights into HCV transmission and early diversification in humans. Previous reports documented a virus population bottleneck associated with HCV transmission, but none of those studies including ones based on 454 deep sequencing captured the broad range in multiplicity of infection or the full spectrum of genetic diversity that exists among transmitted viruses. In our study of 17 acutely infected subjects, we could unambiguously identify and determine the exact nucleotide sequences of one or more T/F virus genomes in each subject. This was true for all subjects whose HCV genomes were sequenced within the initial ∼6–8 weeks of infection; beyond that there were examples of immune selection that confounded the identification of T/F virus genomes (Figures S16 and S17). We estimated the multiplicity of infection (numbers of T/F viruses leading to productive clinical infection) to range from 1 to as many as 37 or more with a median of 4. These are minimum estimates given our sampling limitations, although we note that our median sampling depth of 151 sequences (Table 1) afforded us a 95% likelihood of detecting variants present at 2% prevalence [15]. In subjects productively infected by lower numbers of viruses (<10), where the progeny of each transmitted virus is repeatedly sampled, our estimates (Tables 1 and S1) are likely to be an accurate and precise measure of the number viruses that result in productive infection. In subjects infected by higher numbers of viruses, especially in the setting of acute-to-acute transmission where transmitted viruses are expected to differ by as few as one nucleotide, the accuracy of our estimates are necessarily less. This is because we could not sample deeply enough due to practical constraints of single genome sequencing of quarter and half genomes, and because we could not distinguish between transmitted viruses that differ by one or few nucleotides from single variant transmission followed by early stochastic mutations. However, based on the striking differences in diversity patterns that we observed between subjects with chronic-to-acute versus apparent acute-to-acute transmission, we suspect that the actual numbers of T/F viruses in subjects 10016, 10020, 10003 and 106889 approximate or exceed our estimates of 15, 10, 37 and 30 T/F genomes, respectively (Tables 1 and S2).
The broad range in numbers of T/F viruses responsible for acute HCV infection in our cohort must reflect the different transmission routes and risk practices ofsource plasma donors. Ostensibly, such individuals should be at low risk of acquiring HCV infection since they are qualified as regular source plasma donors only after extensive pre-enrollment screening that consists of medical histories, physical examinations and behavioral questionnaires designed specifically to eliminate from the donor pool individuals at risk for HCV, HBV or HIV infection (http://www.fda.gov/BiologicsBloodVaccines/GuidanceComplianceRegulatoryInformation/default.htm). However, self-reporting of risk behaviors among paid plasma donors is admittedly imperfect [60]. Thus, it is likely that the subjects in the present study represent the broad clinical spectrum of community-acquired HCV infection in the United States, which includes injection drug users, men who have sex with men, heterosexuals, and possibly, household contacts of HCV infected individuals.
Our findings regarding the multiplicity of human infection by HCV are quite different from those obtained by 454 pyrosequencing in seven acutely infected subjects reported by two different investigative groups where the range in T/F viruses was one to four with a median one [19], [43]. Our findings are also substantially different from estimates from other reports that employed reverse transcription, bulk PCR amplification, population sequencing or molecular cloning followed by sequencing [20], [23], [25], [30], [38], [41], [42].The latter studies showed that acutely infected subjects exhibited a spectrum in HCV sequence diversity that could at best be interpreted qualitatively as reflecting ‘few-variant’ versus ‘multi-variant’ transmission. We also note a recent study that used conventional bulk PCR amplification, cloning and sequencing to analyze acute and early HCV sequences consisting of a 225 bp hypervariable region of env from 10 acute infection subjects following IDU, sexual or nosocomial exposures [28]. Thisreportdescribed perplexing findings: 7 of 10 acutely infected subjects seemed to harbor more than one HCV genotype andsequential sequences obtained from these subjects a median of 17.5 days apart throughout the acute infection period suggested fluctuations in the prevalence of different HCV genotypes, subtypes and clades. These findings are at odds with our results (Figures 2, 4–9, S2, S3, S4, S5, S6, S7, S8, S9 and S14, S15, S16) and those of most other studies [24].
The SGA-direct amplicon sequencing strategy used in the present study represents a substantial advance in sensitivity and molecular resolution for distinguishing closely and distantly related T/F HCV genomes and their evolving progeny. Studies of HCV specific CTL recognition and escape [20], [25], [61], [62], neutralizing antibody recognition and escape [41], [42], and DAA drug resistance development [7] have previously been performed without a precise identification of T/F viral genomes and future studies may benefit from such an approach. We could readily distinguish evolving viral lineages that differed from the T/F genome by just 1 nucleotide in 5,000 (0.02%) at sites under selective pressure (Figures S16–17). This discriminating power further revealed evidence of acute-to-acute virus transmission in three subjects (Figures 7, 8 and S9) and DAA drug-induced viral genetic bottlenecking in a donor to a fourth acutely infected subject(Figure 9). This exquisite sensitivity in distinguishing T/F virus genomes and their progeny stands in contrast to the 454-based approaches, which were unable to distinguish between T/F viruses that differed by less than 2.5% [43] and bulk PCR-clone-sequencing methodsthat used a cutoff of 3% diversity to distinguish homogeneous from heterogeneous virus transmission [23]. Both of the latter methods are further confounded by the potential for Taq polymerase-mediated strand transfers leading to recombination artifacts in finished sequences [15], [19], [63].
The ability to identify actual T/F viral sequences and to track virus diversification from these sequences with single nucleotide resolution provided a unique opportunity to assess HCV sequence evolution in vivo. Virus diversification from discrete T/F viruses was generally star-like and conformed well to our HCV adapted model of early virus diversification. The finding of only a single instance of potential CTL escape or reversion among 17 acutely infected subjects at the last sampling time point is consistent with previous reports indicating substantial delays in the onset of adaptive immunity to HCV [9], [10]. The overall nucleotide substitution frequency that we observed among all subjects and including all sampling time points was 1.4×10−4. This substitution frequency is different from the mutation rate since it does not account for time, numbers of replication cycles, or different modes of HCV replication (linear versus geometric) [3], [57], [64], nor does it account for nucleotide substitutions introduced by the MuLV polymerase (Superscript III) during cDNA synthesis. The latter is estimated to occur at a frequency as low as 2×10−6 [65] and thus likely contributes negligibly to the mutation frequencies observed in the present study. Consistent with this interpretation were our results of single genome sequencing performed on the earliest vRNA positive plasma sample from subject 10051 where we found that 43 of 46 5′ quarter 1 genome sequences were identical (Figure S2). Among all 46 sequences, there were only four nucleotide substitutions in 100,050 nucleotides. This corresponds to a combined substitution frequency for the HCV polymerase and the Superscript III MuLV polymerase of 4×10−5, Again, thisresult does not account for the numbers of HCV replication cycles occurring between the moment of virus transmission and the time point of sampling, which in this case was very early during the viral ramp-upperiod when the plasma virus load was approximately 10,000 vRNA molecules/ml (Figure 1). In an accompanying report [50], we describe a new stochastic model of HCV replication and diversification that provides for a more precise estimation of the in vivo HCV RdRp error rate, which was found to be ∼2.5×10−5 per base per generation. This is lower than previous reports for HCV [56], [57] and comparable to the RT error rate of HIV-1 [57], [66].
We found a low dN/dS ratio consistent with early negative or purifying selection and a strong 18 to 1 mutational bias for transitions over transversions in acute infection. The latter finding is consistent with a recent report by Gotte and colleagues [67] who studied sequence evolution in chronically infected subjects and in vitro where a strong preference for G∶U/U∶G mismatches was observed for recombinant HCV RdRp. A mutational bias favoring transitions may be a factor besides RdRp error rate that influences the rate of development of DAA resistance mutations [67]. Importantly, we found no evidence of viral recombination in any subject, which would have been plainly evident in those subjects infected by multiple genetically diverse viral genomes (Figures 5C, D; 6; S3–9). The absence of recombination distinguishes HCV from HIV-1, where early recombination is widespread [12], [16], [46], but is consistent with molecular epidemiological data that suggest that HCV recombination is rare [68]–[71]. In addition, the failure to find plus-plus strand recombination in any of the sequences in the present report shows that strand switching by the MuLV reverse transcriptase (RT) in vitro must be extremely rare. This is important because it demonstrates that MuLV RT-mediated recombination does not confound single genome sequence analyses of HCV or other RNA viruses including HIV-1 andSIV [15], [18], [51]. On the other hand, we did observe seven examples of template switching between plus and minus strands of double-stranded HCV RNA templates (Figure S18). We could not determine if this resulted from strand switching by MuLV RT in vitro or by HCV RdRp in vivo. We note that Branch and colleagues [72] recently reported high levels of double-stranded HCV RNA in hepatic tissue, thus providing a plausible source of dsRNA for the observed template switching events.
A surprising finding of the current study was evidence of acute-to-acute HCV transmission in a relatively high proportion (3 of 17) of subjects. The acute infection period of HCV, like that of HIV-1, is characterized by very high plasma virus loads, absence of neutralizing antibodies, and rapid expansion of biologically fit virus populations that are homogeneous relative to the respective T/F virus genomes [22], [26], [27]. For HIV-1, the acute and early infection period has been shown to be associated with hyper-transmissibility with epidemiological studies and epidemic modeling indicating substantial enhancement in spread of the virus as long as six months post-transmission [73]–[76]. In the simian immunodeficiency virus (SIV) – Indian rhesus macaque transmission model, virus from acute infection plasma is up to 750-fold more transmissible on a per virion basis than is virus from chronic infection plasma [77]. To our knowledge, a clinical predilection for acute-to-acute HCV transmission has not previously been reported. In addition to the three subjects whom we identified with putative acute-to-acute HCV transmission, an argument can be made for an additional potential case in subject 10017 in whom distinct subsets of closely related T/F sequences were found within a context of high overall sequence diversity (e.g., see lineages v1 and v3; Figure S8). In this example, a plausible scenario is that a virus ‘donor’ to subject 10017 was acutely infected by multiple genetically-diverseviruses and that multiple progenyrepresenting several of these lineages were transmitted. The implication of these findings is that if the acute period of HCV infection is characterized by hyper-infectiousness as is the case for HIV-1, it could be a previously unrecognized but important contributing factor to the spread of HCV, potentially contributing to a recently described emerging HCV ‘epidemic’ in HIV-1 positive men who have sex with men [78], [79]. A limitation in our evidence supporting ‘acute-to-acute’ infection is that our study design did not allow us to identify paired donors and recipients of virus in order to analyze virus transmission directly. Future viral sequencing studies involving social networks of HCV transmission partners [80], or analyses of cryopreserved plasma specimens from previously conducted acute-to-acute human-to-chimpanzee HCV transmission studies [81], can provide corroborative evidence. We notethat there is precedent for phylogenetic linkage of HCV sequences in a human-to-human transmission case where clinical epidemiologic linkage between donor and recipient was established [42].
Still another surprising observation in this study was transmission of what we estimated to be as many as 30 NS3 protease-resistant viruses to subject 106889 (Figure 9). These mutations (V36M and R155K) confer high level resistance to both Boceprevir and Telaprevir, which were used in clinical trials near the time when 106889 samples were collected. Recently, we performed single genome sequencing of plasma viral RNA from subjects before and after treatment with a next generation investigational HCV protease inhibitor and observed viral genetic bottlenecking closely resembling that found in subject 106889 (unpublished data). To our knowledge, the data from subject 106889 is the first example of high multiplicity DAA drug resistant virus transmission, and the findings here illustrate how transmission of DAA resistant mutants can be deciphered with single genome specificity and sensitivity.
The identification of T/F genomes of HCV, HIV-1 [18], SIV [51] and potentially other RNA viruses by single genome sequencing is an enabling experimental strategy that captures molecular entities that are wholly sufficient and responsible for productive clinical infection and disease causation. In an accompanying report [50], we use sequences derived by this approach to analyze and mathematically model the early dynamics of HCV replication and diversification in acutely infected humans and derive new estimates of the in vivo mutation rate of HCV. A second application of the single genome sequencing method is to reveal through enumeration of T/F genomes, the challenge that vaccine candidates face in attempting to prevent or constrain HCV transmission. In a third application of the method, we previously demonstrated for HIV-1 that single genome sequencing allows for the molecular identification, cloning and biological characterization of full-length T/F genomes and a comprehensive proteome-wide analysis of autologous, strain-specific patterns of cytotoxic T-cell and neutralizing antibody responses [13], [18], [82]–[84]. By demonstrating that early HCV diversification generally conforms to a model of essentially random virus evolution where sequences coalesce to distinct, unambiguous T/F genomes, the present study has taken the first critical steps to demonstrate the feasibility of similar genome-wide analyses for HCV. An intriguing possibility is that full-length T/F HCV genomes, which by definition possess nucleotide and amino acid sequences sufficient for efficient in vivo replication in humans, can be identified, molecularly cloned and expressed for biological analyses in cell culture and animal models.
This study was conducted according to the principles expressed in the Declaration of Helsinki. It was approved by the Institutional Review Boards of the University of Pennsylvania, the University of Alabama at Birmingham and Duke University. Subjects provided written informed consent for the collection of blood samples and subsequent analyses.
Plasma samples were obtained from 17 subjects with acute HCV infection. These subjects were regular source plasma donors (ZeptoMetrix, Inc.; SeraCare, Inc.) who were HCV and HIV-1 antibody negative but who became HCV infected sometime in the course of their twice-weekly plasma donations as evidenced by the development of HCV viremia on sequential viral RNA testing (Figure 1). The subjects were asymptomatic throughout the collection period and did not receive anti-HCV treatment. By the time of the last sample collection, 5 of the subjects had seroconverted to HCV antibody positivity. Plasma samples from 14 patients from University of Alabama at Birmingham with chronic, treatment-naïve HCV infection were obtained as controls.
Plasma samples were tested for HCV RNA and antibodies by a battery of commercial tests. These included Roche COBAS AmpliPrep/COBAS Taqman HCV vRNA assay; ABBOTT Anti-HCV 3.0 Assay and ORTHO Enhanced SAVe Anti-HCV 3.0 Assay. HCV vRNA analyses were performed according to manufacturer's specifications (http://www.accessdata.fda.gov/cdrh_docs/pdf6/P060030c.pdf) in a CLIA certified laboratory with all assay controls meeting predetermined parameters for assay sensitivity and specificity and with a dynamic linear range of 43 to 6.9×107 vRNA IU/ml.
For each plasma sample, approximately 100,000 viral RNA copies were extracted using the QiagenBioRobot EZ1 Workstation with EZ1 Vrius Mini Kit v2.0 (Qiagen, Valencia, CA). RNA was eluted and immediately subjected to cDNA synthesis. Reverse transcription of RNA to single stranded cDNA was performed using MuLV (SuperScript III) reverse transcriptase using methods recommended by the manufacturer (Invitrogen Life Technologies, Carlsbad, CA). Briefly, each cDNA reaction included 1× RT buffer, 0.5 mM of each deoxynucleoside triphosphate, 5 mM dithiothreitol, 2 units/ulRNaseOUT (RNase inhibitor), 10 units/ul of SuperScript III reverse transcriptase, and 0.25 uM antisense primer. The antisense primers were designed specifically for different genotype. 1.NS4A-R1 5′- GCACTCTTCCATCTCATCGAACTC -3′ (nt 5451–5474 H77 (accession number NC_004102)) for genotype 1, 2NS2-R1 5′-CCCCAGACGATGACTTTCTTCTCCAT-3′ (nt 5445–5467 H77) for genotype 2 and 3aNS3-R2V2 5′–TTACTTCCAGATCAGCTGACA-3′ for genotype 3. The reverse transcription reaction was carried out at 50°C for 60 minutes followed by an increase in temperature to 55°C for an additional 60 minutes. The reaction was then heat-inactivated at 70°C for 15 minutes and then treated with 0.1 U/ul RNaseH at 37°C for 20 minutes. The newly synthesized cDNA was used immediately or kept frozen at −80°C.
cDNA was serially diluted and distributed among wells of replicate 96-well plates (Applied Biosystems, Foster City, CA) so as to identify a dilution where PCR positive wells constituted less than 30% of the total number of reactions. At this dilution, most wells contain amplicons derived from a single cDNA molecule. This was confirmed in every positive well by direct sequencing of the amplicon and inspection of the sequence for mixed bases (double peaks), which would be evidence of priming from more than one original template or the introduction of PCR error in early cycles. Any sequence with evidence of mixed bases was excluded from further analysis. PCR amplification was carried out in the presence of 1× High Fidelity Platinum PCR buffer, 2 mM MgSO4, 0.2 mM of each deoxynucleoside triphosphate, 0.2 uM of each primer, and 0.025 units/ul Platinum Taq High Fidelity polymerase in a 20 ul reaction (Invitrogen, Carlsbad, CA). The nested or hemi-nested primers for generating 5′ half or 5′ quarter genome from different genotypes included: (1) 5′ half genome of genotype 1: 1st round sense primer 1.core.F1 5′-ATGAGCACGAATCCTAAACCTCAAAGA-3′ (nt 342–368 H77) and 1st round antisense primer 1.NS4A.R1 5′-GCACTCTTCCATCTCATCGAACTC-3′ (nt 5451–5474 H77), 2nd round sense primer 1.core.F2 5′- TCAAAGAAAAACCAAACGTAACACCAACCG-3′ (nt 362–391 H77 and 2nd round antisense primer 1.NS3A4A.R2 5′- AGGTGCTCGTGACGACCTCCAGG-3′ (nt 5297–5319 H77); (2) 5′ quarter genome of genotype 2: 1st round sense primer 2.core.F1 5′- ATGAGCACAAATCCTAAACCTCAAAGA-3′ (nt 342–368 H77) and 1st round antisense primer 2.NS2.R1 5′- CCCCACACAATGACCTTCTTCTCCATTG -3′ (nt 5445–5467 H77), 2nd round sense primer 2.core.F2 5′- AATCCTAAACCTCAAAGAAAAACCAAA -3′ (nt 351–377 H77) and 2nd round antisense primer 2.NS2.R2 5′- GGGGAGAGGTGGTCATAGATGTAA -3′; (3) 5′ half genome of genotype 3: 1st round sense primer 3a.core.F1 5′- ATGAGCACACTTCCTAAACCTCAAAGA -3′ and 1st round antisense primer 3aNS3-R1V2 5′-TTACTTCCAGATCAGCTGACA-3′, 2nd round sense primer 3a.core.F2 5′- TCAAAGAAAAACCAAAAGAAACACCATCCG -3′ and 2nd round antisense primer PCR 3a.NS3-R2V2 5′-TTACTTCCAGATCAGCTGACA -3′. PCR was performed in MicroAmp 96-well reaction plates (Applied Biosystems, Foster City, CA) with the following PCR parameters: 1 cycle of 94°C for 2 min; 35 cycles of a denaturing step of 94°C for 15 s, an annealing step of 58°C for 30 s, an extension step of 68°C for 5 min, followed by a final extension of 68°C for 10 min. The product of the 1st round PCR was subsequently used as a template in the 2nd round PCR under same conditions but with a total of 45 cycles. Amplicons were inspected on precasted 1% agarose E-gels 96 (Invitrogen Life Technologies, Carlsbad, CA). All PCR procedures were carried out under PCR clean room conditions using procedural safeguards against sample contamination, including pre-aliquoting of all reagents, use of dedicated equipment, and physical separation of sample processing from pre- and post-PCR amplification steps.
PCR amplicons were directly sequenced by cycle-sequencing using BigDye terminator chemistry and protocols recommended by the manufacturer (Applied Biosystems; Foster City, CA). Sequencing reaction products were analyzed with an ABI 3730xl genetic analyzer (Applied Biosystems; Foster City, CA). Both DNA strands were sequenced using partially overlapping fragments. Individual sequence fragments for each amplicon were assembled and edited using the Sequencher program 5.0 (Gene Codes; Ann Arbor, MI). Inspection of individual chromatograms allowed for the identification of amplicons derived from single versus multiple templates. The absence of mixed bases at each nucleotide position throughout the entire 5′ half or quarter genome sequences was taken as evidence of amplification from a single viral RNA/cDNA template. This quality control measure enabled us to exclude from the analysis amplicons that resulted from PCR-generated in vitro recombination events or Taq polymerase errors and to obtain multiple individual sequences that proportionately represented those circulating HCV virions.
All the sequences alignments were initially made with ClustalW and then hand-checked using MacClade 4.08 to improve the alignments according to the codon translation. All 2922 5′ half or quarter-genome sequences from acute and chronic patients were deposited in GenBank and edited alignments can be accessed at http://www.hiv.lanl.gov/content/sequence/hiv/user_alignments/Li2012.html.
Two thousand three 5′ half genomes and 919 5′ quarter genomes were amplified by SGA-direct amplicon sequencing from the 31 subjects. Half genome sequences were generated so as to obtain longer sequences for linkage analyses, whereas quarter genome sequences were generated to enhance sensitivity of amplification from early samples with lower viral loads. Among the 3168 amplicons generated, the sequences of 2922 were unambiguous at every position. The other 246 amplicons contained one or more “double peaks” representing mixed bases and these were discarded and not included in the analysis. The median number of sequences analyzed per time point was 54 (mean = 54; range = 5–122) for acutely infected subjects and 25 (mean = 27; range = 13–44) for chronically infected subjects. A total of 2922 sequences from all 17 acutely-infected and 14 chronically-infected subjects were analyzed using phylogenetic tree analysis together with a sequence visualization tool, Highlighter (www.HIV.lanl.gov), that allows tracing of common ancestry between sequences based on individual nucleotide polymorphisms. Phylogenetic trees were generated by maximum likelihood methods using PhyML [85] or RAxML-VI-HPC [86]. For subjects productively infected by more than one T/F virus, lineages contained more than 5 closely related sequences were included in the lineage diversity analyses. Lineage-specific sequences were analyzed by the Poisson Fitter program (www.HIV.lanl.gov).
A total of 30 T/F lineages from 17 acutely-infected subjects were analyzed. Each sequence within the lineage was compared with the T/F virus sequence of that lineage. The insertion, deletion, transition and transversion frequencies were counted manually or by computer program. The rates were calculated by taking the ratio of each frequency number and the total number of nucleotides of all the sequences within that lineage.
The SNAP program (www.HIV.lanl.gov) was applied to the codon-aligned sequences of each T/F lineage. Within each lineage, the number of synonymous and nonsynonymous substitutions were derived by comparing to the T/F viral sequences. The accumulation rates of synonymous substitutions per potential synonymous site and nonsynonymous substitution per potential nonsynonymous site were compared to screen for positive selection.
Two mathematical models were used to analyze early HCV sequence diversity. The first was originally designed for HIV-1 and has been previously described in detail [15], [48], [49]. The second is a simplified deterministic model that accounts for the essential differences in replication dynamics between HCV and HIV-1, taking into account HCV's life cycle, that HCV replication occurs via a cytosolic replication complex, and that there can be many replication complexes continuously producing viruses from a long-lived infected cell. In this model, each HCV replication complex was assumed to give rise to a new replication complex at regular intervals by undergoing two RdRp copying events. These complexes, which may reach 40 per cell [53], were presumed to persist and produce viruses for the entire duration of the acute infection sampling period. Because of the sequential creation of the complexes, those at the same generation depth have widely varying number of descendants, unlike the situation in HIV (Figure 3). An average pair of HCV viruses then has a later most common ancestor than does HIV. As a result, the model predicts that sequences with a small number of shared mutations can arise in a subject at detectable frequencies prior to the onset of immune selection. This translates to an expectation of about 3 times larger numbers of stochastically shared mutations in HCV than in HIV, with the potential to violate the star-like hypothesis more often. Furthermore, the persistence of HCV replication complexes of all generations means that, in contrast to the situation in HIV, the Poisson distribution is not necessarily a good model for inter-sequence distances prior to selection, and sequences with distances that deviate from Poisson distribution canbe derived from a single T/F viral genome. Our simple model of HCV diversification accounts for these issues and is amenable to an analytical approach. At the same time, the numerical results from the simplified model are consistent with those from a more detailed agent based stochastic model of early HCV infection that we present in an accompanying paper [50].
The implementation of the clustering algorithm works on a phylogenetic tree describing the evolutionary relations between these sequences and aims to identify monophyletic clusters that could reasonably have arisen by evolution in the infected individual. The modeling in Ribeiro et al. [50] shows that these clusters should satisfy two separate criteria: (a) the total number of mutations that could have accumulated is limited by the mutation rate of the virus and the generation time, and (b) the number of mutations shared by distinct sequences from a single T/F virus is related by coalescent theory to the growth and stabilization of viral load in these acute infections. Starting at the tips of the phylogenetic tree, our algorithm identifies the largest clusters that are consistent with these two criteria. Following from (a), every replication complex present in the body produces a new generation of replication complexes about once a day. Following from this, at time t days into infection, the most divergent replication complex is t generations from the founder, but most replication complexes are of generation t/2. We therefore split the clusters whose average divergence is larger than expected from this scenario. Following from (b), if one samples approximately 30–100 sequences in the initial weeks of infection prior to the onset of immune selection, 1–10 pairs of sequences are expected to coalesce 4–7 generations after the T/F virus.Assuming a mutation rate of 2.5×10−5 per base per generation [50], this corresponds to a probability of more than 5% of seeing 3–4 shared mutations in 5000 bases. Thus, to be conservative in our model estimates of minimum numbers of T/F genomes, wedefined clusters of sequences as having >4 shared mutations in a half genome, or >2 in a quarter genome, as unambiguously to have arisen from a different founder variant.
In the case of acute-to-acute infections, however, the average divergence is a very weak measure for delineating the clusters. In particular, transmission of one or a few highly divergent sequences affects the average little, and they are often not identified as separate variants when they are likely to represent a distinct founder. Thus, mean diversity is a robust measure but occasionally allows one or a few highly divergent sequences to be counted as a part of the cluster. To provide an estimate of the number of T/F variants that takes this into account, we also calculated the number of clusters by implementing a threshold on the most distant tip from the inferred ancestor for each cluster, splitting clusters based on the maximum distance that is improbable. Though this provides a better overall estimate for the number of clusters, such extreme-value statistics are more affected by the approximations made in going from the fully stochastic to the deterministic model. We verified that the main conclusions of this paper, including the identification of three subjects as cases of acute-to-acute transmission, follow from either of the two methods of identifying T/F lineages. Both of these methods provide a minimum estimate for the number of T/F viruses for two reasons. First, the number of shared mutations we have allowed within a cluster provides a highly conservative criterion and lineages could still contain closely related but distinct transmitted viral sequences, a situation made likely by our identification of probable cases of acute-to-acute virus transmission. Second, finite depth of sequencing means we might miss T/F clusters that are represented by a small fraction of sequences. To estimate the impact of this second scenario, we performed direct power calculations as previously described [15] to assess the likelihood of missing infrequent T/F sequence lineages based on sampling depth.
Finally, acute samples are often very homogeneous and can have conflicting phylogenetic signals, and these calculations rely on the phylogenetic tree being a description of the true evolutionary relations. In cases of infection with two highly divergent strains, the descendants of the two strains form distant outgroups from each other, and homoplasy on the long branch linking the two may root the individual clusters suboptimally. To avoid this problem, our algorithm looks for long branches (longer than 30 mutations), deletes them, and applies the rest of the algorithm to each of the then disconnected clusters.After identifying these clusters, the roots of each of these is chosen as tentative transmitted founder viruses. The paths in the phylogenetic tree between these roots represent evolution that happened prior to infection, whereas the paths linking the roots to the tips indicate within host evolution. Occasionally, we find overlap between the two sets of paths indicating convergent evolution between the donor and the recipient infected individuals. Since convergent evolution is unlikely, we further subdivide the previously found clusters to avoid this.After finding the optimal clustering at each time point, the results are mapped on to a tree of all the sequences obtained at various time points from the same individual. The overall number of clusters is then found by merging clusters from different time points that interleave, and leaving the other clusters as distinct. In the scenario of early infection when there are very few mutations, the tree can be ambiguous because of single conflicting mutations causing polytomies; while this could theoretically make the number of transmitted variants ambiguous, this was not the case in our analysis of this sequence set.Codes written in C that implement these two clustering strategies (using either maximum or average distances to define clusters) are available at http://www.santafe.edu/~tanmoy/programs/HCV/.
Standard descriptive statistics including Mann-Whitney and unpaired T-tests with Welch's correction were employed and identified throughout the text. Power calculations to estimate the likelihood of detecting rare sequence variants based on sampling depth were performed as previously described [15]. Estimates of the probability that observed clusters of nonsynonymous mutations in the Env E2 coding region of putative T/F viral genomes from subjects 10003, 10016 and 10020 could have occurred by chance were performed with a binomial expansion as previously reported [87]. We considered ten codon windows because this could generally span the observed clusters as well as typical T-cell epitopes and potential linear neutralizing antibody epitopes. From the binomial expansion, the probability of seeing at least the observed number (k) of clustered mutations within a single 10-mer is:where r is the number of potential 10-mers, m is the total number of mutations relative to the consensus, and p is the probability that a mutation falls by chance in any particular 10-mer. We calculated the probability only for k′ = k, since the results for all k′>k are very much smaller and can be ignored. Only nonsynonymous mutations were included in the analysis. Within the 10-mers of interest, but not elsewhere along the alignment, different amino acid substitutions and combinations of substitutions were identified and analyzed.
|
10.1371/journal.pntd.0006500 | Deworming in pre-school age children: A global empirical analysis of health outcomes | There is debate over the effectiveness of deworming children against soil-transmitted helminthiasis (STH) to improve health outcomes, and current evidence may be limited in study design and generalizability. However, programmatic deworming continues throughout low and middle-income countries.
We performed an empirical evaluation of the relationship between deworming in pre-school age children (ages 1–4 years) within the previous 6 months, as proxy-reported by the mother, and health outcomes of weight, height, and hemoglobin. We used nationally representative cross-sectional data from 45 countries using the Demographic and Health Surveys (DHS) during the period 2005–2016. We used logistic regression with coarsened exact matching, fixed effects for survey and year, and person-level covariates. We included data on 325,115 children in 45 STH-endemic countries from 66 DHS surveys. Globally in STH-endemic countries, children who received deworming treatment were less likely to be stunted (1.2 percentage point decline from mean of 36%; 95% CI [-1.9, -0.5%]; p<0.001), but we did not detect consistent associations between deworming and anemia or weight. In sub-Saharan Africa, we found that children who received deworming treatment were less likely to be stunted (1.1 percentage point decline from mean of 36%; 95% CI [-2.1, -0.2%]; p = 0.01) and less likely to have anemia (1.8 percentage point decline from mean of 58%; 95% CI [-2.8, -0.7%]; p<0.001), but we did not detect consistent associations between deworming and weight. These findings were robust across multiple statistical models, and we did not find consistently measurable associations with data from non-endemic settings.
Among pre-school age children, we detected a robust and consistent association between deworming and reduced stunting, with additional evidence for reduced anemia in sub-Saharan Africa. We did not find a consistent relationship between deworming and improved weight. This global empirical analysis provides evidence to support the deworming of pre-school age children.
| Soil-transmitted helminths are parasitic worms that affect 1.5 billion people. The global public health strategy is regular deworming of children in endemic settings. In this study, we bring an independent dataset with over 320,000 pre-school age children across 45 STH-endemic countries to investigate the relationship between deworming and health outcomes of weight, height, and anemia. We address potential limitations of past studies by examining across diverse settings for improved generalizability, increasing sample size, and focusing on children who received treatment to improve statistical power. We find that deworming is robustly associated with reduced stunting in pre-school age children and find additional evidence that deworming may be associated with reduced anemia in sub-Saharan Africa. Deworming was not consistently associated with improved weight. These findings were robust across multiple statistical models, and we did not find consistently measurable associations with data from non-endemic settings, which improves confidence in the study findings. This global empirical analysis provides evidence to support the continued deworming of pre-school age children in endemic settings.
| The large-scale empiric treatment of children for soil-transmitted helminthiasis (STH), commonly known as “deworming”, has become controversial [1–4]. This strategy of regular deworming (also known as ‘preventive chemotherapy’) with albendazole or mebendazole has been the main STH control strategy recommended by the World Health Organization (WHO) for over a decade [5]. While deworming continues to be implemented throughout low and middle-income countries for the estimated 1.5 billion people with STH (caused by infection with Ascaris lumbricoides, hookworm species of Ancylostoma duodenale and Necator americanus, and Trichuris trichiura), the evidence linking real-world experience of deworming with health outcomes is mixed [6].
Some randomized-controlled trials have demonstrated possible positive benefits of deworming on child health, primarily focused on weight or height, although many trials have not found these relations [3, 7–10]. Observational studies provide evidence for an association between deworming and cognitive improvements, physical growth, and reduced anemia [11–13]. However, many of these benefits have been disputed by recent studies [3, 4, 14, 15]. The Cochrane Collaboration (2015) and Campbell Collaboration (2017) systematic reviews and meta-analyses of randomized trials concluded deworming does not result in improved health outcomes, although they did find some benefit to treating those known to be infected [3, 4]. This was followed by substantial debate on the potential limitations of randomized trials and the Cochrane and Campbell meta-analyses [1, 2, 16–18]. In particular, because a minority of individuals harbors a majority of the worm burden, critics argued that these studies were underpowered to detect a meaningful effect, since only a fraction of the population benefits from treatment in areas of low or even moderate parasite burden. Furthermore, low STH prevalence in some clinical trials, heterogeneity in helminth species and their respective health effects, and the short time period of most trials may obscure potential health benefits. Uncertainty about the benefits of deworming is compounded by clustering of existing randomized trials in a few regions, which may limit generalizability to different contexts.
Periodic deworming of pre-school age children (ages 1–4 years) has been recommended by the WHO in areas where prevalence is ≥20% by stool microscopy since the 2006 WHO guidelines [5]. The strategy is intended to reduce the number of children with moderate and high-intensity infection, which is associated with adverse health outcomes [5]. Since pre-school age children (ages 1–4 years) are not typically reached through school-based deworming programs, they more commonly receive treatment through Child Health days [19], community-based deworming for lymphatic filariasis, or though the healthcare system. Recent modeling studies also have highlighted the importance of treating this population in order to reduce overall transmission and improve the cost-effectiveness of deworming programs [20–23]. From a measurement standpoint, this age group may enable more sensitive detection of health, educational, and economic benefits in response to deworming than the school-age population, which is more often studied. This is because periodic deworming in pre-school age children can occur before significant cumulative exposure to STH and continues throughout a period of rapid growth and development [10, 24].
To better understand the relationship between deworming and health in pre-school age children, we performed an empirical analysis using data from the Demographic and Health Surveys (DHS) across 45 countries. We addressed several limitations of past approaches in terms of generalizability, and draw upon a decade of deworming to measure population-level relationships between deworming and health.
We examined the relationship between reported deworming and the pre-specified health outcomes of malnutrition (underweight, stunting) and anemia among pre-school age children (ages 1–4 years, selected based on WHO classification of this age group) in 45 low- and middle-income STH-endemic countries in Africa, the Americas, Asia, and Europe using cross-sectional data during the period 2005–2016. We also performed an analysis for sub-Saharan Africa alone given the substantial STH burden and scale of ongoing deworming [25].
We used the DHS to examine the relationship between deworming exposure and health outcomes using cross-sectional data. These nationally representative surveys are conducted approximately every 5 years in many low- and middle-income countries in collaboration with in-country partners [26]. We used all available DHS surveys that included data on deworming exposure, health outcomes, and relevant covariates from endemic countries based on WHO classification (Appendix). DHS did not provide person-level information on STH infection status. We included all surveys from eligible countries in order to maximize sample size and examine temporal trends, and selected only children ages 1–4 years. We excluded countries with low-burden STH (defined as countries in which <10% children lived in areas where STH was endemic and WHO recommends mass deworming) or those missing data on deworming, health outcomes, or relevant covariates (see Appendix). A table of excluded surveys is available in the Appendix. We used the WHO preventive chemotherapy database to estimate country-level disease burden for STH based on proportion of children requiring deworming [27]. All data files are available from the DHS database online. This study relied on published data and did not require human subject research approval.
We defined the study exposure using the mother’s report of whether her children received “drugs for intestinal parasites in the 6 months prior to the survey”. We examined three pre-specified binary health outcomes for pre-school age children: underweight, stunting, and anemia, which are the relevant health outcomes available in DHS surveys for this age group. We defined being underweight as a weight-for-age two standard deviations below the WHO global reference standard, stunting as a height-for-age two standard deviations below the WHO global reference standard, and anemia as a hemoglobin below 11 g/dL [28]. We chose to dichotomize health outcome variables in the main analysis to isolate the disease state and to account for possible confounding trends (e.g. increasing obesity), although we also conducted the analysis with continuous outcome variables. We included pre-specified person-level covariates in the analysis to address potential confounding relations, including child’s age and gender, mother’s age (<30 years or older based on prior DHS analysis on health-seeking behavior [29]) and education (defined as no education, some primary school, and completion of primary school or higher education), wealth quintile (a composite measure for each person’s household, relative based on country standard), place of residence (urban or rural, relative based on country’s definition), and binary variables indicating access to improved water, sanitation (based on toilet facility using the Joint Monitoring Programme definitions) [30], and healthcare-seeking behaviors using receipt of third dose of diphtheria-tetanus-pertussis vaccine as an indirect proxy (Appendix).
The primary analyses used three health outcomes (underweight, stunting, or anemia) as dependent variables in logistic regression models where the primary independent variable was an indicator of receipt of deworming. We pre-processed our sample to improve balance between the treated and untreated groups using coarsened exact matching (CEM) within each DHS survey [31]. We used CEM over the more common propensity score matching to relax concerns over balance (i.e., CEM creates strata to ensure that treatment and untreated groups are matched on all covariates, rather than balanced on a propensity score alone), improve the intuitive understanding of the matching process, and because CEM was found to out-perform propensity score matching in balancing groups [32, 33] (see Appendix).
In all models we also used survey fixed effects, implemented as a set of indicator variables that control for all invariant differences between surveys (e.g. level of economic development and national STH prevalence at time of survey). Notably, these fixed effects control for unobserved sub-national (regional) differences in wealth and access to services or nutritional programs that would be related to health outcome, and are in addition to person-level covariates. We also included year fixed effects that control for common time effects across all surveys. Matched models using CEM do not include covariates in regression since they are pre-processed for balance. We estimated the odds ratios (OR) for a given health outcome among children that were dewormed relative to those that were not. We also reported the marginal effects, which represent the change in the probability of the outcome among those exposed to deworming relative to those that were not. We used robust standard errors clustered by country and survey. We also estimated the possible impact of providing deworming in the sub-Saharan African study countries by computing the number of avertable cases of disease states that could be related to deworming for better conceptualization of magnitude of the study results (see Appendix). The analytic code is available on request to the corresponding author.
We conducted sensitivity analyses to examine the stability of our findings to alternative statistical models and assumptions on the data. We repeated the main analysis with alternative model specifications (e.g. ordinary least squares model with continuous outcomes), additional covariates (e.g. vitamin A receipt, use of malaria net, and breastfeeding), and country-specific analyses (see Appendix). We tested region (sub-national) fixed effects to match children within region. We defined relationship robustness in reference to these alternative statistical models.
We performed additional analyses including dose-response, negative controls, and sub-group analyses. The analysis tested stratification by age and estimated STH disease burden, including formal tests of interaction, to estimate dose response. We tested for the possibility of unobserved confounding by testing “negative controls” (factors theoretically unrelated to exposure and outcome, but possibly confounded by those same unobserved factors); we tested negative control exposures unrelated to deworming (e.g., “heard of family planning on radio” and “access to condoms”) and negative control outcomes (e.g., “cough in last two weeks” and “fever in last two weeks”) (see Appendix). We defined a relationship between deworming and health outcomes as “consistent” when the negative controls did not hold comparable associations to the primary study findings. We also examined the relationship between deworming and health outcomes in the sub-group of disease state (e.g. underweight children only). The Appendix contains additional details.
This study included 325,115 pre-school age children from 45 countries in Africa, the Americas, Asia, and Europe from 2005 to 2016 (Table 1). This included a total of 66 DHS surveys, and excluded an additional 26 surveys for lack of data or endemicity for STH (see Appendix, Section 1). Over this time period, the mean proportion of respondents that indicated their pre-school age child received deworming medicine in the previous six months was 43.4%. The country-level deworming coverage was highly variable, ranging from 6.1% (Azerbaijan) to 87.4% (Rwanda), which may partly reflect the proportion of “at-risk” children within the country (Fig 1; Table A1). The overall prevalence of underweight, stunting, and anemia was 29.5%, 35.5%, and 53.0%, respectively. We found that children who received deworming were slightly older, wealthier, more likely to live in urban areas, had mothers that were more educated, had stronger health-seeking behavior, and greater access to toilets (Table 2). The CEM matched dewormed and untreated groups on all specified covariates (Table 2).
Globally, in the matched analysis, children who received deworming treatment were less likely to be stunted (1.2 percentage decline from mean of 36%; 95%CI [-1.9, -0.5%]; OR:0.92; p<0.001). This association was robust and mostly consistent, meaning this relationship was not observed in the negative-exposure controls, relaxing concern for any residual confounding (Table 3, A4, A5). We also found that children who received deworming treatment were less likely to be anemic (1.4 percentage decline from mean of 53%; 95%CI [-2.2, -0.6%]; OR:0.93; p<0.001), but while this finding was robust across specifications, it was also observed in negative-exposure controls suggesting the possibility of residual confounding. We did not measure a robust or consistent relationship between deworming and being underweight in the overall analysis (0.7 percentage decline from mean of 29%; 95%CI [-1.7, 0.2%]; OR:0.94; p = 0.06), but did measure a robust and consistent relationship between deworming and higher weight in the sub-group of underweight children (0.02 standard deviation increase; 95%CI [0.01, 0.04]; p<0.01).
In sub-Saharan Africa, in the matched analysis of 32 countries, we found that children who received deworming treatment were less likely to be stunted (1.1 percentage decline from mean of 36%; 95%CI [-2.1, -0.2%]; OR: 0.94; p = 0.01) and less likely to be anemic (1.8 percentage decline from mean of 58%; 95%CI [-2.8, -0.7%]; OR:0.91; p<0.001) (Table 3). These findings were robust across the majority of alternative statistical models and mostly consistent in overall negative-exposure controls (Table A4, A5). We did not detect consistent associations between deworming and reduced risk of being underweight in the overall analysis, but did find a relationship in the sub-group of underweight children (0.02 standard deviation increase; 95%CI [0.00, 0.05]; p = 0.03). To conceptualize the magnitude of these effect sizes, we estimated that expanding deworming to 100% coverage could be related to 694,100 (95%UI: 134,300–1,298,700) averted cases of stunting and 850,600 (95%UI: 325,300–1,326,000) averted cases of anemia in pre-school age children from study countries in sub-Saharan Africa that are not yet being dewormed due to low coverage (Figure A4; Table A9).
The primary findings reported as robust were supported across the majority of alternative statistical models (Table A4). In some cases, a finding was not consistent because the negative-exposure control was related to the health outcome (e.g. “access to condoms” was associated with less anemia in the global analysis) (Table A5). Deworming was associated with more reports of recent cough and fever as outcomes (see Appendix), which suggests some unexplained relationship, but argues against selection bias of better health in the population that received deworming. When matching with addition of covariate for receipt of vitamin A, use of malaria net, and breastfeeding, the main study finding remained broadly consistent; we did not find evidence for interaction between deworming and vitamin A. We repeated the main analysis restricting to only one child per household (i.e. not including additional children in the household) and found our findings to be broadly consistent.
We tested for dose-response effects through stratification by STH disease burden and age and testing continuous interaction; we found increasing effect size at older ages for the relationship between deworming and reduced stunting but did not find any other consistent relations (Table A6-A7). We did not find any clear evidence for larger effect sizes in higher STH disease burden settings (Table A7). Notably, we did not find any relationships between deworming and better health outcomes using data from non-endemic settings (country list in Appendix), which increases our confidence in our findings since we would not expect deworming benefits in these regions (Table A8). There was a range of effect sizes in the country-specific analyses (Figure A1-3).
Drawing upon data from more than 320,000 pre-school age children in 45 countries, this empirical analysis provides evidence for a global-level association between deworming and reduced stunting, with additional evidence for reduced anemia in sub-Saharan Africa for pre-school age children. We do not find a consistent relationship between deworming and improved weight. While the evidence to support the benefit of mass deworming has recently been questioned [1–4], this study aims to bring new data and study design to better understand the relationship between deworming and health outcomes in the population of pre-school age children. We addressed potential limitations of past studies by examining across diverse settings for generalizability and increasing sample size and focusing on children who received treatment to improve statistical power. Importantly, while a mother’s recall of deworming may be imperfect, any recall bias is unlikely to be unique to the dewormed group so this would bias our overall study findings towards no association and would not be a threat to the validity of our findings. The study results from our global empirical analysis provide evidence to support the deworming of pre-school age children.
Pre-school age children remain a key population for STH burden, because of their potential for long-term health and educational consequences, contribution to ongoing transmission, and need for treatment outside of traditional school-based deworming [5, 21]. Over the past 10 years, treatment in this age group has scaled up across many low and middle-income countries [27]. WHO currently estimates that global deworming has increased substantially to achieve a coverage of 51% of pre-school age children worldwide, although these estimates are only from countries that report this data [25]. Overall, the WHO estimate corresponds well with our estimate that 43% of pre-school age children had received deworming, which lends validity to the proxy-reported deworming exposure. However, we noted substantial within-country heterogeneity for deworming coverage, implying that some regions are not being reached despite a moderately high overall deworming coverage.
In our study, we measured a modest reduced risk of stunting among pre-school age children with a 1.1% decline from 36% prevalence–although this modest association corresponds with an estimated 694,100 cases of stunting in pre-school age children in our study countries from sub-Saharan Africa. This association should be conceptualized as the “diluted” population-level impact. In reality, the benefit of deworming is likely greatest among those children with high parasite burden and is intuitively null among those who are not infected. Therefore, the population-level effect is relatively modest. While stunting is considered to be a chronic indicator of health and may not be expected to change within the relatively short 6-month timeframe of our deworming exposure, reported receipt of deworming is likely correlated with previous deworming so this exposure may represent sustained periodic deworming. Our finding that older children had an increased association for improved height supports this hypothesis that larger benefits may be observed in repeatedly treated children.
We measured a consistent relationship between deworming and reduced risk of anemia (1.8% lower absolute percentage) in sub-Saharan Africa, which corresponds with an estimated 850,600 anemic pre-school age children in our study countries that could benefit from deworming. This relationship was not observed in the global analysis, driven by a non-association in South American countries. The disease burden of hookworm is likely higher in sub-Saharan Africa meaning a higher burden of avertable anemia, where we measured a positive association, than in South America, where we did not [34, 35].
We did not find a robust association between deworming and a reduction in being underweight in the primary analysis, and the negative-exposure control analysis suggested the possibility for residual confounding. However, there was an association between deworming and lower risk of being underweight in children who were already underweight. This suggests the possibility that deworming may associated with weight gain in children who are already underweight, but that the association is diluted out when including all children. A recent systematic review and meta-analysis of school-based deworming programs found positive effects of deworming on weight [16], while other meta-analyses failed to detect a measurable effect, although these have often focused on older children [3, 4].
Our study findings differ from the conclusions of recent meta-analyses of randomized trials, which failed to detect measurable improvements to height, weight, or anemia from deworming at a population-level [3, 4]. Notably, these meta-analyses have found that deworming may improve weight among those who are known to be infected but not at a population-level, which may be explained by a dilution effect [3]. A key difference in our analysis was a focus on pre-school age children, and measurement of only treated individuals rather than an overall population to increase statistical power (since overall population would dilute a relationship by including children who were not treated). A principal feature of this study was its large sample size and adequate statistical power. We included 66 national surveys with 325,000 children to ensure we could detect possible benefit despite the dilution effect (including non-infected and lightly infected children not expected to benefit from treatment). The discrepancy between our study findings and those of previous studies may further be explained by effect heterogeneity that is driven by spatial distribution of disease, differences between worm species, and an important non-linear relationship between disease burden (based on community infection intensity) and health outcomes [17, 18]. Our results find substantial country heterogeneity in the associations between deworming and health outcomes, which further supports the complexities in measuring the impact of deworming and the importance of evaluating real world deworming across diverse settings and epidemiologic conditions. While our deworming treatment record is reported for the previous 6 months, these children appear more likely to have received deworming treatments repeatedly, as recommended by WHO. Therefore, repeated periodic deworming, which consistently reduces worm burden, is more likely to result in long-term health-related outcomes and may also include treatment with different drugs (e.g. albendazole vs mebendazole), multiple treatments, and drugs for other parasitic worms (e.g. praziquantel for Schistosoma spp). Notably, this survey question may not have included treatment given for other infections (e.g. lymphatic filariasis) that would also be effective against STH infections, and this would bias our findings towards no association. We did not find a relationship between increasing country-level STH burden and higher effect sizes (i.e., dose response), although the aggregated country-level nature of the STH burden estimate may obscure a potential relationship. We did not perform a sub-analysis for Asia alone because data were limited.
The findings of this analysis should be interpreted within the context of the study design and constraints of the data. While we controlled for many key differences between the dewormed and control group through the CEM process to adjust for observed differences and through survey fixed effects to adjust for unobserved heterogeneities, residual unobserved differences may not be accounted for in the model (e.g. women who reported deworming may also have different health seeking behavior). We addressed this potential residual confounding by testing negative control exposures and outcomes. Negative-exposure controls that were associated with the health outcome raised concerns about selection bias that could not be eliminated with the observed data, even after matching; in these cases, the primary relationships were deemed inconsistent. While we measured an incompletely explained relationship between deworming and outcomes of increased frequency of reported fever and cough in the previous two weeks, this would bias against our primary study finding since these populations would not be predisposed to better health for unobserved reasons. Furthermore, we repeated the analysis with data from 10 non-endemic and low prevalence countries and found no consistently measurable associations between deworming and health outcomes, which may relax some concern for residual confounding since no relationship would be expected in these regions while similar confounding pathways may still be present. Similarly, while the main exposure variable for deworming of a child was proxy-reported by his or her mother and potentially subject to reporting bias (e.g. increased health knowledge) or recall bias, this would be more likely to bias our findings towards null unless the bias was unique and differential in the deworming group. Importantly, we did not have data on person-level STH infection status, although since mass deworming programs are also being implemented without knowledge of individual infection status, our goal was to quantify the population-level effectiveness of these real-world programs. The substantial overdispersion of the parasite and dilution effect of including all children (the majority uninfected or lightly infected) would significantly push our findings towards null. Since unprogrammed receipt of albendazole is common (e.g. local pharmacy, healthcare facility), it is likely that reported deworming included both programmed (e.g. mass drug administration, child health days) and unprogrammed sources [36, 37]. We also tested including vitamin A receipt as a variable or effect modifier in the model for sensitivity analysis, since vitamin A is often delivered alongside albendazole or mebendazole to pre-school age children during child health days, although our primary analyses remained robust even when considering vitamin A. Our analysis focused on pre-school age children, but future work should also examine school-age children and the broader community that may play a critical role in ongoing transmission [21, 23]. Finally, the estimates for potentially avertable cases of stunting and anemia related to deworming have substantial uncertainty, but are provided for better conceptualization of the magnitude of the study estimates and are not meant to imply causality.
This study found that deworming pre-school age children against STH is related to global reductions in risk of stunting, and evidence for reduced anemia in sub-Saharan Africa. Given recent controversy on deworming against STH, we provide an observational study design grounded in individual-level survey data that overcomes challenges in data and generalizability of prior work, and supports the continued deworming of pre-school age children.
|
10.1371/journal.pcbi.1004562 | A Kinetic Platform to Determine the Fate of Hydrogen Peroxide in Escherichia coli | Hydrogen peroxide (H2O2) is used by phagocytic cells of the innate immune response to kill engulfed bacteria. H2O2 diffuses freely into bacteria, where it can wreak havoc on sensitive biomolecules if it is not rapidly detoxified. Accordingly, bacteria have evolved numerous systems to defend themselves against H2O2, and the importance of these systems to pathogenesis has been substantiated by the many bacteria that require them to establish or sustain infections. The kinetic competition for H2O2 within bacteria is complex, which suggests that quantitative models will improve interpretation and prediction of network behavior. To date, such models have been of limited scope, and this inspired us to construct a quantitative, systems-level model of H2O2 detoxification in Escherichia coli that includes detoxification enzymes, H2O2-dependent transcriptional regulation, enzyme degradation, the Fenton reaction and damage caused by •OH, oxidation of biomolecules by H2O2, and repair processes. After using an iterative computational and experimental procedure to train the model, we leveraged it to predict how H2O2 detoxification would change in response to an environmental perturbation that pathogens encounter within host phagosomes, carbon source deprivation, which leads to translational inhibition and limited availability of NADH. We found that the model accurately predicted that NADH depletion would delay clearance at low H2O2 concentrations and that detoxification at higher concentrations would resemble that of carbon-replete conditions. These results suggest that protein synthesis during bolus H2O2 stress does not affect clearance dynamics and that access to catabolites only matters at low H2O2 concentrations. We anticipate that this model will serve as a computational tool for the quantitative exploration and dissection of oxidative stress in bacteria, and that the model and methods used to develop it will provide important templates for the generation of comparable models for other bacterial species.
| Bacterial hydrogen peroxide (H2O2) response networks contain essential virulence factors for a number of pathogens. Without these systems, infecting bacteria fall prey to host immune cells and cannot establish or sustain an infection. The reaction networks and regulatory features involved are complex, which suggests that computational modeling would facilitate quantitative dissection and analysis of these systems. However, current models of H2O2 reaction networks have been of limited scope. Here, we constructed a systems-level H2O2 detoxification model for Escherichia coli, and used it to understand how the network responds to different H2O2 concentrations and insults. We anticipate that this model and comparable ones for other species that are facilitated by its construction will be useful in identifying and understanding methods to sensitize pathogens to immune attack. Such strategies hold great promise for the development of next generation antibiotics, since agents that impair oxidative stress defense systems would focus selective pressure to infection sites, and therefore exhibit slow resistance development and little impact on commensal bacteria.
| Reactive oxygen species (ROS) are critical immune antimicrobials used in the first line of defense against infections, where phagocytic cells of the innate immune response use NADPH oxidase to generate an “oxidative burst” of superoxide (O2−•) after engulfing pathogens in a phagosome [1, 2]. The O2•− can then be dismutated to H2O2 [2], which readily diffuses across the bacterial membrane where it damages sensitive biomolecules, reacts with ferrous iron to produce the highly deleterious •OH [3], or is detoxified by specialized enzymes. The importance of the oxidative burst to immunity is highlighted by the incidence of recurring infections within and shortened life expectancy of patients with defects in NADPH oxidase, a condition known as chronic granulomatous disease (CGD) [4]. In addition, many pathogens including Bacillus anthracis [5], Coxiella burnetti [6], Chlamydia trachomatis (serovars E, K, and L2) [7], Salmonella enterica (serovar Typhimurium) [8], Mycobacterium tuberculosis [9, 10], Staphylococcus aureus [11], Helicobacter pylori [12], Streptococcus pyogenes [13], and Enterococcus faecalis [14] require H2O2 defense systems to establish or sustain infections. Interestingly, beyond its use by immune cells, bacteria also use H2O2 against each other, such as when Streptococcus pneumoniae stimulates prophage induction and cell death in Staphylococcus aureus by generating H2O2 during niche competitions [15].
Accordingly, bacteria have evolved various pathways to detoxify H2O2. While the importance of these H2O2 detoxification systems has been established [16], there are gaps in knowledge regarding the kinetic interplay between them under different conditions. Escherichia coli K-12 encodes one alkyl hydroperoxidase (AHP) and two separate catalases for detoxifying H2O2, which differ in regulation and/or reaction mechanism. AHP and catalase HPI expression are induced by OxyR during oxidative stress, whereas catalase HPII expression is up-regulated in stationary phase and does not increase in the presence of H2O2 [17–19]. AHP requires one molecule of NADH per reaction cycle, coupling the rate of detoxification achievable by this enzyme to catabolism, whereas H2O2 is the only substrate in the catalase reaction cycle. AHP has been shown to act as the primary scavenger of endogenously produced H2O2, and is efficient at detoxifying low concentrations of H2O2 (<20 μM), whereas catalase is known to dominate clearance at higher concentrations (>50 μM) [20, 21]. Since the result of H2O2 exposure whether that be bacteriostasis, mutagenesis, cell death, or continued growth depends on a kinetic competition for the molecule, it is important to have a quantitative, systems-level understanding of its biochemical reaction network. Due to the complexity of H2O2 biochemical reaction networks, computational models are necessary for interpretation of H2O2 detoxification data and prediction of system behavior.
As a result of its importance as a signaling molecule, the most complete models of H2O2 metabolism currently available were developed for mammalian systems [22–24]. They have included H2O2 elimination by antioxidants (e.g., glutathione and thioredoxin) and enzymes (e.g., catalase, glutathione peroxidase, glutathione reductase, glutaredoxin, and peroxiredoxin), and processing of oxidized protein thiols [22–24]. However, these models were specific to mammalian physiology and did not include transcriptional regulation, enzyme degradation, side reactions of H2O2 with sensitive biomolecules such as methionine and pyruvate, or the related reactive oxygen species O2−• and •OH and their associated reactivity (e.g., •OH rapidly oxidizes all twenty amino acids and glutathione). Although models equivalent to those of mammalian systems have yet to be described for bacteria, there has been progress in modeling subsystems of the H2O2 response network under H2O2 stress, such as the thioredoxin system in E. coli [25].
Here, we have generated a kinetic model of H2O2 stress in E. coli whose components are depicted in Fig 1. The biochemical reaction network is compartmentalized into media and intracellular spaces, includes spontaneous and enzymatic detoxification of H2O2, transcriptional regulation and inactivation of detoxification enzymes, and reactions of H2O2 and its degradation intermediates (e.g., •OH) with biomolecules (e.g., pyruvate, glutathione, all twenty amino acids). Parameters were informed from literature or trained using an iterative and integrated computational and experimental approach (Fig 2). The design criteria we chose to use to develop the model stipulated that consistent discrimination between clearance contributions by the major detoxification systems (AHP, HPI, and HPII) needed to be achieved. Once the design criteria were met, remaining parametric uncertainty was accounted for with use of a Markov chain Monte Carlo (MCMC) procedure to explore the viable parameter space and assemble an ensemble of models that performed comparably well with the training data [26]. The ensemble was then used to quantitatively investigate the importance of carbon availability and translation to H2O2 detoxification, and its predictions were experimentally confirmed.
Our aim was to construct a systems-level kinetic model of H2O2 detoxification in E. coli that could provide consistent predictions of H2O2 distributions among its different detoxification pathways after exposure to a range of initial H2O2 boluses. To accomplish this goal in the most efficient way possible, we adopted the systematic approach shown in Fig 2. Briefly, we began with a minimal number of experiments, wild-type clearance of different initial H2O2 concentrations. After optimizing uncertain parameters, we selected models based on their relative likelihood, also referred to as their evidence ratio (ER) [27–32], discarding models more than ten times less likely than the most-likely model in our set (ER≥10). If the acceptable models did not uniformly attribute H2O2 detoxification to the same pathways, we performed simulations to suggest experiments that could resolve the disagreement. Those experiments were then performed, and data used to arrive at updated parameter estimates. This process was continued until we arrived at a model or set of models that rendered consistent H2O2 distributions. Since some parameters may not have been important to H2O2 clearance under the conditions used here, and therefore, unlikely to be informed by the training procedure, we explored the parameter space using a previously developed MCMC procedure [26] to assemble an ensemble of parameter sets that could all describe the H2O2 clearance data comparably well (ER≤10). In this way, we could ensure that forward predictions were not dependent on ill-defined parameters. We note that this procedure also accounts for cases in which parameter pairings or more complex relationships rather than absolute values are important by varying all parameters simultaneously when walking away from known viable points. Also, before proceeding to forward predictions, we confirmed that all the models within the ensemble still satisfied the design criteria.
We constructed a compartmentalized reaction network that includes spontaneous and enzymatic reactions present in an E. coli culture under H2O2 stress, transcriptional regulation of AHP and HPI, and degradation/inactivation of the major detoxification enzymes AHP, HPI, and HPII. Uncertainty exists with regard to the dynamics of enzyme degradation/inactivation in the presence of H2O2, as well as the possibility of an H2O2 gradient across the membrane. Specifically, the enzymes could be degraded or inactivated in an H2O2-independent manner, either with a fixed degradation constant [33, 34], or optimized to account for the varying degradation rates of different proteins [35, 36]. Alternatively, the H2O2 detoxification enzymes could be poisoned by their own substrate [37–39], following bimolecular [40] or more complex kinetics [37].
In addition to the indeterminacy in degradation/inactivation kinetics, there is evidence supporting [41] and opposing [42] the presence of an H2O2 gradient across the cell membrane. Models accounting for these various possibilities are presented in Table 1, along with their corresponding number of uncertain parameters. The introduction of unknown parameters has the potential to improve agreement between model simulations and experimental data solely by increasing the flexibility of the model. For this reason, we calculated the relative likelihood of models, otherwise known as their evidence ratio (ER) [27–32] based on their respective Akaike Information Criterion (AIC), which is a commonly used statistical metric that weighs goodness of fit against model complexity when discriminating between competing models [27, 43–45]. Models with a relative likelihood of ten times less than the best model in the set (ER≤10) were considered acceptable, whereas others were discarded.
When optimizing parameters simultaneously on clearance of 10, 25, 100, and 400 μM boluses in wild-type cultures, 35 of the 10,000 models had an ER≤10 and were considered viable models. The windows of simulation results of these 35 models are presented in Fig 3A–3D along with the experimental clearance data they were trained on. None of these models contained a gradient; 30 were structure 2 models, and 5 were structure 3. We note that structures that contain a gradient could be favored over those with no gradient under different conditions (e.g., training on data from a single H2O2 concentration); however, our goal was to arrive at a model that could describe a wide range of bolus concentrations, and the gain in simulation accuracy for gradient models did not justify addition of the extra parameter as determined by the ER for the experimental conditions considered here. When the H2O2 distributions of the acceptable models were analyzed, the utility of AHP and the catalases separated into two distinct groups at all bolus concentrations (Fig 3E–3H). The reaction fluxes through AHP and HPI+HPII can be found in S2A–S2D Fig, and those also separated into two distinct groups. We found that these two groups represented predictions made by the two model structures. At all bolus concentrations, structure 2 models predicted a greater contribution by AHP than did structure 3 models. Indeterminacy in catalase null mutant simulations (Fig 3I–3L) suggested that experiments on a strain lacking both HPI (katG) and HPII (katE) would resolve this discrepancy.
Simultaneous training of models on wild-type and ΔkatE ΔkatG clearance data was able to resolve the uncertainty between structures 2 and 3. This training iteration resulted in 965 models that all had an ER≤10 (Fig 4A–4D), and all acceptable models were structure 3, which suggests that bimolecular H2O2-dependent enzyme degradation is an important feature of the detoxification network. We note that clearance of 400 μM H2O2 by ΔkatE ΔkatG was omitted because significant cell death was observed (S1H Fig), and the models were not designed to simulate cell death and possible lysis. All models predicted similar distributions across the major pathways (Fig 4E–4H), but diverged when we looked more closely at the individual catalase contributions (Fig 4I–4L). Reaction fluxes through the major pathways and individual catalases can be found in S2E–S2H Fig and S3A–S3D Fig, respectively. The different parameter sets predicted a range of clearance profiles after removal of either catalase (S4 Fig), suggesting that data obtained from these mutants would resolve the disagreement between models.
Training uncertain model parameters on wild-type, ΔkatE ΔkatG, ΔkatE, and ΔkatG data resulted in 40 parameters sets from the 1,000 random initializations that were within an ER of 10 (Fig 5A–5D). All of these models agreed regarding how H2O2 distributes across not only the major pathways (Fig 5E–5H), but also the individual catalases (Fig 5I–5L). Reaction fluxes through the major pathways and individual catalases can be found in S2I–S2L Fig and S3E–S3H Fig, respectively. The consistent distributions satisfied our design criteria, so we proceeded with the generation of an ensemble of viable parameter sets with which to make forward predictions.
The identification of universal “sloppiness” in computational biological models [46], meaning many parameters are poorly constrained after fitting on experimental data, led to the development of a number of methods designed to identify ensembles of parameter sets that could comparably describe the data and be used to assess the robustness of forward predictions [26, 46]. Methods such as “brute force” uniform sampling or Gaussian sampling become impossible with increasingly complex models, so computational biologists have turned to the use of Monte Carlo techniques to explore the viable parameter space efficiently (e.g., HYPERSPACE [26] and SloppyCell [46]). Here, we used a previously developed MCMC method [26] to explore the parameter space, initiating a random walk away from each of the 40 acceptable parameter sets and keeping 100 viable sets with an ER≤10 for each point. This resulted in an ensemble of 4,000 parameter sets that could all capture our experimental observations, and allowed us to assess robustness of our predictions to parametric uncertainty. In addition, before proceeding, we ensured that all models in the ensemble satisfied our design criteria (S5 Fig).
Based on the tight predictions that AHP, HPI, and HPII would dominate clearance (Fig 5), we sought to determine the minimal reaction network required to capture all of our data. To do this, we adopted a previously used two-tiered approach that first deletes reactions in a random order, and then re-optimizes uncertain parameters to determine if adjusted parameters would allow deletion of additional reactions [33]. Beginning with the best model in our ensemble and using this method, we determined that 70 out of our 75 reactions could be removed without increasing the ER beyond a threshold of 10. The essential reactions to the network were the major detoxification enzymes (AHP, HPI, and HPII) and degradation of AHP and HPI. In the case of AHP, a drop in active enzyme could indicate degradation, or alternatively a decrease in available NADH, which is held constant during simulation. On the other hand, H2O2 is the only substrate of catalase, which suggests that the importance of degradation reflects a decrease in concentration of functional enzyme.
In addition to identifying reactions in the network that are dispensable to capturing the H2O2 clearance data presented in Fig 5, we identified those uncertain parameters that influenced the simulations. Using the optimal parameter set, we individually varied each of the 13 optimized parameters within their bounds. Changes to the Fenton reaction rate constant and Fe2+ and Fe3+ initial concentrations never increased the ER to beyond 10. All other uncertain/trained parameters perturbed simulations to varying degrees, and their impact was quantified and displayed in S6 Fig
With the model developed and an ensemble of viable parameter sets identified, we sought to assess its predictive capabilities on a physiologically-relevant environmental perturbation. There is growing evidence that microbial killing within macrophages is a combined effect of the toxic environment and a scarcity of nutrients [47–49]. We therefore chose to investigate how H2O2 detoxification changes during carbon starvation. In the absence of an exogenous source of energy and carbon, the abundance of reducing equivalents can fall to limiting levels [50] and energetic processes such as translation can be hampered [51]. These effects could impact the stress response network by limiting AHP activity and inhibiting H2O2-dependent induction of AHP and HPI. To determine if carbon starvation in the media used here leads to NADH depletion, we directly measured NADH and NAD+ in M9 media cultures with and without glucose and found that a significant reduction in NADH occurred in carbon-starved cultures (S7 Fig). To see if carbon starvation depresses NADH to levels that inhibit enzyme activities, we measured respiration, which is an NADH-driven process, in M9 media in the presence and absence of glucose and found it to be significantly impaired when glucose was omitted (S8 Fig). In addition, to see if a lack of carbon reduces NADH to levels that impair AHP activity, we monitored clearance of 10 μM H2O2 in a strain with AHP as the lone major detoxification enzyme (ΔkatE ΔkatG), and found that omission of glucose completely inhibited H2O2 clearance in this strain (S9 Fig). In accordance with these results, model predictions indicated that if NADH was not held constant, AHP would drain it from the system in less than a second (S10 Fig). The impact of glucose starvation on induction of AHP and HPI expression was also assessed with the use of GFP reporter plasmids. Omission of glucose completely inhibited H2O2-dependent induction (S11 Fig). Therefore, to simulate the impact of carbon deprivation, NADH concentrations were no longer held constant and protein production was set to zero.
Ensemble predictions for carbon deprivation (- glucose) were made using the complete reaction network and are shown in Fig 6A–6D, along with the carbon-replete control (+ glucose). These predictions were experimentally confirmed and the data are presented in Fig 6I–6L, orange). To quantify how the different elements of glucose deprivation (NADH limitation, inhibition of translation) contributed to the observed phenotypes, we investigated the individual effects of NADH depletion or translation inhibition with simulation controls (Fig 6E–6H). At lower treatment concentrations (10 and 25 μM H2O2), starvation was predicted to slow detoxification as a result of NADH depletion, whereas inhibition of protein synthesis was predicted to have a negligible effect. At 100 μM H2O2, glucose-deprived cultures were predicted to clear H2O2 comparably to glucose-fed cultures, with neither reducing equivalent availability nor enzyme production substantially hindering detoxification. The impact of starvation at 400 μM H2O2 was predicted to be largely mediated by translation. We note that although selective inhibition of NADH production and usage was not feasible due to the wide variety of sources and sinks, targeted inhibition of translation was experimentally tractable with the use chloramphenicol (CAM) (S11 Fig). Experimental confirmation of clearance by CAM-treated cultures is shown in Fig 6I–6K. Unfortunately, CAM-treatment led to cell death at 400 μM H2O2 (S1X Fig), which prevented direct confirmation of the prediction at that concentration. Interestingly, this cell death suggested that translation of some protein other than AHP or HPI is important to survival at 400 μM H2O2, because we demonstrated that carbon deprivation inhibited induction of AHP and HPI at 400 μM H2O2 (S11 Fig), and it is known that carbon deprivation can have promoter-specific effects [51] and CAM stops synthesis of all proteins.
The toxic nature of H2O2 makes it an ideal weapon in inter-species warfare, and it is used as such by the immune system during infection [2] and even by other bacteria in niche competitions [15]. Bacteria have evolved numerous defense systems, which can differ significantly in their substrate requirements, reaction mechanisms, and regulation, and the complexity of these defense networks and the broad reactivity of H2O2 necessitate the use of computational modeling for quantitative interpretation and prediction of H2O2 distributions in cells [52]. Due to its importance as a signaling molecule, models of H2O2 metabolism in mammalian systems have been constructed, and they have included enzymatic detoxification of H2O2 [22–24], oxidation of cysteine residues [22], transport of H2O2 across membranes [22–24], and oxidation of targets involved in signaling [24]. However, beyond their specificity for mammalian systems, none have accounted for uncertainty in optimized parameters or included synthesis or inactivation of enzymes, side reactions of H2O2, or other reactive oxygen species present in the network. In bacteria, modeling efforts have focused on subsystems affected by H2O2 stress, such as that of the thioredoxin system in E. coli, which included the oxidation of thioredoxin and the reduction of oxidized thioredoxin, methionine sulfoxide, protein disulfides, and 3’-phosphoadenosine-5’-phosphosulfate [25]. These previous efforts inspired us to construct a quantitative, systems-level model of H2O2 stress in E. coli that includes media and cellular compartment-specific species and reactions; H2O2-dependent transcriptional regulation, inactivation, and activity of H2O2 detoxification enzymes; reductases to reduce oxidized species; O2−• and •OH and their related reactions (e.g., oxidation of all twenty amino acids by •OH); and reactions of H2O2 with other metabolites such as glutathione and the α-keto acid pyruvate. In addition, we addressed structural uncertainty in the model using an iterative computational and experimental methodology, and assessed parametric uncertainty using an MCMC procedure, which enabled the robustness of model predictions to be assessed. Similar ensemble approaches have become popular methods to account for parametric uncertainty [53–62], and several techniques have been developed to efficiently explore parameter spaces [26, 46].
One power of quantitative computational modeling is its ability to predict emergent systems behavior [33, 44, 63–65]. For instance, Schaber and colleagues used an ensemble of possible models describing different hypotheses regarding the mechanism of the high osmolarity glycerol (HOG) pathway in yeast to uncover novel features of the pathway [44]. Here, we leveraged our model to gain a quantitative understanding of how carbon deprivation, which bacteria encounter in phagocytes [47–49], affects H2O2 detoxification by E. coli. Accounting for the NADH limitation and translational inhibition that occurs with carbon source starvation, our simulations were able to correctly predict H2O2 clearance dynamics. Upon dissection of simulation results, delayed clearance at lower concentrations was attributed to reduced AHP activity from NADH depletion, whereas at higher concentrations carbon-starved cultures resembled carbon-replete cultures because pre-expressed HPI dominated H2O2 detoxification, suggesting that both NADH availability and induction of AHP and HPI synthesis at H2O2 concentrations > 25μM were of minor importance. We note that changes in concentrations of other metabolites, such as ATP, occur in carbon-starved cultures [66], and that they were not explicitly accounted for here because they did not directly act as a substrate in any of the reactions of the model. Rather we anticipate that some of the metabolite perturbations were implicitly accounted for because they contributed to the inhibition of translation and/or depletion of NADH that were included. These data demonstrate that the nutritional status of the environment can have a major impact on bacterial H2O2 defenses, but the extent of that impact depends on the quantitative level of H2O2. Beyond carbon deprivation, it would be interesting to see how other types of starvation (e.g., sulfate, iron) influence H2O2 detoxification, since bacteria are subjected to oxidative stress in various scenarios [1, 15, 67], and we expect that dependencies distinct from those of carbon source starvation could be observed if other types of limitation influence NADH availability and translation differently.
In immune cells, phagocytized bacteria are exposed to ROS with an oxidative “burst” from NADPH oxidase, which then tapers over time [68–70]. In this work, we examined detoxification of a burst of H2O2 with bolus treatments, and note that more complex treatment dynamics could be handled by the model developed here. For instance, the model could be adjusted for continuous treatment by adding an H2O2 delivery reaction, which could be achieved experimentally with a fed-batch reactor. Alternatively, H2O2 could be provided through indirect means, such as with exposure to redox-cycling agents like paraquat [71]; though obtaining accurate estimates of H2O2 production from such compounds could be a challenge, because generation would be cell-dependent. Different H2O2 delivery dynamics could have a profound impact on the kinetic competition for H2O2, and as long as H2O2 influx can be accurately accounted for the model developed could prove invaluable for interrogating its distribution.
One area of growth for the platform we developed is adaptation of the model to allow for analysis of lethal H2O2 concentrations. In its current form, the size of the cellular compartment is fixed and enzymatic reactions do not occur in the media compartment. To model lethal concentrations of H2O2, cell lysis has to be accounted for in terms of reduction in the volume and surface area of the cellular compartment and the addition of certain enzymatic activities to the media compartment, such as catalase, which can function when released from cells. In addition, it might be necessary to diversify the cellular compartment if all cells that die do not lyse but contain compromised translational and/or catabolic activities. Despite this added complexity, the ability to accurately simulate lethal damage, such as that involving DNA and the membrane, could provide insight into H2O2-induced death.
Models akin to the one described here will improve understanding of bacterial defenses against host immune responses, and possibly suggest targets for novel anti-virulence therapies [52]. For example, an existing model of nitric oxide (NO•) defenses in E. coli [33] has provided valuable predictions regarding NO• delivery rates that maximize antimicrobial activity [63]. Additionally, it provided a framework that allowed for a model-guided investigation of the underlying mechanism of an NO•-sensitizing mutation, ΔclpP, in E. coli [72]. The success of the NO• model provided inspiration to develop a similar model for H2O2, which is another toxic, diffusible metabolite used by immune cells when fighting infection [1, 2]. We anticipate that the H2O2 model developed here will yield novel quantitative insight into the kinetic competition for H2O2 in E. coli and provide a framework for the mechanistic investigation of perturbations that affect clearance, while illuminating targets to sensitize bacteria to immune attack.
E. coli K-12 MG1655 was used in all experiments. ΔkatE and ΔkatG mutations were transduced into MG1655 from their respective strains in the Keio collection [73] by the P1 phage method. The ΔahpCF mutation was generously provided by Michael Kohanski and transduced into MG1655 using the P1 phage method. All antibiotic markers were cured out using pCP20 [74]. The known antioxidant pyruvate (25 mM) was included in all LB agar plates to prevent a toxic build-up of H2O2 [75]. Deletions were PCR checked for proper chromosomal integration with a forward primer external to the gene and reverse primer within the kanamycin resistance cassette (kanR) before curing. Internal primers were used to check for gene duplication. In cured strains, external primers before and after the gene were used to check for proper scar size. All PCR primers are listed in S4 Table.
Overnight cultures were inoculated from −80°C stocks and grown for 20 hours in 1 mL LB + 30–75 U/mL catalase (bovine liver catalase at 2,000–5,000 units/mg protein: Sigma Aldrich), then used to inoculate 20 mL M9 10 mM glucose medium + 30–75 U/mL catalase to an OD600 of 0.01 in 250 mL baffled flasks. Catalase was added to prevent the possibility of H2O2 accumulation in strains lacking major detoxifying enzymes, and added to wild-type cultures to maintain consistency across strains. The catalase concentration was chosen based on the amount required to maintain growth in a mutant lacking all major detoxification enzymes (ΔkatE ΔkatG ΔahpCF), beyond which increasing the catalase concentration no longer increased growth rate or terminal cell density in the case of overnights. Cultures were grown at 37°C with shaking at 250 rpm for 8 h (OD600 0.3–0.6, 0.15 for the slower growing ΔkatE ΔkatG ΔahpCF). After the 8 hour growth period, 12 mL of culture was removed to a pre-warmed 15 mL Falcon tube and centrifuged at 37°C and 4,000 rpm for 10 min. 10.8 mL of spent media was removed, the cell pellet resuspended, and 1 mL transferred to a warm 1.5 mL microcentrifuge tube. Cells were washed a total of four times to remove all catalase. Washes consisted of spinning down at 14,000 rpm for 2 min, removing 980 μL of media, and resuspending the cell pellet with 980 μL fresh warm media. For samples lacking glucose during challenge with H2O2, glucose was omitted during the final wash step and in the inoculated flask. For CAM treatment assays, all wash steps were performed with 100 μg/mL CAM.
Prior to inoculation with washed cells, 20 mL fresh M9 10 mM glucose media in 250 mL baffled flasks were warmed to 37°C. A bolus of 10, 25, 100, or 400 μM H2O2 was added to the flasks, and the time 0 point was measured, after which flasks were inoculated to an OD600 of 0.01. At desired time points, 200 μL was removed to a 1.5 mL microcentrifuge tube and centrifuged at 15,000 for 3 min. 150 μL of the supernatant was moved to a sterile microcentrifuge tube and stored at 4°C until H2O2 concentration could be measured. Samples were assayed for H2O2 within 2 h of harvesting. H2O2 in the supernatant was measured using the Amplex Red Hydrogen Peroxide/Peroxidase kit (Life Technologies) according to the manufacturer’s instructions after dilution to below 10 μM H2O2. A standard curve spanning 0 to 10 μM H2O2 was used to calculate H2O2 concentrations. A fresh standard curve was produced for each Amplex Red assay to account for increasing background fluorescence over the course of the day due to the sensitivity of Amplex Red to both light and air [76].
To assess whether centrifuging to remove cells was sufficient, we compared values from this method to sterile filtering (0.22 μm pore size) of samples, as well as centrifuging + filtering, for the 30 min point in the 400 μM H2O2 clearance assay (S12 Fig). Centrifuging alone was not significantly different from filtering (p = 0.45) or centrifuging + filtering (p = 0.40) based on a two-sample t-test with unequal variance. To determine if the exogenous catalase that was added to the pre-culture steps affected clearance profiles, we performed identical experiments for wild-type propagated without catalase in all steps (S13 Fig). The presence of catalase in the pre-culture did not affect wild-type clearance of H2O2 at any concentration.
To determine whether H2O2 treatment resulted in cell death, we quantified CFUs throughout the clearance assays. After isolating the H2O2-containing supernatant for Amplex Red assays as described above, an additional 30 μL supernatant was removed from centrifuged samples and discarded, to achieve a greater fold-dilution of H2O2 during the first wash step. In the first wash step, 980 μL of PBS was added and the cell pellet was resuspended. Samples were centrifuged again at 15,000 rpm for 3 min, 980 μL of the supernatant was removed, and the cell pellet was resuspended a final time in 80 μL PBS. Plating was performed using the serial dilution method, and samples were plated on LB agar containing 25 mM pyruvate to scavenge any residual H2O2 remaining in the pellet and any endogenously produced H2O2 in scavenging-deficient strains. Plates were incubated for 16 h at 37°C prior to counting colonies.
MG1655 cultures were grown and washed identically to the H2O2 clearance assay. After the final wash, the resuspended cells were used to inoculate 10 mL of pre-warmed M9 with or without glucose in a 50 mL Falcon tube containing a sterile magnetic stirring bar, immersed in a stirred water bath at 37°C, to an OD600 of 0.1. Cells were allowed to consume oxygen for ten minutes before being treated with 5 mM KCN to halt respiration, which consumes the majority of O2 in E. coli cell cultures under these conditions [63]. The percent oxygen saturation was measured at a frequency of one reading per second using the FireStingO2 fiber-optic O2 meter with the OXROB10-CL2 robust oxygen miniprobe (PyroScience, GmbH). Temperature fluctuations were compensated for using the TDIP15 temperature sensor (PyroScience GmbH) and the FireSting Logger Software. The equilibrium oxygen concentration was used to convert the percent saturation to concentration, and was determined by calibrating the probe in ultrapure Milli-Q water at 37°C, which has an oxygen concentration of 210 μM [77], and transferring the probe to air-saturated M9 media. The equilibrium concentration of the media matched that of the ultrapure water.
Overnight cultures were inoculated and grown identically to the H2O2 clearance assay, and used to inoculate 20 mL M9 10 mM glucose medium + 30–75 U/mL catalase to an OD600 of 0.01 in 250 mL baffled flasks. Cultures were grown at 37°C with shaking at 250 rpm to an OD600 of 0.2 (~6.5 h). Four 1 mL aliquots were transferred from the flask to warm, 1.5 mL microcentrifuge tubes and centrifuged at 15,000 rpm for 3 min. The media was removed, and the pellets were resuspended with 1 mL fresh M9 with or without 10 mM glucose and transferred to warm test tubes. The tubes were then incubated at 37°C with shaking at 250 rpm for 60 min. The time 0− point was taken directly from the flask prior to centrifuging and resuspension.
NAD+ and NADH were measured using the EnzyChrom NAD/NADH Assay Kit (BioAssay Systems) following the manufacturer’s protocol, except for a brief sonication step. For each measurement, 400 μL (NAD+) or 800 μL (NADH) of cell culture was transferred from the flask (time 0) or test tube (60 min) to a 1.5 mL microcentrifuge tube. The tubes were centrifuged for 3 min at 15,000 rpm, and 380 μL (NAD+) or 780 μL (NADH) of supernatant was removed and discarded. The cell pellets were resuspended in 100 μL of either NAD or NADH extraction buffer and sonicated for 20 s at room temperature at an amplitude of 10 using a Fisher Scientific Model 50 Sonic Dismembrator. The extracts were heated at 60°C for 5 min, before adding 20 μL of assay buffer and 100 μL of the opposite extraction buffer. Samples were then vortexed briefly and centrifuged for 5 min at 14,000 rpm. The supernatants were used for the NAD/NADH assay following manufacturer’s instructions. An NAD+ standard curve from 0–2 μM was generated each day and used to convert absorbance to concentration. The standard curve underwent extraction protocols identical to cell samples, including sonication and heating. Since NAD+ and NADH produce identical standard curves, and NADH is more unstable, only NAD+ was provided in the kit and was used to convert both NAD+ and NADH absorbance to concentration.
MG1655 was transformed with pUA66 PahpC-gfp, pUA66 PkatG-gfp, and pUA66 PkatE-gfp, which were all obtained from a pre-existing library [78]. Overnight cultures were inoculated from −80°C stocks and grown for 20 h in 1 mL LB + 30–75 U/mL catalase + 30 μg/ml kanamycin for plasmid retention, then used to inoculate 20 mL M9 10 mM glucose medium + 30–75 U/mL catalase + 30 μg/ml kanamycin to an OD600 of 0.01 in 250 mL baffled flasks. Cultures were grown, washed, and treated with H2O2 identically to the protocol described in the Amplex Red assay protocol. Washed cells were fixed before inoculation to the H2O2-containing flasks to provide a time 0− sample for each condition. Final points were sampled and fixed when ~90% of the H2O2 had been cleared by wild-type in 10 mM glucose M9 media: 30 min for the 10 μM H2O2 flask, 40 min for the 25 μM H2O2 flask, 1 h 15 min for the 100 μM H2O2 flask, and 2 h for the 400 μM H2O2 samples. Fixing involved removing 1 mL culture to a microcentrifuge tube and centrifuging at 15,000 rpm for 3 min, removing 980 μL supernatant, and resuspending with 480 μL 4% paraformaldehyde (PFA). After 25 min at room temperature, the samples were again centrifuged at 15,000 rpm for 3 min, 480 μL of the PFA was removed, and the pellet was resuspended with 980 μL 1X PBS. Samples were stored at 4°C until analysis by flow cytometry on an LSR II flow cytometer (BD Biosciences, San Jose, CA), where green fluorescence was measured on a per cell basis. Fluorescence was measured using 488 nm excitation and a 525/20 bandpass filter, and data were acquired using FACSDiVa software (BD Biosciences, San Jose, CA).
The modeling framework used in this work largely followed that used by Robinson and Brynildsen [33]. It is composed of a system of ordinary differential equations that are numerically integrated to provide predicted species concentration over time. We begin with a mole balance for all species in the model:
dNdt=S⋅v
(1)
where N is an s x 1 vector representing the total amount of given species in moles, S is the s x r stoichiometric matrix, and v is an r x 1 vector representing reaction rates in moles per time. Here, s indicates the number of species in the model, and r indicates the number of reactions in the model. Since most reaction rates are calculated on the basis of concentration per time, the rate vector was converted into these units in the following manner.
dNdt=S⋅Vrxn⋅r
(2)
where Vrxn is an r x r diagonal matrix representing the volumes of the compartments in which the reactions are taking place: Vcell for intracellular reactions, Vmedia for reactions taking place only in the media, and Vtotal for exchange reactions. Due to experiments also being performed on a concentration basis, N was converted into units of concentration by performing the following operation.
Vspec-1dNdt=Vspec-1⋅S⋅Vrxn⋅r
(3)
where Vspec is a diagonal s x s matrix of species compartment volumes: Vcell for intracellular species, Vmedia for species in the media compartment, and Vtotal for species that freely diffuse across the cell membrane (e.g., O2, H2O2 in non-gradient models). The left-hand side of the equation is equivalent to dC/dt when the volume does not vary appreciably over the course of the experiment. To avoid having a culture-volume-specific model, we transformed the volume dependencies into volume fractions by multiplying and dividing by Vtotal (Vcell/Vtotal, Vmedia/Vtotal, Vtotal/Vtotal) and rearranging the equation with the use of the commutative property of scalar multiplication.
dCdt=VtotalVtotal⋅Vspec-1⋅S⋅Vrxn⋅r
(4)
dCdt=Vtotal⋅Vspec-1⋅S⋅1Vtotal⋅Vrxn⋅r
(5)
dCdt=Fspec-1⋅S⋅Frxn⋅r
(6)
where Fspec is an s x s diagonal matrix of the volume fractions for species and Frxn is an r x r diagonal matrix of the volume fractions for reactions. By making this adjustment, we avoid the need of requiring total culture volume as an input into the model, and simplify the input to optical density (OD600) that can readily be converted to volume fractions.
Most initial species concentrations were obtained from literature (S1 Table)[21, 25, 79–90]. The equilibrium concentration of oxygen was determined by calibrating a FireStingO2 fiber-optic O2 meter with the OXROB10-CL2 robust oxygen miniprobe (PyroScience, GmbH) in ultrapure Milli-Q water at 37°C and transferring it to air-saturated M9 10 mM glucose medium at 37°C. We calculated the concentration in our media by comparing it to the known value in deionized water [77], which is 210 μM. The value in our media was equivalent. The initial H2O2 concentration was set to the initial average value of the data when optimizing parameters (e.g., 25.89 μM instead of 25 μM for wild-type) to avoid penalizing model fit for experimental error. When making forward predictions, the concentration was set to the anticipated initial concentration (e.g., exactly 25 μM).
Initial species that were trained on experimental data included AHP, HPI, HPII, Fe2+, and Fe3+. While experimental measurements on AHP [21], HPI [82], and HPII [82] are available, their concentrations vary with environment and growth phase, as shown in our reporter assays (S11 Fig). AHP, HPI, and HPII are the major H2O2 detoxification systems in E. coli. We therefore allowed flexibility in their initial concentrations, constraining them to be within the range of 0–20 μM. Because individual concentrations of Fe2+ and Fe3+ are unresolved, but experimentally measured to have a combined concentration of 10 μM, we allowed their initial concentrations to vary from 0 to 10 μM.
Catalase activity follows a ping-pong mechanism, reacting with one H2O2 to form a reactive intermediate, followed by reaction with a second H2O2 molecule to return the enzyme to its original form [91, 92]. Given that the substrate in the first and second reaction is the same, the rate equation simplifies to a Michaelis-Menten type structure. While inhibition of catalase activity by H2O2 becomes apparent at high concentrations (e.g., greater than 100 mM H2O2 for E. coli HPII [92]), it is assumed negligible at the concentrations used in our experiments, and Michaelis-Menten kinetics are appropriate. Rate equations and constants can be found in S3 Table (Reactions 69 and 70). While HPI has the ability to utilize other reducing agents at low H2O2 concentrations, this activity is significantly slower than its catalase activity (about 1% the kcat of its catalase activity [21]), so the peroxidase activity was assumed to negligibly contribute to H2O2 detoxification in this study.
The AHP reaction cycle begins when the peroxidatic cysteine of AhpC reacts with H2O2 to form a sulfenic acid, which resolves to form a disulfide bond with another cysteine residue. The active AhpC is regenerated by its reductase partner AhpF, which uses NADH as an electron donor [38]. We modeled this cycle using ping-pong reaction kinetics, with H2O2 and NADH as the substrates, a structure which has been used previously [38]. Kinetic parameters were not available for E. coli, but a protein BLAST search [93] revealed 98% protein sequence identity for AhpC and 95% identity for AhpF between E. coli MG1655 and S. Typhi, so available parameters for S. Typhi were used [38, 94]. Additional information and rate constants can be found in S3 Table (Reaction 71).
In this work, the main experimental variable was the concentration of H2O2, and therefore, we opted for simplicity and only H2O2-dependent regulation of gene expression was considered. The expression of catalase HPI and AHP increase in response to H2O2 [17, 95], whereas the expression of catalase HPII is not dependent on H2O2 [17, 18]. These dependencies were confirmed using transcriptional reporters for ahpC, katE, and katG, and measuring fluorescence on a per cell basis using flow cytometry (S11 Fig). Following previous dynamic models, gene expression was modeled using a Hill equation with a coefficient of n = 1 [33, 34], except for HPII which had initial concentration but was not expressed further. In addition, transcription was assumed to be limiting in the production of active enzyme, as assumed previously [33, 34], and the bioavailability of ferroheme b, which is an essential cofactor of HPI, was assumed to not be rate limiting. HPI and AHP are expressed according to Reactions 67 and 68, respectively, in S3 Table. The maximum expression rate and Hill equation constants KAHP-exp,H2O2 and KHPI-exp,H2O2 are informed during parameter optimization. The bounds on the maximum expression rates are based on the highest and lowest maximum expression rates found in the work of Kotte and colleagues [34], and have been used previously when optimizing unknown expression rates [33]. Bounds on KAHP-exp,H2O2 and KHPI-exp,H2O2 were approximated by the work of Kotte and colleagues [34], which varied from approximately 2 nM to 1 mM. Here, we allowed variation from 0 to 1 mM. We note that while the parsimonious treatment of H2O2-dependent expression was sufficient in this work, exploration of new environments could necessitate a more comprehensive modeling of transcriptional regulation, since expression of the detoxification enzymes in this work are known to depend on numerous regulators (e.g., OxyR, RpoS, FIS, Fur) [96].
In previous studies, enzymes have typically been modeled as undergoing first order degradation with a universal constant [33, 34]. However, rates can vary greatly among different proteins [35, 36], and some evidence suggests that H2O2 detoxification enzymes, such as catalase and alkyl hydroperoxidase, can be poisoned by their own substrate [37–39]. Inactivation of catalase has been described using bimolecular kinetics [40], but it has also been suggested that poisoning should be of the same general form as the enzyme’s reaction kinetics [37]. The AhpC component of alkyl hydroperoxidase is reduced by AhpF after oxidation by H2O2. If there is not enough NADH present or AhpC reacts with more H2O2 before encountering AhpF, the cysteine sulfenic acid formed by the first oxidation can be further oxidized to sulfinic or sulfonic acid, rendering it inactive [38, 97]. Whether AhpC inactivation is significant at the concentrations used in this work was uncertain [39]. We therefore allowed it to be degraded in a first order or bimolecular manner. Due to the indeterminacy of how these enzymes are degraded, we included all of these possible degradation scenarios. Rate equations can be found in S2 Table. We note that since AHP requires a co-substrate, NADH, and that we assume that NADH is constant (unless otherwise noted), if AHP inactivation were found to be important it could reflect degradation of AHP or reduced availability of NADH (Reaction 71).
Parameters or bounds on parameters for enzyme degradation/deactivation rate equations varied with the method of degradation. For constant degradation with a constrained constant, we used the "general" protein degradation rate reported by Kotte and colleagues [34]. When optimized, the constant degradation rate had bounds set by the longest [35] and shortest [36] protein half-life we found in literature. For both bimolecular and the more complex inactivation, bounds were based loosely on rate constants found for Aspergillus niger and bovine catalase [37, 40]. Based on the gross difference between the organisms tested in those studies (fungus and mammal) and our own (bacteria), as well as the orders of magnitude difference between the Aspergillus niger and bovine catalase rates in both the bimolecular and more complex kinetics studies, these parameters were constrained to two orders of magnitude lower than the lowest reported value, and two orders of magnitude higher than the highest reported value.
While H2O2 rapidly diffuses across bacterial membranes at a rate similar to that of water, there is evidence for and against the existence of a gradient across the membrane [41, 42]. For example, an Ahp−Kat+ strain cocultured with Ahp−Kat− in the presence of a low H2O2 concentration can outcompete its scavenging deficient neighbors and multiply under the stress, whereas the deficient strain only grows after the catalase proficient strain has cleared the environmental H2O2 [41]. On the other hand, dilute suspensions of catalase proficient strains are readily killed by high concentrations of H2O2 similarly to scavenging deficient strains, while high-density catalase proficient strains can not only survive challenge with H2O2 but also protect deficient neighbors [42]. There have been strides in the ability to measure H2O2 intracellularly, with the introduction of genetically-encoded indicators (HyPer) [98] and the ability to use Amplex Red intracellularly with expression of a mutated ascorbate peroxidase [99]. However, the dependence of HyPer fluorescence on reductase activity, the impact of HyPer on cellular scavenging capacity [100], and the difficulties associated with converting measurements from either method to absolute H2O2 concentrations led us to use measurements of external H2O2 and statistical metrics (AIC) to assess the suitability of modeling the system with or without a gradient. In one set of models, the intracellular and extracellular H2O2 concentrations were equal. In the other set, we allowed for a gradient by modeling transport across the membrane as a convective mass transport process, with the effective mass transport coefficient being an additional parameter for optimization. The lower bound on the effective mass transfer coefficient was set as the permeability coefficient of H2O2 across E. coli cell membranes in unstirred culture [41], adjusted for cell area and the cell density for our system. The upper bound was set two orders of magnitude higher than the permeability, to account for increased mass transfer in our chaotic shake flask system.
The rate of spontaneous degradation of H2O2 into H2O and O2 was determined using the MATLAB function lsqcurvefit after monitoring H2O2 concentration over time in cell-free controls for each media condition (M9 10 mM glucose with and without CAM and M9 lacking glucose). Samples were collected at time 0, 20 min, 40 min, and 60 min for 10 and 25 μM H2O2-containing flasks; time 0, 1 h, 2 h, and 3 h for 100 μM H2O2-containing flasks; and 0 h, 1 h, 2 h, 3 h, and 4 h for 400 μM H2O2-containing flasks. All control data for each media condition was fit simultaneously (e.g., 10, 25, 100, and 400 μM H2O2 in M9 10 mM glucose) while accounting for experimental error by weighting points in the sum squared of residuals (SSR) calculation by the inverse of the variance for that data point [43, 101–103]. The optimized rate constants were 0.0324 h-1 for M9 10 mM glucose, 0.0331 h-1 for M9 10 mM glucose with 100 μg/mL CAM, and 0 h-1 for M9 lacking glucose. The spontaneous degradation rate was assumed to be equivalent in the intracellular and media compartments.
Other reactive oxygen species in the reaction network include O2-• and •OH. The model includes endogenous production of O2-• (Reaction 4) and its dismutation to H2O2 (Reactions 3, 74–75), as well as its reactions with other molecules (Reactions 1, 25, 26, 55) and production by other reactions in the network (Reaction 49). Additionally, the Fenton reaction produces •OH (Reaction 24), which can react with amino acids (Reactions 5–23) and other compounds in the network (Reactions 1, 27, 36, 37, 48). The rate constant for the Fenton reaction were variable in literature, so bounds were set as the slowest and fastest reported rates [21].
Other reactions include glutathione oxidation, reduction, and reaction with other molecules (Reactions 29–32, 42–55), and oxidation of methionine (Reaction 2) and pyruvate (Reaction 57) by H2O2. All spontaneous reactions, rate equations, and rate constant can be found in S2 Table [21, 34–37, 39–41, 81, 85, 88, 91, 92, 104–111]. Enzymatic rate equations and rate constants can be found in S3 Table [38, 91, 92, 94, 111–117], and include methionine sulfoxide reductase (Reaction 64), thiol peroxidases (Reaction 65–66), thioredoxin reductase (Reaction 72), and glutathione reductase (Reaction 73).
Parameters were optimized by minimizing the SSR using the built-in MATLAB function lsqcurvefit. Because experimental error varied for each time point, we weighted each data point’s contribution to the SSR by the inverse of the variance of that point [43, 101–103]. The initial concentration of H2O2 in the model was changed to match the experimental data before optimizing. Due to the nonlinear nature of the optimization, each model structure was initialized with random parameter sets within the defined bounds a total of 1,000 times. The progression of minimum SSR found by each of these iterations is shown in S14 Fig. A plateau suggests that additional iterations would not lead to substantial improvement in the fit. The slowest converging optimization reached a plateau after 639 optimizations.
The number of parameters optimized varied according to model structure, as presented in Table 1. Parameters related to HPI and AHP expression (4 parameters total) were optimized in all structures. The Fenton reaction rate constant varied in literature (1 parameter), and intracellular Fe2+ and Fe3+ concentrations (2 parameters) were unresolved. Additionally, initial concentrations of the major detoxifying enzymes (3 parameters) were unknown and can vary with growth environment and stage of cell growth. Beyond these 10 parameters, all structures that did not have constant enzyme degradation with a universal degradation constant added 3 parameters, and including an H2O2 gradient added 1 parameter (a convective mass transfer coefficient).
The introduction of additional parameters, such as enzyme degradation rate constants and mass transport coefficients, has the potential to improve fit solely by increasing the flexibility of the model. To account for the utility of additional parameters, we ranked models based on their evidence ratios (ER), or likeliness relative to the most likely model in the set. For each model, we calculated its Akaike Information Criterion corrected for small sample size (AICc) [43]:
AICc=n⋅ln(SSRn)+2K+2K(K+1)n−K−1
(7)
where n represents the sample size and K is the number of estimable parameters. Here, n is the number of data points used in the fitting procedure, and K is the number of model parameters plus 1 because regression estimates SSR and parameter values [43]. We account for unequal variances within the data by using weighted least squares, where each point is weighted by the inverse of its variance.
The weight of evidence for a given model in a set of M models is given by the following [43]:
wi=e−Δi/2∑i=1Me−Δi/2
(8)
where Δi = AICi-min(AIC). With this, an ER can be calculated, which represents the relative likelihood of a model compared to the best model in the set [27]:
ERi=wbestwi
(9)
A larger ER indicates a more unlikely model. In this work, models with an ER greater than 10 were discarded, a cutoff that has been used previously to discard models during model selection [29].
To account for parametric uncertainty when making predictions, we generated an ensemble of parameter sets that all predicted the data within an ER≤10. We initially attempted to use the software HYPERSPACE [26], which is a three-step process that provides a uniform sampling of the viable parameter space. However, the calculation of our cost function was computationally expensive, which lengthened the time required per iteration, and the method had not converged within 100,000 iterations. Therefore, we utilized the pre-existing Markov chain Monte Carlo function within the HYPERSPACE software, but started it from all 40 viable parameter sets that met our design criteria. We allowed approximately 200 random steps away from each viable point, and randomly selected 100 of those parameters sets with an ER≤10 from each random walk. This process generated 4,000 parameter sets that had an ER≤10.
We identified the minimal model that was able to capture our data (ER≤10) using a previously developed two tiered approach [33]. In the first tier, reactions were removed from the best model in a random order and the ER was calculated. If the deletion of a reaction increased the ER above its threshold of 10, it was returned to the model and the process continued through the remaining reactions. This random deletion process was repeated 100 times. In the second tier, parameters were re-optimized after the deletion of each of the remaining reactions. If the optimization produced a model that returned below an ER of 10, the parameters were changed and the process continued.
|
10.1371/journal.pcbi.1004884 | A Network Biology Approach Identifies Molecular Cross-Talk between Normal Prostate Epithelial and Prostate Carcinoma Cells | The advent of functional genomics has enabled the genome-wide characterization of the molecular state of cells and tissues, virtually at every level of biological organization. The difficulty in organizing and mining this unprecedented amount of information has stimulated the development of computational methods designed to infer the underlying structure of regulatory networks from observational data. These important developments had a profound impact in biological sciences since they triggered the development of a novel data-driven investigative approach. In cancer research, this strategy has been particularly successful. It has contributed to the identification of novel biomarkers, to a better characterization of disease heterogeneity and to a more in depth understanding of cancer pathophysiology. However, so far these approaches have not explicitly addressed the challenge of identifying networks representing the interaction of different cell types in a complex tissue. Since these interactions represent an essential part of the biology of both diseased and healthy tissues, it is of paramount importance that this challenge is addressed. Here we report the definition of a network reverse engineering strategy designed to infer directional signals linking adjacent cell types within a complex tissue. The application of this inference strategy to prostate cancer genome-wide expression profiling data validated the approach and revealed that normal epithelial cells exert an anti-tumour activity on prostate carcinoma cells. Moreover, by using a Bayesian hierarchical model integrating genetics and gene expression data and combining this with survival analysis, we show that the expression of putative cell communication genes related to focal adhesion and secretion is affected by epistatic gene copy number variation and it is predictive of patient survival. Ultimately, this study represents a generalizable approach to the challenge of deciphering cell communication networks in a wide spectrum of biological systems.
| In the current era of cancer research, stimulated by the release of the entire human genome, it has become increasingly clear that to understand cancer we need to understand how the many thousands of genes and proteins involved interact. Modern techniques have enabled the collection of unprecedented amounts of high quality data describing the state of these molecules during cancer development. In cancer research particularly, this strategy has been particularly successful, leading to the discovery of new drugs able to target key factors promoting cancer growth. However, a large body of research suggests that in complex organs, the interaction between cancer and its surrounding environment is an essential part of the biology of both diseased and healthy tissues, therefore it is of paramount importance that this process is further investigated. Here we report a strategy designed to reveal communication signals between cancer cells and adjacent cell types. We apply the strategy to prostate cancer and find that normal cells surrounding the tumour do exert an anti-tumour activity on prostate cancer cells. By using a statistical model which integrates multiple levels of genetic data, we show that cell-to-cell communication genes are controlled by DNA alterations and have potential prognostic value.
| Prostate Cancer is the most common cancer in males. It is characterized by a considerable molecular and phenotypic heterogeneity that results in radically different clinical outcomes [1].
The role of tumour microenvironment in the development of cancer is crucial. More specifically, the expression of growth and motility factors, extracellular matrix components produced by stromal cells, is linked to the pathophysiology of the tumour and it often predictive of clinical outcome. Stromal cells, such as fibroblasts and endothelial cells secrete many factors that influence the expansion of the tumour. For example, they secrete most of the enzymes involved in extracellular matrix breakdown and produce growth factors that control tumour cell proliferation, apoptosis, and migration [2]. They also secrete pro-inflammatory cytokines, which play a major role in a wide spectrum of pathophysiology mechanisms (e.g. chemo attraction, neoplastic transformation, angiogenesis, tumour clonal expansion and growth, passage through the ECM, intravasation into blood or lymphatic vessels and the non-random homing of tumour metastasis to specific sites) [3]. In addition to tumour promoting factors, they also secrete tumour suppressor factors that can potentially have an anti-tumour effect on adjacent tumour cells [4]. Current research on the role of stroma is principally focused on immune cells fibroblasts and cells of the vasculature such as endothelial cells. However, since other cell types, such as normal epithelial cells, also produce a number of these factors, such as IL-6 [5], TNFα [6] [7] and TGFβ1 [7] it is reasonable to hypothesize that they may also play an important role in influencing the molecular and physiological state of tumour cells.
The intricacy in the biology of cell-to-cell communication and the relatively small amount of available knowledge makes understanding the biological networks underlying the development of tumour microenvironment a suitable challenge for a systems-level approach. The powerful combination of functional genomics and computational biology have contributed to the discovery of novel signaling networks in the biology of cancer [8] [9], including cell communication networks [10]. However, so far there has been no attempt to develop a completely data-driven systems biology approach to discover novel cell-communication networks.
Here we describe a data-driven strategy we developed to address this challenge. Our approach is designed to “learn” the underlying structure of cell-to-cell communication networks from functional genomics datasets, representing the transcriptional state of normal and adjacent tumour cells.
The application of this novel analysis strategy to prostate cancer revealed genes whose expression is associated to directional signals linking normal and tumour epithelial cells. Remarkably, experimental validation of our predictions using an in vitro co-culture system recapitulated the predicted transcriptional response and revealed that normal epithelial cells have the potential to revert some of the phenotypic traits of tumour cells. Moreover, by integrating genetics, gene expression and tumour features in a single conceptual model, we were able to show that putative cell communication networks, involved in focal adhesion and protein secretion are perturbed by genetic mutations and that are linked to survival.
Ultimately, the experimental validation of the hypothesis generated from the model support the approach we have developed, which explicitly search for candidate directional signals between different cell types. Its application to a wider range of biological systems is likely to have a profound impact in the field of functional genomics.
Our study is based on a data analysis workflow which includes reverse engineering techniques to identify gene expression signatures that may be involved in cell to cell communications. The strategy we followed, which is summarized in Fig 1, is based on several cycles of computational analysis, hypothesis generation and experimental validation. The workflow consisted of five distinct but interconnected steps.
The overarching goal of this project was to develop a data driven strategy to identify molecular pathways involved in cell-to-cell crosstalk. We first set to test whether gene expression profiles across normal samples may correlate with the gene expression profiles from the matching tumour samples. We reasoned that if such correlated profiles exist they might be a manifestation of the signaling events between normal and tumour epithelial cells and may shed new light on the role of normal epithelia in prostate cancer.
With this in mind, we first applied relevance networks [11], a relatively simple network inference procedure, to link genes differentially expressed in normal and in tumour epithelia. We used a dataset developed by Singh et al. [12], representing the transcriptional state of 47-paired prostate tumour and adjacent normal cells samples. The resulting network (NT network) is composed of 2581 positively and negatively correlated genes (Fig 2). These were subdivided in 1600 gene expression profiles in normal epithelia (referred from now on as ‘normal-expressed’ genes) and 981 gene expression profiles in tumour epithelia (referred from now on as ‘tumour-expressed genes’. The NT network was grouped into 68 modules by using GLay [13], a community detection method that maximizes inter-module connectivity. Only three modules contained more than twenty nodes and thus were selected for further investigation (Fig 2). This arbitrary threshold was used to make sure that a sufficiently large number of genes was present in each module for subsequent functional analysis. The NT network and its modules fitted a power law node connectivity distribution (p<10−2), consistent with the existence of a relatively small number of genes with a very large number of connections.
Module 1 displayed a marked enrichment in normal-expressed genes (Fig 2B, p<10−4) and module 2 showed enrichment in tumour-expressed genes (Fig 2C, p<10−4). In module 3, the frequency of normal- and tumour- expressed genes was as expected by random chance (Fig 2D, p = 0.41). Interestingly, the most connected genes in module 1 and 2 represented profiles from the tissue that was less represented (p<10−4). The most extreme case was module 1 where 19 of the 20 most connected genes were tumour-expressed genes (expected frequency was 1). Although module 3 showed no preferential tissue distribution, it still showed a higher than expected frequency of tumour-expressed genes among the 20 most connected genes (p<10−2). Functional analysis of the genes represented in each module showed that these were enriched in a wide spectrum of biological functions (Fig 2A and S1 Table).
The results described above (Fig 2) are consistent with the notion that a relatively small number of genes expressed in either normal or tumour epithelial cells may control communication signals that can either modify or respond to the molecular state of the adjacent tissue.
In order to mine the NT network for such signals we developed the polarization index (pol), a novel gene connectivity metric. We design this index to represent genes that may exert an effect on the adjacent cell type only when expressed in one specific tissue. This scenario implies a directional signal, which is for example typical of soluble factors encoded by tumour suppressor genes or oncogenes. In the case of tumour suppressor genes, these may have lost the ability to control tumour cell proliferation via autocrine signaling but they may retain that function when expressed in adjacent stromal cells by a paracrine signal.
We formalized this scenario as follows:
Considering that a given gene gi can be expressed in both normal and tumour tissue, we define fi as the number of tumour-expressed genes that correlate with the normal-expressed gi. Similarly, we define bi as the number of normal-expressed genes that correlate with the tumour-expressed gi. We define the polarization coefficient for gene gi as:
poli=fi−bifi+bi+ε
(1)
ε is a small positive constant designed to stabilize the ratio when fi and bi are small.
Poli has a number of desirable properties: its value is proportional to the asymmetry in the number of correlated genes with gene i in the two tissues while its sign gives the direction of the effect. This metric tend to 1 or -1 for fi >>bi or fi << bi, respectively.
We computed this index for all genes represented in the NT network (Fig 2) and discovered that, independently of the threshold used, it is distributed accordingly to a multimodal distribution with three peaks (Fig 3A). The highest frequency of the distribution is centered on zero whereas a smaller number of genes show polarization coefficients close to +1 and -1.
We focused subsequent analysis on genes with pol>|0.75|, a very stringent threshold that we found to have less than 1 in 8000 false positives (Fig 3B and S2 Fig). This very stringent threshold identified 146 and 244 positively and negatively polarized genes, respectively (Fig 3C and S2 Table).
Functional analysis of the polarized genes using Gene Ontology and the Ingenuity database shows a statistically significant enrichment in functions key to cancer biology (Fig 3D and 3E). The main functions significantly enriched in the negatively polarized genes are cell death of tumour cell lines, migration of tumour cell lines, necrosis and proliferation of PC cell lines (Fig 3D). The main functions enriched in the positively polarized genes are proliferation of cells, migration of cells, invasion of cells and apoptosis of tumour cell lines (Fig 3E). Moreover, 107 positively polarized genes (53% of the 204 genes that had functional annotation) are linked to the Gene Ontology term cell communication and therefore represent a class of proteins potentially mechanistically involved with cell crosstalk (S3 Table). Interestingly, only positively polarized genes are significantly enriched in this functional term (FDR<10−2). Manual curation into the role of the positive and negatively polarized genes using available literature and online databases was consistent with the computational analysis. In Table 1 we report the positively and negatively polarized genes that are either secreted factors (potential paracrine signals) or factors partitioned at the cell surface (potentially involved in cell-cell communication via direct contact) or transcription factors that may regulate the expression of cell communication genes.
Additionally, almost all the network hubs described in Fig 2 are characterized by a high polarization coefficient (either positive or negative).
In order to investigate the potential role of polarised genes in cell-to-cell communication we first identified their first neighbours in the NT network and then we tested the resulting gene lists for functional enrichment. We could identify 1223 normal-expressed genes as targets of tumour-expressed negatively polarised genes and 794 tumour-expressed genes as targets of normal-expressed positively polarised genes (S4 Table). We discovered that there was a significant overlap between them (520 genes, p<10−3) (S3 Fig) suggesting that although positively and negatively polarised genes are by definition different, they may ultimately target the same biological processes, in tumour and normal cells respectively. This hypothesis was supported by the functional analysis, which identified a set of terms enriched in the overlapping set of gene targets. Among these there were regulation of cell death, response to growth factor, cell adhesion and extracellular region part (S3 Fig and S5 Table).
We reasoned that if the cell-to-cell communication model we developed around the gene polarization index is correct, we should be able to modulate the putative targets of polarized genes by reconstructing an in vitro system where normal and tumour prostate cells share the same micro-environment. We performed such experiment by using a trans-well co-culture system where normal (RWPE1) and tumour (DU-145) epithelial cell lines are separated by a semipermeable membrane. In these experiments either tumour or normal cells were inserted into dishes already containing tumour cells. This experimental set up represents the aspect of the prostate tissue in which cancer epithelial cells sits in proximity but not necessarily are in direct contact (paracrine signals). Four sets of samples were processed for expression profiling 24 hours after the start of the experiment. These were: 1) RWPE1 cultured with RWPE1, 2) DU-145 cultured with DU-145, 3) RWPE1 cultured in the presence of DU-145 and 4) DU145 cultured in the presence of RWPE1. Genes whose expression in tumour cells is influenced by the presence of normal cells were identified by direct comparison between gene expression in DU-145 cultured on their own and gene expression in DU-145 cultured in the trans-well system in the presence of RWPE1. Similarly, we identified genes whose expression in normal cells depended on the presence of tumour cells by direct comparison between gene expression in RWPE1 cultured on their own and RWPE1 grown in the trans-well system in the presence of DU145. We considered the two sets of genes identified by this simple differential expression analysis as the experimental equivalents of the predicted targets of positively and negatively polarized genes, respectively.
Consistent with the analysis of the targets of polarized genes (S3 Fig) we found a significant overlap between genes differentially expressed in normal and tumour cells as a result of co-culture (Fig 4A, p<0.01). We also discovered that a significant percentage of genes up regulated in tumour cells were down regulated in normal cells and vice versa (Fig 4A). This is consistent with the results of a principal component analysis of these data showing that the variation between normal and tumour cells following co-culture followed anti-parallel trajectories (Fig 4B).
Next, we compared the predicted targets of polarized genes and the experimentally determined transcriptional signatures. The overlap between the differentially expressed genes in the co-culture system and the predicted targets of polarized genes was significant both at gene (S4 Fig) and at functional level (Fig 4C and 4D). We concluded that remarkably, the in vitro system was able to recapitulate a significant component of the transcriptional network inferred from the clinical study.
The functional analysis of these gene signatures revealed enrichment in several important cellular functions that are very relevant in cancer (e.g. regulation of growth, apoptosis and cell adhesion). Since we could not identify a specific direction in differential gene expression we set to determine whether change in the transcriptional state of co-cultured cells impact a relevant cancer phenotype. We therefore performed a battery of in vitro tests on tumour cells, using the same trans-well co-culture system described above. Here we assessed whether the transcriptional signatures defined by our computational analysis and validated by the in vitro co-culture system may truly reflect a cancer relevant phenotype. We found that the presence of normal epithelial cells induced several phenotypic changes in tumour cells. More specifically, population doubling time (PDT) in tumour cells cultured in the presence of normal cells was considerably longer than in tumour cells cultured on their own (30 hours against 18 hours, Fig 5A). Cell numbers at the end of the experiment were consistent with this finding and also revealed that additional tumour cells in the trans-well promoted survival (Fig 5B). The apoptosis test revealed that normal cells did not have any effect but tumour cells surprisingly increased the number of tumour apoptotic cells (Fig 5C). We then tested the formation of cell clusters and recorded the number of cell clusters (Fig 5D), the size of clusters (Fig 5E) and the area of the dish occupied by single cells (Fig 5F). Normal cells reduced the number and size of clusters and increased the area occupied by single cells whereas tumour cells had the opposite effect (Fig 5D–5F).
Consistent with these findings, conditioned media from COS cells overexpressing the tumour suppressor gene SLIT2, one of the most positively polarized genes (pol = 0.99) which is expressed at higher levels in normal prostate tissue compared to tumour (S5A Fig), was able to dramatically reduce tumour cell clone formation in a Matrigel in vitro Clonogenic assay (Fig 5G).
All of this data is consistent with the normal cells effectively 'normalising' the phenotypic characteristics of the tumour cells.
Having inferred and experimentally validated a transcriptional network representing the interaction between normal and tumour prostate epithelial cells we then hypothesised that expression of genes within the network may be influenced by genetic/epigenetic modifications and/or correlate to tumour features and clinical outcome.
We first checked whether the expression of polarised genes might be influenced by DNA methylation, a common mechanism for transcriptional silencing in cancer. By mapping genes known to be re-expressed in prostate cancer cell lines, following exposure with DNA hypomethylating agents [14][15][16], we could show that methylation significantly affect the expression of 30 of the 245 positively polarized genes and 12 of the 146 negatively polarised genes in tumour cells (S6 Fig and Fig 6 and S6 Table). Although the percentage of genes affected by methylation is relatively small, the number of positively polarised genes whose expression is affected by methylation was significantly higher than expected by random chance (S6 Fig).
Next we assessed the role of copy number variation (CNV). We selected an independent dataset [17], which included genetics (CGH), gene expression and clinically relevant variables (S7 Table). First we tested whether the expression of polarised genes was directly affected by CNV. We could only identify 9 polarised genes with significant correlation (p<0.01) between their CNV and expression (S8 Table). Next we developed a hierarchical Bayesian model to identify whether epistatic CNV could explain the expression of polarized genes in tumour cells. We were able to show that the expression of 70 polarised genes could be explained by CNV in seven genomic regions (S9 Table and Fig 6). Three of these included genes with known function (ATAD1, GRHL2 and KCNB2) (Fig 6 and S10 Table). Interestingly, the large majority of polarized genes whose expression was linked to CNV were mainly positively polarised (59 out of 70).
Finally, we tested whether the expression of polarised genes was related to tumour features and clinical outcome. Indeed we found that the expression of a large number of polarised genes (130) was linked to Gleason score. A smaller number of genes (18 and 1) were linked to PSA antigen and T stage, respectively (Fig 6 and S11 Table).
The integration of these associations using a network representation revealed 173 polarised genes linked either to regions affected by CNV (89 genes) and/or to tumour features (84 genes) (Fig 6). Remarkably, while the expression of none of the polarised genes could be linked to survival, 132 of them were linked to time free of recurrence (FDR<5%, S7A Fig and S12 Table). Interestingly we could also show that polarised genes linked to CNV did show significantly lower p-values than polarised genes only linked to Gleason score (S7B Fig) supporting the clinical relevance of the epistatic effects identified by the computational model.
A group of 58 positively polarised genes and 6 negatively polarised genes (Fig 6 and S7B Fig and Table 2) were linked to both CNV and Gleason score. We found that the large majority of genes in this group (57/58) were negatively associated to Gleason score and positively correlated to time free of recurrence (S7B Fig and S12 Table).
Intriguingly, these were highly enriched in Cytoskeleton proteins (24 out of 51, over-represented in the GO term Cytoskeleton at FDR<10−8) (Table 2).
We then tested the expression of 36 out of the 58 genes that were profiled in a dataset representing normal and tumour cells which were laser micro-dissected from prostate cancer specimens [18] (Fig 7). This analysis showed that 11 out of 36 are differentially regulated and that all except 1 were down regulated in the tumour tissue (Fig 7), an observation that is consistent with the direction of correlation with the survival free of metastases.
Among these 58 genes, 8 represented genes involved in formation of cell projections (ACTN1, CALD1, CLIC4, DPYSL3, DBN1, ILK, PAFAH1B1 and RTN4) and six (ACTN1, CCND2, FLNA, FLNC, ILK and MYL9) mapped on the KEGG pathway focal adhesion (FDR<1%). Also, five were proteins known to be associated to the Golgi apparatus and involved in protein secretion (SEC23A, CRYAB, FLNA, NUCB1 and PRNP). Among these were several genes with known tumour suppressor activity (e.g. FLNA [19], FBLN1 [20], MYL9 [21], CLIC4 [22] and SEC23A[23]).
Here we have described a relatively simple network inference and analysis procedure, explicitly designed to learn cell communication networks from observational data. This is the first example of an open ended reverse engineering strategy that explicitly searches for cell communication networks from observational data. Our approach also provides clues on the role of normal epithelial cells in prostate tumour progression. The application of our analysis strategy (Fig 1) to prostate cancer revealed that normal epithelial cells may have a more important role in controlling tumour expansion than previously suspected. The applicability of this approach is broader and indeed it opens important avenues for better understanding the whole network of signals regulating cell communication in both normal and pathological scenarios.
Since the large majority of efforts have focused on understanding the role of fibroblast and endothelial cells in cancer, the interface between normal and transformed epithelial cells is still not clearly understood. Our analysis suggests that normal epithelial cells exert a “normalizing” effect on tumour cells, up to an advanced stage of tumour progression.
A number of recent studies have suggested that at the initial phase of tumour expansion, normal epithelia could provide a tumour suppressive environment that cancer cells need to overcome to develop a tumour. So far, tumour suppressor activity of normal epithelial cells has been studied in cell culture systems replicating early transformation events in epithelia [24]. These models include kidney and mammary epithelial cells in culture where only a few cells are selectively transformed by oncogenic transformation or inhibition of tumour suppressor genes [24]. In these conditions, transformed cells are excluded from the epithelia and out grown by normal epithelial cells. It has been suggested that additional mutations and/or alterations in the adhesion properties of tumour cells may be needed to overcome the tumour suppressive effects and allow for clonal expansion [24]. However, the precise molecular events underlying this process are still unknown.
Our work therefore provides further evidence of the tumour suppressor effects of normal epithelial cells and supports the concept that although tumour cells obviously eventually overcome these normalizing signals, the effect of normal epithelia may be relevant for the entire clinical history of prostate cancer.
The models we have developed provide a link between genetic mutations and the expression of polarized genes in tumour cells. Remarkably, the functional profile of mutated genes is consistent with a pivotal anti-tumour role of the apical junctional complex and the protein secretion machinery.
Among the three genes we have identified as potential epistatic regulators, GRHL2 is known to be a transcription factor known to play a pivotal role in cancer progression [25][26][27]. GRHL2 regulates epithelial cell differentiation by effectively regulating the expression of genes of the epithelial apical junctional complex [28]. It controls the expression of the adherents junction gene E-cadherin and the tight junction gene claudin 4 (Cldn4) and has been linked to both pro and anti-tumour activity [25]. Moreover, GRHL2 up regulates the human telomerase reverse transcriptase (hTERT) gene during cellular immortalization of oral squamous cell carcinoma cells [29]; it is a proto-oncogene in breast cancer cells [25]; it regulates proliferation of hepatocellular carcinoma cells [30] and is a suppressor of epithelial-to-mesenchymal transition in breast cancer [31].
Our model predicts that increase expression of GRHL2 due to CNV down regulates the expression of a set of polarized genes that precisely encode for components of the cytoskeleton and are involved in focal adhesion and cell migration. Several of these genes are extracellular factors and one of them (SEC23A) has been found to control secretion of anti-tumour factors in breast cancer [23].
These observations lead to the hypothesis that increased expression of GRHL2 in tumour cells may result in the deregulation of at least two different types of tumour suppressor signals, one dependent on the establishment of focal adhesion junctions and the other directly affecting secretion of anti-tumour factors. This chain of events may contribute to tumour transformation and metastases formation and at the same time could make tumour cells sensitive to the same tumour suppressor signals that continue to be produced by adjacent normal epithelial cells. The in vitro system we have used to validate our model shows that normal epithelial cells are able to exert anti-tumour effects even if normal and tumour epithelial cells were separated by a semi-permeable membrane, suggesting that soluble factors may be playing a major role in tumour suppression. Secretion of the highly positively polarized gene SLIT2 from normal epithelial cells has the potential of exerting a tumour suppressor activity as shown by our clonogenic assay on tumour cells exposed to diluted conditioned media.
More broadly, there is strong support in the literature linking several of the positively polarized genes to tumour suppression.
More precisely, FLNA[19] FBLN1[20], MYL9 [21], CLIC4 [22] all have demonstrated tumour suppressor activity. It has been shown that Filamin A (FLNA) exerts anti-tumour activity via at least three different mechanisms. It represses MMP-9 expression reducing cell migration in prostate cancer. It controls focal adhesion and androgen-related cell migration in human fibrosarcoma [19] and Cyclin D1/cyclin-dependent kinase 4 mediated cell migration in breast cancer [32]. The myosin light chain (MYL9) in stroma has been shown to predict malignant progression and recurrence-free survival in prostate cancer [21]. Fibulin 1 (FBLN1) is down regulated in a number of tumours, including prostate [33]. CLIC4 was first characterized as intracellular chloride channel, later shown to be involved in signaling, cytoskeleton integrity and differentiation [34] and is a tumour suppressor gene in cutaneous squamous cell cancer [22].
The reverse engineering approach we have adopted is based on the assumption that gene co-expression is either directly or indirectly a reflection of important underlying mechanisms of gene regulation and as such it can reveal novel biological networks. While this concept is well accepted in the scientific community, it remains true that correlation does not necessarily imply causation, hence the importance of experimental validation.
However, for a number of candidates, it is possible to hypothesize a mechanism whereby highly polarized genes may directly affect adjacent cells. For example, a number of them are secreted factors that can work as paracrine signals or membrane proteins known to be involved in cell communication (Table 1). This is the case for Slit-2 that we have experimentally verified by treating prostate cancer cells with conditioned media derived from cells over-expressing the recombinant protein. Others may indirectly control cell communication. This is for example the case with transcription factors (e.g. GATA2 control of IGF1 signalling [35]), proteins controlling secretion (e.g. LTBP1 control of TGFB1 secretion [36]) or proteins involved in cell migration.
Interestingly, the gene expression profiling analysis we have performed to validate our predictions, suggest that the polarization coefficient may have the ability to capture directional signals that are triggered by normal and tumour cells. However, the experimental system we have used is based on a trans-well system, which only validate paracrine signals.
We believe our approach could have a broad impact. Although, at present there are not many suitable datasets containing both disease and adjacent normal tissue, we have verified that the distribution of polarization coefficient in two additional, datasets representing kidney and liver adjacent normal and tumour tissues are similar to the one observed in prostate cancer (S9 Fig). In the future we envisage that tissue laser micro dissection and mRNA sequencing technologies may provide a very powerful combination for the identification of genome wide cell communication networks.
The approach we have developed has the advantage to reverse engineer cell communication networks in the absence of any prior information. In this respect, the method is different from the recently developed computational method developed by Choi et al [10]. The latter has been successfully applied to understanding the relationship between stroma and cancer cells in a model of lung tumour metastases and is based on comprehensive ligand-receptor network information, which can be extracted from several knowledge databases. We envisage that the integration of these knowledge driven approach within the framework of statistical learning will allow the development of a more powerful set of methodologies.
The study we have performed relies on cross-sectional data and therefore the correlations we estimate do not take into consideration the hierarchical sequence of events that characterize cell communication dynamically. However, such dynamics can be captured using in vivo models of tumor expansion [37]. In such scenarios, different computational methodologies may be used to reverse engineer underlying gene regulatory networks [38]. For example, suitable approaches may include ordinary differential equation (ODE) or state space models [39].
In conclusion, the approach we have pioneered is likely to provide a general strategy to ‘learn’ the structure of cell-to-cell communication networks in diseased and physiologically normal tissues. We anticipate that the availability of a viable strategy to infer cell communication networks will stimulate the development of experimental studies representing the molecular state of adjacent tissues and their functional interactions in physiology and disease.
This analysis initially focus on the dataset developed by Singh et al.[12] representing the transcriptional state of 47 paired prostate tumour and adjacent normal cells samples. Raw Affymetrix microarray data were normalized and processed before analysis to remove low variant and low expressed genes. Further details of the procedures can be found in S1 File.
The analysis linking copy number variation (CNV), gene expression, tumour features and clinical outcome was performed on the dataset developed by Taylor et al. [17], which consist of 231 tumour samples. Raw comparative genomic hybridization Agilent data was processed as detailed in S1 File.
The dataset developed by Tomlins et al [18] was used to test the expression of polarized genes in laser capture micro-dissected low and high-grade tumour and normal prostate tissue. In the Tomlins et al study, tumour grading was determined by Gleason score. A scores of 3 determined a low-grade tumour, a score of 4 or greater determined a high-grade tumour. Raw data was downloaded and normalized using the “marray” BioConductor package in R [40].
All data processing was performed in the statistical environment R.
Network inference was performed using a relevance network approach [11]. Non-linear Spearman ranking correlation (rs) was used to infer gene-to-gene correlations. In order to estimate the number of significant correlations, 100 bootstrap versions of the original dataset were used for each dataset to draw the null distribution of rs expected by chance. The bootstrap distribution was used to estimate a p-value, which was subsequently corrected for multiple-test using an FDR correction procedure [41] (S1 Fig). The use of the relevance networks based on the Spearman correlation coefficient has advantages respect to more complex reverse engineering methods such as the mutual information based ARACNE [42] algorithm. Spearman correlation measure both positive and negative correlations and is better suited for datasets with a smaller number of samples. We used a threshold of rs >|0.75| (FDR<10−2) to select significant connections (NT network).
The NT network, representing statistically significant correlations between genes expressed in normal and tumour tissues, was modularized using the community finding algorithm GLay [13], as implemented in the network analysis tool Cytoscape [43].
The algorithm begins by setting each node into a separate community and progressively merges those with the maximum increase to the modularity score. The hierarchical merging tree is cut at the point where maximum modularity is achieved.
Connectivity analysis of the whole network and of the three largest modules (defined as larger than 20 nodes) was performed using the network analysis tool NetworkAnalizer [44], also implemented as a Cytoscape plug-in.
The general definition of the polarization index for a given gene i, have been given in the result section (Eq 1). The analysis described in this paper has been performed with the parameter ε set to 1. Additionally, pol was set to 0 if the absolute difference between f and b was lower than 20 to avoid high pol values for low number of connections.
In order to acquire confidence in the biological relevance of the polarization index we derived a null hypothesis distributions for estimating the likelihood that a given polarization value derives by random chance. This represented a scenario where the overall properties of the data are conserved in the absence of any interaction between normal and tumour samples.
Random data sampled from the Singh et al. dataset were used to compute normal and tumour correlation matrices. Each matrix was fitted by a multivariate Gaussian model to generate a synthetic dataset. Synthetic datasets were then used to compute the correlation matrix whose distribution predictively resembles that of the original dataset. Subsequently, the polarization index was estimated from this correlation matrix. The multivariate fitting and subsequent random dataset generation was performed using the function rmvnorm within mvtnorm packages [45] in the R statistical software environment (S2 Fig). Significantly polarized genes have been defined as poli > |0.75|. At this threshold we did not observed any false positives in the 8000 random simulations performed.
Although the expected level of contamination of tumour tissue with normal cells is expected to be very low, we devised a computational strategy to ask whether the polarization index could arise as a result of contamination of tumour samples with normal cells. We computed the polarization index between two simulated datasets that reproduce a situation where both tumour and normal samples are derived from normal tissues with added noise, thereby simulating variation that is consistent with a true microarray experiment (S8A Fig). Firstly, an adapted model of the type derived by Jain et al [46] was used to estimate the experimental noise across replicates (S8B Fig). Random Gaussian noise [47] derived from this noise model was then added to the normal tissue dataset to create a synthetic normal and synthetic contaminated tumour dataset. The intensity of the added noise was controlled by adding a scaling factor γ, which was chosen to match the distribution of correlations between genes in the synthetic datasets with the distribution observed in the real data (S8C Fig). The distribution of the polarization coefficient is consistent with the notion that even high levels of contamination cannot explain the observed distribution of polarization coefficient (S8D Fig).
In order to test whether the trimodal distribution of the polarization coefficient could be observed in other cancer types in addition to prostate cancer, we analyzed two additional public domain datasets representing kidney [48] and liver [49], respectively. Only paired data corresponding to tumour and normal from the same tissue were used. Only one pair of samples per individual was used. In general, normal tissue was adjacent to the tumour. Datasets were normalized and processed before analysis as for the main prostate cancer dataset. Results are shown in S9 Fig.
Lists of polarized genes or their correlated genes were analyzed for enrichment of curated functional categories using the QIAGEN Ingenuity Pathway Analysis tool (IPA, www.qiagen.com/ingenuity). Enrichment of Gene Ontology (GO) terms and KEGG pathways was determined using the web-based tool gprofiler [50]. In order to reduce redundancy in the functional terms we used REVIGO and selected the functional terms with dispensability index equal to zero. Unless stated otherwise gProfiler functional clusters were considered for further investigation if they had a FDR<1%.
Normal (RWPE1) and tumour (DU145) prostate cell lines were co-cultured in a transwell system (transwell I used was from Nunc, Loughborough, UK, Cat. 12-565-286; Pore size, 0.2 μm.) for 24 hours in the presence of DMEM containing 10% fetal calf serum. The experiment was performed in triplicate with DU145 alone or DU145 co-cultured with either RWPE1 or DU145 in the insert. Cells from all compartments were harvested and RNA extracted using a Qiagen RNeasy kit according to the manufacturer's instructions (Qiagen, USA). Custom-made oligonucleotide arrays were manufactured using the Operon Human Oligo set, version 3.0 [51] and then hybridized with Cy3 labeled probe, as described in Sarti et al. [52]
Phenotypic cell analysis was carried out in Becton Dickinson TC treated 96-well plates. 2.5 x 103 cells were seeded per well in DMEM containing 10% fetal calf serum. 24 hours later, some wells were fitted with inserts also seeded with 2.5 x 103 cells per insert. Two days later inserts were removed, media was aspirated from the wells and cells were fixed with 85% ice-cold ethanol for at least two hours. After fixation cells were stained with propidium iodide (10 μg per ml propidium iodide, 100 μg per ml RNase A, 0.1% Triton X-100 in PBS, 100 μl per well). Plates were incubated at 37°C for 20 min in the dark and then analysed by laser scanning cytometry (Acumen Explorer, TTP Labtech.).
The intensity of the propidium iodide fluorescence was proportional to the DNA content of the cells and was measured on a linear scale. Single healthy and apoptotic cells were identified based on nuclear size and DNA content [53]. Cell clusters were defined as single scanned objects that contained multiple nuclei. The size of the clusters was defined as the ratio of the total nuclear area within a cluster divided by the size of an average nucleus in the same population.
Single-cell suspensions for either PC-3 or DU 145 cells, were prepared from 80% confluent cultures. The cells were counted and plated onto 24-well flat-bottomed plates using a two-layer soft agar system with 1x103 cells in 400 μl of media per well, as described previously [54]. The feeder layer was prepared with agar (1%) equilibrated at 42°C. On top of the agar layers conditioned media from COS-7 cells stably transfected with a SLIT2 expression vector, or mock transfected control, was added. After 14 days of incubation, the colonies (>50 cells) were counted using an inverted microscope. All experiments were done at least three times in triplicate per experimental point and all statistical analyses were performed using the Student's t-test.
Genes differentially expressed were first identified using SAM multi-class test [22], with a threshold of FDR<1%. Differentially expressed genes were then used as input for principal component analysis (PCA) and the first two components representing 68% of variability were plotted to visualize the relationships between the different samples (Fig 5A). Genes differentially expressed in a given cell type as a result of co-culture were identified by a 2-class SAM procedure (FDR<1%) by directly comparing RWPE1 co-cultured with DU145 (RWPE1DU145) and RWPE1 cultured in isolation (RWPE1) or by comparing DU145 co-cultured with RWPE1 (DU145RWPE1) and DU145 cultured in isolation (DU145).
Predicted targets of polarized genes and the differentially expressed genes were then compared using a Fisher exact test. The comparison of these gene lists at the functional level was performed by plotting the frequency of genes in each functional term for predicted targets (x axis in Fig 5C and 5D) against differentially expressed genes (y axis in Fig 5C and 5D).
In order to address the hypothesis that disease linked genetic mutations such as copy number variation (CNV) may influence the expression of polarised genes in tumour cells and that this, in turn, may be predictive of tumour features and clinical outcome we implemented a data analysis pipeline based on a number of advanced statistical procedures.
We used an independent dataset [17], which had comparative genomic hybridisation (CGH), gene expression and comprehensive information on tumour features and clinical outcome for a total of 231 tumour samples.
Firstly, in order to prioritise relevant genetic abnormalities we used ANOVA to rank CGH signals linked to tumour features and/or one of the clinical outcome variables (see S1 File for further details). The top 2017 probes in the ranked list were selected as an input of the modelling procedure.
We then mapped the 391 polarised genes we originally identified on the independent dataset. Next, we used the selected CGH data and the polarised gene expression profiling dataset as an input of a hierarchical Bayesian model [55] to find association between polarized gene expressions and CNV (see S2 File for details of the modeling procedure as applied here). Next, we fit an ANCOVA model for each gene expression on the tumour features. We then computed correlations for the significant associations (p<0.05) and integrated all information in a network format using the Cytoscape [43] software (Fig 6). Finally, we selected all polarized genes represented in the network and performed a survival analysis testing the hypothesis that their expression in tumour cells could be linked to clinical outcome (survival and time free of recurrence).
Survival analysis was performed as below. Briefly, for each gene we defined an optimal cutoff to separate patients in two groups of low and high-expressing tumours, using procedure described in Budczies et al [56]. Using this cut off, we dichotomized each gene expression level that was then used to fit a cox regression model.
|
10.1371/journal.pgen.1005898 | Role of the BAHD1 Chromatin-Repressive Complex in Placental Development and Regulation of Steroid Metabolism | BAHD1 is a vertebrate protein that promotes heterochromatin formation and gene repression in association with several epigenetic regulators. However, its physiological roles remain unknown. Here, we demonstrate that ablation of the Bahd1 gene results in hypocholesterolemia, hypoglycemia and decreased body fat in mice. It also causes placental growth restriction with a drop of trophoblast glycogen cells, a reduction of fetal weight and a high neonatal mortality rate. By intersecting transcriptome data from murine Bahd1 knockout (KO) placentas at stages E16.5 and E18.5 of gestation, Bahd1-KO embryonic fibroblasts, and human cells stably expressing BAHD1, we also show that changes in BAHD1 levels alter expression of steroid/lipid metabolism genes. Biochemical analysis of the BAHD1-associated multiprotein complex identifies MIER proteins as novel partners of BAHD1 and suggests that BAHD1-MIER interaction forms a hub for histone deacetylases and methyltransferases, chromatin readers and transcription factors. We further show that overexpression of BAHD1 leads to an increase of MIER1 enrichment on the inactive X chromosome (Xi). In addition, BAHD1 and MIER1/3 repress expression of the steroid hormone receptor genes ESR1 and PGR, both playing important roles in placental development and energy metabolism. Moreover, modulation of BAHD1 expression in HEK293 cells triggers epigenetic changes at the ESR1 locus. Together, these results identify BAHD1 as a core component of a chromatin-repressive complex regulating placental morphogenesis and body fat storage and suggest that its dysfunction may contribute to several human diseases.
| The importance of epigenetics in regulation and dysfunction of metabolic pathways is increasingly recognized but the underlying mechanisms and molecular actors involved remain incompletely characterized. Here, we provide evidence that the heterochromatinization factor BAHD1 cooperates with MIER proteins to assemble chromatin-repressive complexes that control a network of metabolic genes involved in placental and fetal growth and in cholesterol homeostasis.
| Chromatin-based transcriptional repression is mediated by macromolecular complexes containing proteins involved in chromatin writing, reading, erasing and remodeling activities. The combinatorial assembly of subunits with transcription factors affects cell-specific gene expression in response to developmental, physiological or environmental stimuli. Chromatin-repressive complexes control key pathways during embryonic development and adult life; as a consequence, deregulation or abnormalities in their components can lead to a wide range of pathological processes [1, 2]. The importance of chromatin-modifiers in development, cell differentiation and disease is well illustrated for three complexes containing the histone deacetylases HDAC1 and HDAC2: NuRD [3, 4], Sin3A [5] and CoREST [6, 7] (For reviews, see [2, 8–10]). By a proteomic approach, we found that the Bromo-Adjacent-Homology domain-containing 1 (BAHD1) protein co-purifies with HDAC1/2, together with heterochromatin proteins HP1 and KAP1 (or TRIM28) in human embryonic kidney (HEK) 293 cells, suggesting that BAHD1 is a core component of a novel HDAC1/2-associated complex [11]. BAHD1 also interacts with the Methyl-CpG-binding protein MBD1 and the H3K9 methyltransferases (KMT) SETDB1 [12] and SUV39H1 [13] and acts as a repressor, pointing to a role of BAHD1 in heterochromatin-mediated transcriptional repression [12]. In agreement with this, overexpression of BAHD1 in human cells induces large-scale chromatin condensation [12] and changes in the DNA methylation landscape [14].
BAHD1-associated heterochromatic domains lack acetyl histone H4 and partially overlap with HP1α, a marker of constitutive heterochromatin, and/or with H3 trimethylated at lysine 27 (H3K27me3), a marker of facultative heterochromatin [12]. Furthermore, when overexpressed in human female cells, BAHD1 is enriched at the inactive X chromosome (Xi), a paradigm of facultative heterochromatin [12]. A study in mouse embryonic stem cells (mESCs) recently reported that BAHD1, HDAC1 and HDAC2 are pulled-down by CDYL, a transcriptional co-repressor that may play a role in the maintenance of the Xi [15]. Taken together, these data suggest that BAHD1 is a component of HDAC1/2-associated complexes involved in a variety of epigenetic mechanisms.
A single gene encodes BAHD1 in vertebrates and no ortholog is found in invertebrates or plants, suggesting that BAHD1 has vertebrate-specific functions. However, the low expression of BAHD1 in cell lines has hampered its functional characterization and, so far, the BAHD1 regulatory gene network is poorly characterized. We identified the insulin-like growth factor II (IGF2) transcript and its antisense transcript (IGF2AS) as BAHD1 targets in HEK293 embryonic cells [12]. We also demonstrated that infection by a bacterial pathogen triggers BAHD1-mediated repression of Interferon-Stimulated Genes (ISGs) in epithelial cells [11]. However, apart from bacterial infection, signals that control the expression and/or activity of BAHD1 are unknown.
The aim of the present study was to determine the physiological functions of BAHD1. We show that disruption of the Bahd1 gene in the mouse leads to a placental growth defect associated with low birth weight and neonatal death, hypocholesterolemia and decreased body fat in surviving adults. Proteomic studies of BAHD1-associated proteins identify MIER proteins as novel BAHD1 partners. Our extensive characterization of the transcriptome strongly suggests that BAHD1-MIER complexes repress genes involved in the control of steroid/lipid metabolism both in mouse and human cells.
We searched for tissue-specific expression of the BAHD1 gene by a survey of referenced transcriptome datasets (S1 Table) and found that BAHD1 mRNA levels are low and do not vary much between tissues, when compared to a set of housekeeping or tissue-specific genes. This observation is consistent with the lack of detection of the endogenous BAHD1 protein in mammalian cultured cells and suggests that BAHD1 basal levels are low. In order to identify functions of BAHD1 in biological processes, we studied the physiological consequences of BAHD1 inactivation by performing a large-scale phenotyping of Bahd1 haplo-deficient (Bahd1+/-) mice [11]. The analysis was carried out on 18 heterozygous (HET) and 16 wild type (WT) littermates (results are detailed in S1 Text and S2 Table). Mice were first fed a standard chow diet (CD) for 14 weeks and then a high fat/high carbohydrate diet (HFHCD) for 16 weeks. HET Bahd1+/- mice did not show any morphological, sensory or cardiac abnormality and no change in bone density, body weight and fat, when compared to WT littermates, and the blood chemistry and hematology parameters were within the normal range. However, some slight and gender-specific modifications of glucose or cholesterol parameters were observed in HET mutants. At the age of 10 weeks, HET male mice displayed a significant hypoglycemia when compared to WT mice (Fig 1A). After the switch to HFHCD, blood levels of glucose remained lower in 30-week-old HET than in WT male mice, although difference did not then reach statistical significance (S2A Table). In contrast to males, glycemia was not affected in female HET mice whatever the diet (Fig 1A) but female HET fed with HFHCD displayed a slight decrease in blood levels of total cholesterol, high-density lipoprotein (HDL) and low-density lipoprotein (LDL) compared to WT littermates (Fig 1B). However, values were statistically significant only for LDL concentrations (S2B Table).
We next crossed Bahd1-heterozygous mice to evaluate potential metabolism defects in the Bahd1-null context. Most Bahd1-/- (KO) pups died within the first days of life, indicating that the Bahd1-null mutation leads to perinatal death. Only 5 Bahd1-/- out of 200 genotyped pups (2.5%) survived birth. At six weeks, these 5 survivors, all males, displayed reduced sizes and weights compared to control littermates (Fig 1C). After several months, they reached the length of Bahd1+/+ littermates (Fig 1C and 1D), but kept a significant lower body weight (Fig 1E), characterized by reduced fat and lean mass (Fig 1F). Quantification of blood parameters highlighted lower levels of total cholesterol, HDL and LDL in plasma of Bahd1-/- mice than in Bahd1+/+ littermates at 3–5 months (S1A Fig), 7 and 18 months of age (Fig 1G and 1H). Glucose and leptin levels were also significantly lower in 18 month-old Bahd1-/- mice than in Bahd1+/+ littermates (Fig 1H). The other measured blood parameters were not significantly different between KO and WT mice (i.e. insulin, adiponectin, Fig 1G and 1H; triglycerides, free fatty acids, glycerol, aspartate and alanine amino transferases, urea, creatinine, albumin, glucagon, GIP, S1B Fig). Together, these results demonstrate a biological function of BAHD1 in controlling lipid and carbohydrate metabolism.
The high neonatal mortality rate of BAHD1-KO pups suggested that BAHD1 could have important functions during fetal life. In order to address this question, we examined embryos just before birth at the embryonic day 18.5 (E18.5). Bahd1-/- fetuses were present at Mendelian frequency, had normal morphology, were alive and exhibited the breathing reflex. However, their weight was decreased by 30% (Fig 2A) when compared to WT or HET fetuses. In addition, Bahd1-/- embryos exhibited a smaller placenta, with a reduction of 30% in the circumference and 55% of the area (Fig 2B), when compared to placentas of WT or HET embryos. At an earlier stage (E16.5), Bahd1-/- placentas were also smaller than Bahd1+/+ ones (Fig 2B). Histology of placenta sections with hematoxylin and eosin (HE) staining showed that Bahd1-/- placentas comprised the two fetal compartments (labyrinthine zone (Lz) and junctional zone (Jz)) and the maternal compartment (decidua basalis (Db)). However, the Lz surface area was significantly reduced and the Jz and Db were thinner in Bahd1-/- placentas than in control littermates (Fig 2C). The junctional zone plays an important role in hormone synthesis, while the labyrinthine zone is critical for materno-foetal nutrient exchange [16, 17]. The Jz comprises fetal spongiotrophoblast cells and trophoblast glycogen cells (GCs). The exact origin and function of GCs is unknown, but they are believed to provide an important glucose supply for fetoplacental development. Evidence indicates that they differentiate early (E.6.5) in the ectoplacental cone at the origin of the Jz and migrate into the maternal decidua at about E12.5 [18, 19]. We used periodic acid-Shiff (PAS) staining to examine if GCs were altered by the Bahd1 null mutation. PAS staining confirmed the reduced thickness of Jz and Db in Bahd1-/- relative to Bahd1+/+ placentas at E16.5 and showed a severe reduction in the number of PAS-positive GCs both in the Jz and the Db (Fig 1D and S2A Fig). In contrast, histology of the fetal liver at E16.5 did not reveal any noticeable difference between Bahd1-KO and WT embryos (S2B Fig). Together, these results indicate that BAHD1 is required for a normal placental development and intra-uterine fetal growth. BAHD1 mutants may die after birth from metabolic defects, or secondary to altered placental exchange prenatally, impairing the development of energy stores.
To examine the consequence of BAHD1 deficiency on the placental transcriptome, we isolated RNA from Bahd1-/- placentas at stages E16.5 (n = 6 per genotype) and at stage E18.5 (n = 3 per genotype) to perform a comparative microarray analysis using Affymetrix mouse arrays. The Bahd1-null mutation altered expression of 397 and 1396 genes (FDR-BH < 0.05) at E16.5 and at E18.5, respectively, with a much higher proportion of up-regulated genes (70–80%) than down-regulated genes (20–30%), consistent with a role for BAHD1 in transcriptional repression (Fig 3A). 65% of genes deregulated in E16.5 Bahd1-KO placentas remained similarly deregulated at the E18.5 stage (214 up-regulated and 46 down-regulated genes, S3 Table). We used the DAVID software to classify genes based on Gene Ontology (GO) [20] and highlight the most significant biological processes that could be altered by BAHD1 deficiency. At the E16.5 stage, the most significant gene cluster (P value < 5.10−4) was a group of 16 genes involved in steroid metabolic processes (Fig 3B and S4A Table). Notably, 11 of these genes remained deregulated in Bahd1-/- E18.5 placentas: 7 genes up-regulated (Apoc3, Atp8b1, Cyp11A1, Insig2, Osbpl5, Pbx1, VldlR) and 4 genes down-regulated (Fabp6, Hsd17b2, Hsd17b7, LepR) (S3 Table). In addition, two steroid hormone receptor genes, the estrogen receptor alpha gene Esr1 and progesterone receptor gene Pgr, were up-regulated in Bahd1-/- placentas, both in males and females embryos and at both gestational stages, as confirmed by Real time quantitative PCR (RT-qPCR) analysis (Fig 3C).
Groups of genes involved in the response to oxidative stress, skeletal system development and blood vessel development were also over-represented in Bahd1-KO placentas (S4 Table). Several of these genes are known to be involved in placental morphogenesis, such as Gja1/Cx43, Cxcl14, Mmp2, Mmp14, Runx1 (up-regulated) and Ada, Adm, Gcm1, Gjb5 (down-regulated). Altered expression of these genes in Bahd1-KO mice is consistent with an abnormal placentation. RT-qPCR analysis of a set of transcripts confirmed the transcriptome data (Fig 3C and 3D).
Several genes previously shown to be imprinted in the mouse during development or in the adult [21] were also up-regulated in Bahd1-KO placentas. Imprinted genes, expressed from only one of the parental alleles, play important functions during mammalian development, particularly in the placenta [22]. The parental conflict hypothesis of imprinting proposes that maternally− or paternally−expressed imprinted genes restrict or increase the energy expenditure, respectively [23]. We have previously shown that in human embryonic cells BAHD1 repressed IGF2 and IGF2AS [12], which are imprinted in many tissues. BAHD1 deficiency did not affect Igf2 expression in murine placenta, but, the paternally−expressed Igf2as and ten maternally−expressed putative imprinted (Ampd3, Ano1, Dcn, Gatm, Htra3, Osbpl5, Qpct, Tfpi2, Tnfrsf23 and Wt1) were up-regulated in Bahd1-deficient E16.5 placentas, while one, Gpr1, was down-regulated. RT-qPCR assays confirmed these results using several other imprinted genes as negative controls (Fig 3E). Most of these genes are imprinted only in the placenta. However, determining placental specific imprinting can be confounded by maternally derived placental tissue and there is controversy over the imprinted status of some of these genes [24]. Strikingly, Htra3, Tfpi2, Ampd3, Gatm, Osbpl5, Qpct and Wnt1 were also consistently up-regulated in Bahd1-deficient E18.5 placentas (S3 Table) and may be functionally implicated in the impaired placental growth.
Bahd1-KO fetuses exhibited a lower weight than Bahd1-WT. This phenotype could result from placental dysfunction but also from deregulation of BAHD1 target genes in intra-embryonic tissues. To address this point, we analyzed transcriptomes of mouse embryonic fibroblasts derived from Bahd1-WT and -KO embryos at stage E13.5 (n = 3 per genotype). The results of this analysis revealed 353 up-regulated and 139 down-regulated genes (FDR-BH < 0.05) in Bahd1-KO MEFs, when compared to Bahd1-WT MEFs (Fig 3A). We focused on up-regulated genes, as these likely contain those genes directly repressed by BAHD1. Strikingly, the most significant biological process associated with up-regulated genes in Bahd1-KO MEFs was sterol metabolism (Fig 3B). However, genes associated with this process were different to those affected in Bahd1-KO placentas (S4C Table), suggesting that BAHD1-dependent gene networks are tissue-specific even though they may have a common function.
We next assessed whether BAHD1 has a similar role in repressing sterol metabolism genes in human cells, by comparing transcriptomes of HEK293 cells stably overexpressing BAHD1 (HEK-BAHD1) to that of parental isogenic cells (HEK-CT), in which the BAHD1 protein is undetectable [14]. This analysis identified 1148 transcripts (FDR-BH < 0.05) that were down-regulated in cells constitutively expressing the BAHD1 repressor. Once again, GO term classification analysis of these transcripts highlighted sterol metabolism as the most significant biological process (Fig 3B and S4D Table). Genes involved in lipid and hexose metabolism also grouped as significant clusters. Although biological functions of BAHD1-associated chromatin repressive complexes are likely to be cell type- and species-specific, and a transcriptome will identify direct and indirect gene expression changes, BAHD1 bona fide target genes may be identified by comparing different biological systems (Fig 3F). Overlapping HEK-BAHD1 and Bahd1-KO E18.5 placenta transcriptomes identified 107 potential BAHD1 targets (S5A Table and Fig 3F), the most significant gene groups where those involved in the regulation of hormone levels (ALDH1A2, BACE2, CAMK2G, CRABP2, LY6E, SCARB1, SLC16A2) and lipid biosynthetic processes (ALDH1A2, CD81, EBP, ELOVL7, LASS4, LPL, LPCAT2, LTA4H, SCARB1)(S4E Table). Overlapping HEK-BAHD1 and Bahd1-KO MEFs transcriptomes identified 44 potential BAHD1 target genes (S5B Table and Fig 3F), the most statistically significant group in term of biological processes including genes involved in sterol/steroid metabolism (DHCR24, HMGCS1, LDLR, NSDHL, SC4MOL, SREBF2; S4F Table). Only six genes were altered consistently between all the three transcriptome datasets (S5C Table). Using a tetracycline-inducible BAHD1 HEK293 line (HPT-BAHD1 cells, [11]), we confirmed that induction of BAHD1 expression (S3 Fig) is sufficient to repress five transcripts involved in lipid/steroid metabolism (HMGCS1, LDLR, NSDHL, CRABP2 and LASS4) (Fig 3G). Of note, no imprinted gene was found in this analysis, indicating that the effect of the Bahd1-null mutation on this gene category is specific to the placenta. Altogether, these findings show that changes in BAHD1 levels alter expression of gene networks controlling steroid/lipid metabolism and hormone signaling pathways, which are key players in the development of the placenta and in control of energy in the body.
Efforts to perform chromatin immunoprecipitation (ChIP) of BAHD1 in murine or human cells using custom [12] or commercial BAHD1 antibodies repeatedly failed, preventing straightforward identification of genomic loci targeted by BAHD1 (BAHD1 ChIP is further discussed in S1 Text). To get further clues on BAHD1 function, we searched for novel BAHD1-associated proteins in cells stably expressing His6-Protein-C-tagged-BAHD1 (HPT-BAHD1) by using tandem affinity chromatography purification (TAP) and Mass spectrometry (MS), as described previously [11]. We found that ~10 proteins reproducibly co-purified with BAHD1 after two successive purification steps (Fig 4A and S6 Table): the mesoderm induction early response (MIER) proteins (MIER1, MIER2 and MIER3), HDAC1/2, HP1γ/β, KAP1, CDYL1, KAP1, PPP2R1A, RUVBL2 and DDX17. Other polypeptides were consistently detected in the first-step purification (HP1α, RUVBL1, CDYL2, DDX21, G9a, CHD3, S6 Table). Peptides matching with MBD1 were found in one TAP, in agreement with yeast two-hybrid, co-immunoprecipitation and GST pulldown assays that previously showed that BAHD1 interacts with MBD1 and HP1 [12]. Western blot analysis confirmed that BAHD1 co-purifies with MBD1 and HP1γ together with MIER1, HDAC1, HDAC2, KAP1 and the H3K9 KMT G9a (Fig 4B and S4 Fig). These data suggest that BAHD1 and MIER proteins are subunits of a macromolecular complex containing chromatin writers (e.g. G9a), readers (e.g. HP1α/β/γ, CDYL1/2, MBD1), erasers (e.g. HDAC1/2) and remodelers (e.g. CHD3, RUVBL1/2).
We noticed that BAHD1 and MIER1 displayed conserved domains found in scaffolding proteins of other HDAC1/2 complexes: the metastasis-associated protein (MTA) subunits of NuRD [8, 25] and RERE/atrophin-2, a transcriptional repressor belonging to a poorly characterized HDAC1/2 complex [26–28] (Fig 4C). BAHD1, MTA and RERE share a Bromo-Adjacent Homology (BAH) domain, which is known to promote protein-protein interactions and binding to nucleosomes [25, 29, 30], while MIER1, MTA and RERE proteins contain juxtaposed ELM2 and SANT domains that recruit HDACs [25]. Thus, BAHD1 and MIER1 could cooperate to fulfill a function similar to that of MTA and RERE in other HDAC1/2 complexes (Fig 4D). To investigate this hypothesis, we performed a series of co-immunoprecipitation experiments on HEK293-FT cells transiently expressing tagged-versions of BAHD1 from plasmid vectors, since endogenous BAHD1 is undetectable in HEK293 cells. Reciprocal co-IP assays with nuclear extracts confirmed association of endogenous MIER1 with V5-tagged BAHD1 in HEK293-FT cells and their co-immunoprecipitation with HDAC2 and HP1γ (Fig 4E and S5 Fig). Likewise, BAHD1 tagged with the fluorescent protein citrine (YFPc-BAHD1) pulled-down HDAC2, MIER1 and HP1γ (Fig 4F, lane 1).
BAHD1 harbors a N-terminal proline-rich region with the highest density of prolines found between residues 239 to 361, a region termed the cPRR [31], and a C-terminal BAH domain that interacts with the N-terminal tail of histone H3 [12] (Fig 4C). To assess the requirement of specific protein domains in BAHD1 for productive interaction with MIER and HDACs, we expressed truncated forms of YFPc-BAHD1 in HEK293-FT cells. As shown in Fig 4F and S5E Fig, the deletion mutant YFPc-BAHD1-ΔcPRR lacking residues 239 to 361 co-immunoprecipitated with HDAC2, MIER1 and HP1γ, as the full-length YFPc-BAHD1. In contrast, the deletion mutant YFPc-BAHD1-ΔBAH lacking residues 592 to 780 encompassing the BAH624-780 domain failed to pull-down HDAC2 and MIER1, while retaining the ability to bind HP1γ. We conclude that BAHD1 interaction with MIER1 and HDAC2 requires the integrity of the BAH domain of BAHD1.
We also found that varying BAHD1 amounts has an impact on MIER1 subcellular localization. First, MIER1 shifted from cytosolic to chromatin-bound fraction in response to induction of BAHD1 expression in HPT-BAHD1 cells (Fig 5A). Second, microscopy experiments showed that BAHD1 overexpression increased MIER1 nuclear staining and induced enrichment of MIER1 at the heterochromatic inactive X chromosome (Xi) (Fig 5B and 5C). This effect was not due to changes in MIER1 transcript levels (S3 Fig). The fact that BAHD1 levels dictate the localization of MIER1 strongly suggests that BAHD1 and MIER1 form a stable chromatin-bound complex. This hypothesis is also supported by results from the Heard laboratory [15] showing that BAHD1, MIER1/2, HDAC1/2 and G9a co-immunoprecipitate with CDYL, a protein recruited to Xi in mouse ES cells and also found in our TAP assays (S6 Table).
We propose that BAHD1 and MIER are co-repressors that, like MTA subunits of NurD complexes [9], exert their function through association with sequence-specific DNA binding transcription factors (TFs). With the hypothesis that such TFs should regulate the same gene networks as BAHD1, we used the Ingenuity Pathway Analysis Upstream Regulator software to predict TFs responsible for differential gene expression in BAHD1-deficient or -overexpressing cells. This analysis identified seven TFs that were consistently found as upstream regulators of a set of BAHD1-associated genes in all transcriptome datasets: ESR1, ESR2, EPAS1 (HIF2α), PPARG, FOS, TP53 and SP1 (S7 Table), the latter being previously reported to associate with BAHD1 [12] and MIER1 [32]. It is striking that these TFs have relevance both in placental development and regulation of genes involved in lipid/steroid metabolism ([33–36] and other references in S1 Text).
Since ESR1 has been shown to bind MIER1 [37] and to cooperate with SP1 [38], we hypothesized that BAHD1-MIER could act as co-repressors for ESR1-mediated transcriptional regulation. Based on this assumption, BAHD1 should target the ESR1 gene itself, because ESR1 autoregulates its own expression [39]. Accordingly, ESR1 was up-regulated in Bahd1-KO placentas (Fig 3C), as well as in HEK293-FT cells depleted of BAHD1 with siRNA (Fig 6A). siRNA-mediated knockdown of individual MIER genes did not significantly change expression of ESR1 in HEK293-FT cells, but combined knockdown of MIER1 and MIER3 (MIER2 was undetectable) increased ESR1 expression (Fig 6A). We also found that expression of PGR (encoding the progesterone receptor), which has an ESR1 binding site in its proximal region [40] significantly increased in BAHD1- and MIER1/3-knockdown cells (Fig 6A), whereas the androgen receptor transcript (AR) was unaffected. These results indicate that BAHD1 and MIER1/3 act as repressors for the steroid hormone receptor genes ESR1 and PGR.
Human ESR1 forms a large and complex genetic unit that spans approximately 300 kb of chromosome 6 (chr6), of which 140 kb containing 8 protein-coding exons [41]. It is transcribed from at least seven promoters into multiple transcripts that vary in their 5’-UTRs and whose expression is tissue-specific. To find out whether the BAHD1-associated complexes targets ESR1, we analyzed the DNA captured with BAHD1 and its associated partners in the TAP assays by using real-time PCR with primers targeting the proximal region of ESR1 at, or in the vicinity of ESR1-binding sites identified by ChIP-seq (Fig 6B, [40], ENCODE ChIP-seq data). Regions in C6orf211 (Fig 6B) and GAPDH loci were used as controls. Three sites upstream of ESR1 were identified as BAHD1 binding sites, when compared to input or control cell DNA (E2, B2, B1; Fig 6C). To determine whether any change in the chromatin structure occur as a consequence BAHD1 binding at these sites, we tested the effect of BAHD1 depletion on histone H3 modifications at lysine 9 (H3K9) by using a ChIP-qPCR assay. We observed a reproducible increase in histone H3K9 acetylation and decrease in H3K9 dimethylation (H3K9me2) and trimethylation (H3K9me3) in cells with siRNA−induced knockdown of BAHD1 expression, relative to cells treated with control siRNA (Fig 6D and S6 Fig). This effect occurred at all sites of the ESR1 proximal region surveyed by PCR, but not at the GAPDH control site. This result strongly suggests that a BAHD1-associated complex containing HDACs and KMTs contributes to the epigenetic silencing of ESR1.
H3K9 methylation is often linked to DNA methylation and several BAHD1-associated partners are known to interact with DNA methyltransferases [42–45]. To test whether BAHD1 level also affects DNA methylation patterns at ESR1, we exploited the DNA methylome datasets of HEK-CT and HEK-BAHD1 cells that we recently obtained by whole-genome bisulfite sequencing (BS-seq) [14]. In this study, we found that BAHD1 overexpression in HEK293 cells stimulates de novo DNA methylation in autosomes at ~80,000 regions that become hypermethylated when compared to control cells. These “BAHD1-DMRs” group into large (0.3–6.5 Mb) chromosomal domains, termed BAHD1-Associated Domains (BADs, [14]). We observed that one such BAD was present on chr6 in a large region containing the ESR1 locus (Fig 6E). The high density of BAHD1-DMRs located at the ESR1 proximal and gene body regions suggests that BAHD1 overexpression stimulates large-scale epigenetic changes. Taken together, these results provide evidence that BAHD1-associated complexes induce histone and DNA modifications that shape repressive chromatin structures.
The functions of BAHD1-mediated chromatin reorganization in developmental and physiological processes were hitherto unknown. Deletion of Bahd1 in mice does not lead to embryonic lethality or anatomical malformations in fetuses. However, murine Bahd1-KO placentas are small and display histomorphological alterations, including reduced surface area of the labyrinthine zone and a thinner junctional zone and decidua. The lack of trophoblast glycogen cells is particularly striking, suggesting a failure in the differentiation program. The growth restriction of Bahd1-KO fetuses is likely to be secondary to defective placental nutrient exchange resulting from these placental abnormalities. In addition, as there is substantial deregulation of metabolic genes in Bahd1-KO MEFs, the perinatal death of Bahd1-KO pups likely results from metabolic defects in embryos. The rare mice surviving beyond birth are males with a lower weight than controls, associated with decreased fat mass and lower levels of circulating cholesterol, glucose and leptin. In light of our finding that the transcriptional co-repressor MIER1 is a key partner of BAHD1, it is striking that Mier1 null mice also show decreased body weight and reduced levels of circulating glucose and cholesterol (data from The Mouse Phenotyping Consortium and The Wellcome Trust Sanger Institute, references in S1 Text). This phenocopy is a strong argument in favor of BAHD1 and MIER1 cooperating to control the expression of genes involved in energy metabolism in somatic tissues.
Separate studies have shown that MIER1 binds to HDAC1/2, G9a and CDYL [27, 28, 46], while BAHD1 interacts with proteins involved in heterochromatin formation (e.g. HP1, MBD1) and its overexpression is sufficient to compact chromatin [12]. We now bring evidence that BAHD1 and MIER act in partnership within a novel histone deacetylase complex involved in gene silencing. First, BAHD1 co-purifies with MIER, HDAC1/2, G9a and CDYL in very different cellular models: human HEK293 cells, as shown here, and mouse ES cells [15]. Second, as in other subunits of HDAC1/2-associated complexes, MTAs of NuRD [8–10, 25](Fig 4D) and RERE [26–28] (Fig 4C), a BAH domain is present in BAHD1 and HDAC-interacting ELM2-SANT domains are present in MIER [27, 46]. We found that a truncation of the BAH domain in BAHD1 abrogates BAHD1 coimmunoprecipitation with MIER1 and HDAC2. From this, we propose that MTA, RERE and BAHD1-MIER define related macromolecular complexes of the “NuRD superfamily” and to name “BAHD1 complexes” those containing a BAHD1 subunit, as BAHD1 has no homolog or isoform. MIER1, MIER2 and MIER3 could be incorporated into distinct BAHD1 complexes leading to functional redundancy and/or context-dependent functions, as described for MTA1, MTA2 and MTA3. MIER1 is the most abundant partner of BAHD1 in HEK293 cells and its nuclear localization, including at the Xi, depends on BAHD1 expression levels. This translocation is likely to have functional consequences, as MIER1 nuclear/cytoplasmic distribution varies with cell type and stage of differentiation [47, 48]. Bahd1- and Mier1-KO adult mice share common phenotypes but a prenatal growth restriction is not reported for Mier1-KO; hence, the BAHD1-associated regulatory gene network in embryonic tissues likely involves other partners than MIER1.
Genes altered in Bahd1-KO placentas or embryonic fibroblasts are mainly up-regulated compared to wild type controls, consistent with BAHD1 acting as a repressor. The most significant biological process associated with these genes, and with those repressed in BAHD1-overexpressing human HEK293 cells, is steroid metabolism. This functional convergence is striking and consistent with the observed hypocholesterolemia and hypolipidaemia in Bahd1-KO mice. We observed that genes associated with this biological function in Bahd1-KO MEFs and HEK-BAHD1 cells (e.g. DHCR24, HMGCS1, LDLR, NSDHL, SREBF2, SC4MOL) mostly differ with those in Bahd1-KO murine placentas (e.g. Apoc3, Atp8b1, Cyp11A1, Insig2, Osbpl5, Pbx1, VldlR). Thus, like many other chromatin repressors [2], BAHD1 regulates distinct targets in different tissues and at different developmental stages. However, some of the BAHD1 bona fide target genes can be common to different biological systems, as exemplified by the steroid hormone receptor gene ESR1. BAHD1 binds the proximal region of human ESR1 and represses ESR1 both in human cells and murine placentas. In addition, changes in BAHD1 expression levels in HEK293 cells alter the patterns of H3K9 modifications and DNA methylation at the ESR1 locus, which is consistent with the interaction of BAHD1 with HDACs, KMTs, HP1 and MBD1 and the crosstalk between histone and DNA modifications [49].
It is worth mentioning that MTA1 [50] and SIN3a [39] have also been shown to represses ESR1 in other cell types, indicating that different HDAC complexes control the epigenetic silencing of ESR1. Recruitment of BAHD1 at specific sites of the genome might rely on the combinatorial assembly of BAHD1-MIER subunits with transcription factors. The product of ESR1 itself (ESR1/ERα) could be one such TF targeted by BAHD1 complexes because ESR1 autoregulates its own expression [39] and MIER1 has been shown to bind ESR1 [37]. In line with this hypothesis, ESR1 is predicted to control several genes differentially expressed in BAHD1-KO tissues, such as PGR. In addition, ESR1 cooperates with SP1 [38], a TF known to bind BAHD1 and MIER1 [12, 32], and which is also predicted to drive BAHD1-associated transcriptional changes. Our analysis also suggests that BAHD1 could act as a co-repressor with PPARγ and HIF2α, which like ESR1 are known to play roles in the regulation of energy metabolism and placental cell differentiation [33–36]. This raises the plausible hypothesis that the BAHD1-associated chromatin complex could act as a transcriptional co-repressor in synergy with different TFs in the context of placental functions and lipid metabolism.
Placenta morphogenesis depends on the correct balance of cytotrophoblast proliferation and differentiation, into either syncytiotrophoblast involved in nutrient/gas exchange or invasive extravillous trophoblast involved in establishment of blood flow to the placenta. It is proposed that ESR1 controls the proliferation of estrogen-dependent cells, while ESR2 controls their maturation, hence trophoblast differentiation is associated with the transition from Esr1 to Esr2 expression [51, 52]. By maintaining unbalanced expression of Esr1 at an inappropriate time of the gestation, BAHD1 deficiency could disturb trophoblast differentiation. Pgr and Lepr genes are also deregulated in Bahd1-KO placentas; this should affect progesterone and leptin signaling, also important in placental development [53, 54]. Therefore, BAHD1 could play a role in trophoblast differentiation, in particular in the formation of glycogen-producing cells, by controlling hormone signaling in a timely manner, particularly in the junctional zone, an important endocrine region. Knockdown of Bahd1 in the mouse placenta results in up-regulation of several other genes that also have relevance to placental development. Among them, Htra3 and Tfpi2 are two confirmed imprinted genes expressed from the maternally inherited allele [24]. Maternally expressed genes have been proposed to limit maternal resource provision; thus up-regulation of such genes is consistent with a placental growth restriction observed upon BAHD1 deficiency. In fact, HtrA3 and Tfpi2 are highly transcribed in the placenta and involved in the regulation of endothelial function, trophoblast migration and invasion [55–58]. Further studies will be required to explore whether BAHD1 directly targets these genes and contributes to their imprinting.
With the importance of the placenta in the feto-maternal exchange processes, as well of steroid signaling in the body, alterations in the amounts or activity of BAHD1 may lead to various pathological processes. We previously identified a role for BAHD1 in infection of epithelial cells with the bacterial pathogen Listeria monocytogenes [11, 31]. The fact that BAHD1 is involved in placental function now opens the possibility that BAHD1 contributes to the fetoplacental step of listeriosis, as L. monocytogenes has a tropism for the placenta. More generally, a connection between BAHD1 and placenta-associated pathologies, such as Intrauterine Growth Restriction and Pre-Eclampsia, should be carefully examined. BAHD1-mediated regulation of ESR1 is also enticing as ESR1 is a key regulator in a variety of biological processes and ESR1 deregulation has been implicated in several diseases, including breast cancer [59]. A shift from nuclear to cytoplasmic localization of MIER1 during breast cancer progression has been observed, suggesting that nuclear MIER1 contributes to the repression of genes involved in invasive breast carcinoma [37]. Similar to chromatin-repressive complexes Polycomb and NuRD [9, 60], BAHD1 could be implicated in the regulation of transcriptional events involved in diverse oncogenic pathways. Of note, insertion events in BAHD1 and MIER1 genes were identified in a screen for genes that cooperate with oncogenic KRAS (G12D) to accelerate tumorigenesis and promote progression in a mouse model of pancreatic ductal preneoplasia [61]. Our data also implicate BAHD1 in mammalian metabolic regulation and several BAHD1 candidate target genes have been associated with metabolic diseases in humans. For instance, genetic polymorphisms in CRABP2 are associated with changes in plasma cholesterol levels [62], SNPs in LPL and LASS4 are associated with dyslipidemia [63–65], changes in expression levels of HMGCS1, LDLR and SC4MOL correlated with obesity-related type 2 diabetes and cardiovascular diseases [66]. Together, these results indicate that dysfunction of BAHD1 complexes could promote aberrant epigenetic phenomena at the origin of different disorders. Detailed understanding of how and where BAHD1 complexes establish repressive chromatin states could be instrumental for the development of new strategies for selective treatment of metabolic disorders in the future.
Mice were bred and maintained in the animal facilities of the Institut Clinique de la Souris (ICS, Illkirch, France) under pathogen-free conditions. The ICS facilities are licensed by the French Ministry of Agriculture (agreement #A67-218-37). All animal procedures were approved by the local ethical committee CREMEAS (registered under the reference “C2EA– 35”), and were supervised by M.F.C. and O.W. who are qualified in compliance with the European Community guidelines for laboratory animal care and use (2010/63/UE Directive).
Human cell lines derive from HEK293 cells (ATCC CRL-1573): the HPT-BAHD1 inducible line and its isogenic HPT-control are described in [11]; the HEK-BAHD1 constitutive line and its isogenic HEK-CT control are described in [14]; HEK293-FT are from Invitrogen (ThermoFisher Scientific). Plasmid pcV5-BAHD1 (BUG2289), pYFPc (also named pEYFP-Citrine-N1, BUG2897), pYFPc-BAHD1 (BUG2897) and pYFPc-BAHD1-ΔBAH (BUG2740) are described in [12]. pYFPc-BAHD1ΔcPRR (BUG2897) is described in [31]. Antibodies were against BAHD1 (Abcam, 46573), MIER1 (Sigma, HPA019589), HDAC1 (Abcam, ab7028), HDAC2 (Abcam, ab7029), HP1γ (Euromedex, 2MOD-1G6-AS), KAP1 (Abcam, ab10483), G9a (MBL Cliniscience, D141-3), MBD1 (Abcam, ab3753), tubulin α (Santa Cruz, sc5546), H3K9me2 (Abcam ab1220), H3K9me3 (Abcam ab8898), H3K9ac (Upstate/Millipore 07–352), V5 and V5-HRP (Invitrogen R960-25, R961-25), GFP (Mouse anti-GFP Sigma/Roche 11814460001 used in IP and Rabbit anti-GFP Santa-Cruz sc-8334, used in WB, which both recognize YFPc) and control IgG mouse (Santa-Cruz sc-2025) and IgG rabbit (Santa-Cruz sc-2027). Fluorescent secondary antibodies were from Jackson ImmunoResearch or Molecular Probes, and HRP-conjugated secondary antibodies were from AbCys or Abcam (IgG Veriblot for IP, ab131368). siRNAs were purchased from Dharmacon (ThermoFisher Scientific) as follows: on-TARGETplus Non-targeting pool (D-001810-10-20), BAHD1 (L-020357-01), MIER1 (M-014201-02), MIER2 (M-023917-01), MIER3 (M-015618-01). Cells were transfected 72h with siRNA using Lipofectamine RNAimax (Life Technologies, Grand Island, NY) according to the manufacturer's instructions. Immunofluorescence and XIST FISH assays were as described in [12].
Bahd1+/- mice have been described in [11]. Mice were bred and maintained in the animal facilities of the Institut Clinique de la Souris (ICS, Illkirch, France) under pathogen-free conditions. Throughout the experiment mice were housed in the same climate-controlled stable with a 12h/12h dark-light cycle and handled identically. For wild type and Bahd1-/- production, Bahd1+/- mice were mated and the day on which a vaginal plug was found was designated 0.5. Genotyping protocols are described in [11]. Knockout of Bahd1 expression was verified by RT-qPCR on placenta, MEF and embryo liver samples. Phenotyping methods of Bahd1 heterozygous and knockout mice and isolation of primary mouse embryonic fibroblasts are described in S1 Text. All the data are expressed as mean ± SE. Statistical analysis were performed using a one way ANOVA tests followed by a Fischer’s PLSD test with significance set at P<0.05. * P<0.05, ** P<0.01, *** P<0.0001.
Total RNAs were extracted from Bahd1+/+ and Bahd1-/- placentas at E16.5 (n = 6/genotype) or E18.5 (n = 3/genotype), as well as from Bahd1+/+ and Bahd1-/- MEFs (n = 3/genotype) and HEK-CT and HEK-BAHD1 cells (n = 3/cell line) using RNeasy Kit (Qiagen), treated TURBO DNA-freeTM kit (Ambion). RNA concentration and integrity were tested with RNA quality was monitored on Agilent RNA Pico LabChips (Agilent Technologies, Palo Alto, CA). 100 ng of RNA per sample were used as templates for the synthesis of hybridization probes for Affymetrix GeneChip Microarrays (Genechip HuGene 1.0 ST for HEK-CT and HEK-BAHD1 cells; Mouse gene 1.0 for E16.5 placentas; Mouse Exon 1.0 ST for E18.5 placentas and MEFs). Hybridization was carried out with biological replicates according to the expression analysis technical manual with wash and stain kit (Affymetrix). Gene-level expression values were derived from the CEL file probe-level hybridization intensities using the model-based Robust Multichip Average algorithm (RMA) [67]. RMA performs normalization, background correction and data summarization. An analysis is performed using the LPE test [68] and a p-value threshold of p<0.05 is used as the criterion for expression. The estimated false discovery rate (FDR) of this analyze was calculated using the Benjamini Hochberg approach in order to correct for multiple comparisons. Results were annotated using information provided by Affymetrix. Full data sets were reduced by discarding genes with “EST” and “unknown” annotation labels. To generate functional clusters of genes, we used the DAVID program (http://david.abcc.ncifcrf.gov, 2015 version; [20]) for selected gene sets according to gene ontology (GO) of biological process categories. To search for transcriptional regulators driving the differential gene expression changes we used the Ingenuity Pathway Analysis Upstream Regulator software [69]. Datasets have been deposited in Gene Expression Omnibus (GEO) and are accessible through accession number, as follows: GEO series GSE51868 (transcriptomes of HEK-CT and HEK-BAHD1); GSE53443 and GSE53442 (transcriptome of Bahd1-WT and Bahd1-KO placentas at E16.5 and at E18.5) and GSE73816 (transcriptome of Bahd1-WT and Bahd1-KO MEFs).
Total RNA from HEK293-CT and HEK-BAHD1 cells, HEK293-FT cells treated 72h with siRNAs, or from mouse placentas or MEFs was extracted using the RNeasy Kit (Qiagen), from three to six biological replicates. Genomic DNA was removed by treatment with TURBO DNA-freeTM kit (Ambion). cDNAs were generated from 1 to 2 μg total RNA using the RT2-HT first strand kit (Qiagen/SABiosciences). Quantitative Real-Time PCR was performed on Biorad MyiQ device (Biorad), using SsoFast Evagreen supermix (Biorad), as specified by the supplier. Each reaction was performed in triplicate. Data were analyzed by the ΔΔCt method. Target gene expression data were normalized to the relative expression of human GAPDH or mouse Gapdh (and Hprt, for imprinted genes) and YWHAZ was used as a control gene. Statistical significance of the difference in mean expression of genes was evaluated using the Student t test; a P value <0.05 was considered significant. Primer sets are provided in S1 Text.
The TAP-TAG protocol to purify the partners of His6-Protein-C-tagged BAHD1 in HPT-BAHD1 cell is described in [11]. Modifications to this protocol, Mass spectrometry analysis and precipitation of the associated DNA are detailed in S1 Text. Immunoprecipitation (IP) of nuclear proteins was performed as described in [70] with the following modifications. Nuclear soluble and insoluble fractions were sonicated, mixed and incubated overnight at 4°C with 1–3μg of the indicated antibodies and then with Dynabeads Protein G (#10004D, Novex) at 4°C during 2h30. IPs were washed 4 times with washing buffer (Tris 20mM pH = 7,65, NaCl 150mM, 0,05% IGEPAL, 2,5% Glycerol, 0,5mM EDTA, 0,6mM DTT). Samples were resuspended in 1X Laemmli and boiled at 95°C for 10 min. Proteins were separated on 8–10% SDS-PAGE gels, transfered to nitrocellulose membrane, probed with primary and secondary antibodies and detected by chemiluminescence (SuperSignal West Femto Substrate #34094, ThermoFisher Scientific). Blots were visualized on Films (Amersham) or using a ChemiDoc MP Imaging system (Bio-Rad). ChIP of modified H3K9 were performed in three independent biological replicates, using the ChIP-IT Express Enzymatic Kit (Active Motif) according to manufacturers instructions. Briefly, HEK293-FT cells were grown in T150 flasks for 72h with either siRNA against BAHD1 or non-targeting control siRNA (up to 70–90% confluency). One flask was kept for RNA extraction and quantification of BAHD1 knockdown by RT-qPCR. Cells from other flasks were harvested and cross-linked with a final concentration of 1% formaldehyde (Sigma-Aldrich) for 10 minutes. Fixation was stopped with 0.125 M glycine and cells were washed twice with ice-cold PBS. Collected cells were lysed and nuclei pellet were resuspended in shearing cocktail and incubated for digestion for 10 minutes at 37°C. Shearing efficiency was tested by agarose gel electrophoresis and DNA concentration was quantified with a Nanodrop 2000 (ThermoFisher Scientific) to normalize the quantity of chromatin per ChIP. For each ChIP, 3 μg of antibody were used. 10 μL of chromatin was kept as input and processed as ChIP samples. After washing and reverse crosslinking of precipitated samples, DNA was purified by two extractions with equal volumes of phenol:chloroform:isoamylalcohol (25:24:1, pH = 8), assisted by phase lock heavy gel tubes (5Prime), followed by ethanol precipitation. Pellets were washed once in 75% ethanol, then resuspended in 50 μL DNAse-free water. H3K9ac, H3K9me2 and H3K9me3 enrichment levels were measured by qPCR with primers matching in the ESR1 locus and the non-target control GAPDH region (primer sets are provided in S1 Text). A standard curve was generated using 10%, 1%, 0.1% and 0.01% of input DNA. We first determined the fold enrichment of the ChIP sample relative to the IgG sample and then the effect of BAHD1 knockdown on H3K9 modifications was calculated as the ratio of enrichment in cells treated with BAHD1 siRNA to that in cells treated with control siRNA (presented in Fig 6 and S6 Fig as a Log2 ratio).
|
10.1371/journal.pcbi.1007126 | With an eye on uncertainty: Modelling pupillary responses to environmental volatility | Living creatures must accurately infer the nature of their environments. They do this despite being confronted by stochastic and context sensitive contingencies—and so must constantly update their beliefs regarding their uncertainty about what might come next. In this work, we examine how we deal with uncertainty that evolves over time. This prospective uncertainty (or imprecision) is referred to as volatility and has previously been linked to noradrenergic signals that originate in the locus coeruleus. Using pupillary dilatation as a measure of central noradrenergic signalling, we tested the hypothesis that changes in pupil diameter reflect inferences humans make about environmental volatility. To do so, we collected pupillometry data from participants presented with a stream of numbers. We generated these numbers from a process with varying degrees of volatility. By measuring pupillary dilatation in response to these stimuli—and simulating the inferences made by an ideal Bayesian observer of the same stimuli—we demonstrate that humans update their beliefs about environmental contingencies in a Bayes optimal way. We show this by comparing general linear (convolution) models that formalised competing hypotheses about the causes of pupillary changes. We found greater evidence for models that included Bayes optimal estimates of volatility than those without. We additionally explore the interaction between different causes of pupil dilation and suggest a quantitative approach to characterising a person’s prior beliefs about volatility.
| Humans are constantly confronted with surprising events. To navigate such a world, we must understand the chances of an unexpected event occurring at any given point in time. We do this by creating a model of the world around us, in which we allow for these unexpected events to occur by holding beliefs about how volatile our environment is. In this work we explore the way in which we update our beliefs, demonstrating that this updating relies on the number of unexpected events in relation to the expected number. We do this by examining the pupil diameter, since—in controlled environments—changes in pupil diameter reflect our response to unexpected observations. Finally, we show that our methodology is appropriate for assessing the individual participant’s prior expectations about the amount of uncertainty in their environment.
| The role of the noradrenergic (NA) system in decision making [1] and encoding uncertainty [2] has been explored in great depth, with many studies using pupillary dilation as an index of changes in central adrenergic signalling [3–6]. A central theme of this work is the role of NA in contextualising perceptual inference and planning. This has an interesting connection to the P3 evoked response potential seen in EEG paradigms [7]. The amplitude of the P3b wave increases following presentation of surprising stimuli [8,9] and is thought to signal a change in beliefs about the underlying environmental contingencies [9]–i.e., the updating of context [10]–and might be mediated by NA [1]. However, these accounts of the role of NA in signalling surprise often focus on transient responses following a single unexpected event. Here, we extend this work to show that pupil dilatation tracks a subject’s long-term beliefs (that is, tonic changes to the baseline pupil diameter), spanning multiple aberrant events and how these allow the participant to infer the precision of their environmental dynamics. Precision here refers to the predictability of the next state of the world, given the current state.
Formally, we appeal to the notion of a generative model. This is central to theoretical treatments of the Bayesian brain and, more generally, active inference. These accounts frame brain function as a process of inference that depends upon an internal generative (predictive) model comprising prior beliefs about variables in the world, and (likelihood) beliefs about how these give rise to sensory data. On observing sensory data, creatures can use their internal model to update posterior beliefs about their environments. Posterior beliefs can then be used to compute empirical prior beliefs about the future (i.e., planning as inference), using temporal contingencies in the generative model. Our focus here is how the brain handles uncertainty about these contingencies.
The encoding of uncertainty is essential in enabling animals to predict confidently (or not) what might happen next [11]. From the perspective of the Bayesian brain, this is the process of using beliefs about the past to form (empirical) prior beliefs about the present. Crucially, the confidence (precision) in these priors determines the relative weighting of prior and sensory influences on perceptual or state inference. This has relevance for the role of abnormal prior beliefs in pathology, where under-confident priors fail to contextualise inferences drawn from sensory data or where excessively confident priors support false inferences in the presence of contradictory sensory data [12]. Of particular relevance to this work are those conditions that have been associated with abnormal NA signalling—for example, autism [13,14].
We start by introducing a few technical concepts. Following this, we describe our experimental design and data collection. With these data, we test the hypothesis that the pupil diameter closely tracks the precision inferred by the participant in a volatile setting. We build on this formulation to propose a technique that quantifies prior beliefs, which could be used in a clinical context to phenotype individuals, in terms of their prior beliefs about precision and volatility (e.g., that might underwrite autistic symptoms).
Adaptive engagement with the world requires an understanding of our sensations in terms of the latent (hidden or unobservable) states that generated them. This requires an internal (generative) model of the world that can be used to make predictions about sensory input [15]. These generative models are necessarily complicated (i.e., usually deep, dynamical and nonlinear), to capture the subtleties of our (deep, dynamical and nonlinear) environment. Despite the complexity of such models, they can be constructed by combining relatively simple models [16]. The simplest that accounts for perceptual inference and planning—in a changing environment—is a Markov decision process (MDP) [17]. Technically, in this paper we use a hidden Markov model (as we do not model any decisions), but we retain the MDP rhetoric to emphasise that these results generalise to situations that require active sensing of the world [17,18]. In the following section we provide a brief outline of the inversion of this type of generative model. Readers familiar with this sort of modelling are invited to skip this section.
An MDP treats the world as comprising a series of states (s) that are hidden from an observer. The transitions among these states over time represent the (stochastic) dynamics of the environment, and are defined by a (square) transition matrix that we denote by B (see Fig 1 for a Bayes net representation of this process). These states give rise to observable outcomes that act as the observer’s sensory stimuli. The relationship between the hidden states and the outcomes they generate is expressed as a likelihood matrix, A. These probability distributions are not trivial: to motivate their importance we appeal to the Good Regulator Theorem [19]. This theorem says that ‘every good regulator of a system must be a model of that system.’ From this, one may intuit that if a creature inhabits, and wishes to interact with, an environment defined by stochastic state transitions, this creature must be able to estimate the precision (i.e., the negative entropy) of these transition densities. The ensuing inference about environmental dynamics is intertwined with beliefs regarding the likelihood mapping from states to outcomes, since it is only these outcomes that an agent can observe [17]. If the dynamics of the environment are deterministic, and state-to-outcome mappings are well understood, the agent is likely to have precise beliefs about the nature of its environment, and is therefore able to accurately predict what it may expect to see in the future [20]. However, when these mappings from states to outcomes are not deterministic and where state transitions are themselves stochastic—the agent is presented with a confound, since poor inferences about the nature of state-to-outcome mappings may have a detrimental effect on inferences about state transitions [20,21].
We represent these imprecise state-to-outcome mappings and state-to-state transitions in the A and B matrices, respectively (Fig 2). Imprecise state transitions (a non-deterministic B matrix) define a volatile environment. In other words, volatility is equivalent to the inverse precision of the B matrix. In a volatile world, even if an animal accurately infers the likelihood mappings and state transitions, the stochastic nature of these dynamics means the agent’s beliefs about what will happen next are necessarily imprecise [22]. Imprecise beliefs over state transitions leave the animal with no way of predicting what might come next. This has been referred to as ‘unexpected uncertainty’, in contrast to ‘expected uncertainty’ (that maps to imprecision of A) [2,6].
Previous work has suggested that acetylcholine (ACh) and noradrenaline (NA) act as the neurochemical analogues of the precision over state-outcome mappings (i.e. the A matrix) and the precision over state transitions (defining the empirical prior and B matrix) [2] respectively. In other words, ACh is thought to be involved in signalling confidence in our beliefs about the likelihood of what we might see, in a given state, while NA moderates our confidence in prior beliefs about the state we may find ourselves in next [2,3,5,23,24]. In this work, we use these proposed relationships—between central noradrenergic signalling and volatility signalling—to motivate predictions about changes in pupil diameter based on environmental volatility. Given that circuits in the brain—that update beliefs over the likelihood mappings from states to outcomes—are thought to use ACh as a transmitter [2,23,25–29], a potential confound arises: pupil diameter, particularly the tonic changes examined in this work, may also depend on beliefs about the likelihood of certain stimuli. We therefore consider the possibility that beliefs about sensory mappings may confound the effects of beliefs about state transitions on the pupillary response; acknowledging that beliefs about the likelihood might also vicariously influence beliefs about transitions. Put simply, an unexpected observation could plausibly be explained by imprecision in either the A or the B-matrix.
To optimise beliefs about environmental uncertainty (i.e., precisions), we must first specify how precisions are parameterised. To pursue this formally, we express the precisions as inverse temperature parameters, such that the precision of state transitions is given by ω. This adjusts a (source) transition B matrix by virtue of a Gibbs measure (i.e., a softmax function), as shown in Eq 1. Here, precision is an exponent on the elements of the transition matrix, which is then normalised, to produce the agent’s beliefs about state transitions [22].
Eq 1 describes how transition matrices are generated from a source matrix. This produces the transition matrices shown in Fig 3, with all 4 derived from the same ‘source’ matrix. Intuitively, a high prior precision reflects high confidence in prior beliefs, and would be represented by a large value of ω. If ω were to equal infinity, this would represent absolute confidence, and results in a purely deterministic transition matrix, as is shown in Fig 3a. Smaller values of ω (Fig 3b–3d) represent increasingly less precise beliefs, resulting in increasingly stochastic transition matrices, and a greater propensity to accommodate randomness in the environment.
The updated, normalised (denoted by the bar notation), B matrix is then used to update the expected precision given new sensory outcomes. Under ideal Bayesian observer assumptions [22] this update can be cast as a gradient ascent on variational free energy (a lower bound on log model evidence). Specifically, this scheme updates the volatility (inverse precision, β = ω−1) using the sum of prediction errors, weighted over all possible transitions, as shown in Eq 2. These prediction errors represent the difference between the observed state transition and the expected state transition, where the expected transition is calculated using the updated B matrix calculated in Eq 1. The ensuing error term is shown in Eq 3.
Eqs 2 and 3 show how the inferred volatility is inextricably linked to violations of expected transitions, as inferred by the subject [22]. Here, σ refers to the softmax function. Importantly, the non-italic β in Eq 2 represents the prior beliefs an agent has regarding the volatility of the environment (so β-1 = ω are the prior beliefs over precision). Eq 2 therefore shows how the agent’s posterior beliefs about the current environmental volatility (β) depends on their prior beliefs. The formulation in Eq 3 provides a useful intuition on belief updating for volatility or precision. It says that, for every possible current state, we compute the difference between the expected next state and the posterior beliefs about that state. These errors are weighted by the posterior probability of the current state. Larger errors then induce greater updates in beliefs about the volatility. Framing belief updating in terms of state prediction errors connects this aspect of active inference to recent work suggesting that much of noradrenergic phenomenology can be reproduced by appealing to similar error terms [30].
Eq 4 shows how we estimate posterior beliefs about the states, and the influence of the volatility on this belief updating [22]. This shows that, in a highly volatile world (low ω), the influence of beliefs about the future and past should have little influence over beliefs about the present, and we should rely to a greater extent upon sensory evidence. In contrast, in minimally volatile environments, we should depend more upon empirical priors from the past and future. When inferring state trajectories, we can use these posterior beliefs to evaluate Eq 3 and update beliefs about volatility.
There is a large literature on modelling of volatility in dynamical systems that rely upon autoregressive or Kalman filter like models. While important for cognitive sciences and psychology [31,32], these also find application in the domain of financial modelling and economics [33,34]. Some of these approaches rely upon the use of stochastic differential equations for continuous variables (or their associated density dynamics [35]), while others rely upon probability transition matrices. In the former, volatility is simply the variance of random fluctuations, while in the latter it takes the form of a temperature parameter. Common to all, is the notion that the current value of a latent variable does not deterministically predict the next value. All explicitly or implicitly appeal to the imprecision of predictions about the next state, given the current state, as a measure of volatility.
Previous work has considered the updating of precisions in continuous state space models, using a hierarchical gaussian filter [36]. In this scheme, beliefs are held at multiple hierarchical levels, with belief updating driven by prediction errors. The precisions at each level are dynamic, and encode the uncertainty (or the volatility) about fluctuating continuous states of the environment [36]. Other approaches have considered a delta-rule style belief updating, which has been combined with Bayesian approaches to form a Bayesian delta rule [37,38]. These formulations have previously been used to examine the relationship between noradrenergic signalling and the estimation of volatility in both a neurotypical setting, with and without reward, [6,38] and in the case of patients with autism [14]. Indeed, optimising beliefs about the uncertainty of state transitions is an essential feature of cognitive flexibility, allowing us to anticipate changes in task contingencies. This means we can assess the relevance of recent events in predicting what might come next. This regulation of beliefs is synonymous with the learning rate in reinforcement learning [6,24,38,39].
In this work we focus on the uncertainty about the environmental contingencies. By formalising the Bayesian updating thought to occur in the brain [6,22,40], we can quantify the prior precision (i.e., confidence) participants afford their beliefs about environmental volatility by examining the effect on the belief-updating when presented with a unpredictable outcomes [22]. Crucially, our model makes predictions about the online encoding of uncertainty and accompanying pupillometric responses. This allows us to examine the tonic responses of the pupil without using summary statistics, as in previous work [6,41]: usually, trial-to-trial fluctuations in the pupil diameter are measured by taking the average dilation or the change relative to baseline. In this work, we generalise this examination of the trial-to-trial fluctuations in pupil diameter by explicitly parametrising it as a function of inferred precision [22]. This allows us to quantify a participant’s prior precisions over environmental volatility based upon observable (pupillometry) responses—a capability that holds promise for applications in theoretical, computational and clinical neuroscience [14,22,42].
To test the hypothesis that pupillary responses are, in part, mediated by the encoding of uncertainty, we assessed the evidence for alternative models of pupillary dilatation afforded by pupillometry data from 9 healthy participants. We generated plausible models to account for pupillary responses, considering optical factors, our formulation of the inferred environmental volatility, and possible interactions between these factors (i.e. how beliefs about environmental volatility moderate the pupillary response to luminescence). In this section, we explain the design and rationale of the stimuli, the pupillometry data collection protocol, stimulus generation and model specification.
The study was approved by the UCL Research Ethics Committee (Project ID Number 4356/002). Both oral and written informed consent was obtained from all participants.
We recruited 9 participants between the ages of 18–35 with no reported psychiatric history and neurotypical development. All participants’ data are included in the analysis, and all participants completed the 16 blocks. Each block lasted for just over 2 minutes; allowing for short breaks between blocks (to avoid discomfort). Each session lasted for around 1 hour. Participants rested their head in a chinrest 0.5m away from a stimulus presentation screen. Dark numbers on a grey background were used to reduce the effect of pupillary dilatation in response to changes in illumination [37], and the lights were dimmed (high illumination leads to a constricted pupil and restricted dilatation). Quiet was maintained during each experiment to avoid dilation in response to auditory cues [43], which would effect the level of surprise [6]. Pupil area was monitored in the left eye using an EyeLink 1000 desktop mount (SR Research, sampling rate: 1000Hz), with calibration performed before the first experiment, as well as after any periods where the participant moved from the chin rest.
While fluctuations in pupil diameter can be attributed to a luminance effect, due to the presentation of numbers on the screen, we note that this pupil dilation could alternatively (and perhaps also) be due to increased attention [44]. This effect has been reported during auditory paradigms: a recent study showed that the conscious processing of regularities in an auditory paradigm induces a pupillary dilation [4,45,46].
To measure the pupillary response to changes in environmental volatility, we presented sequences of numbers increasing from 1 to 8. These were generated using a probability transition matrix (details below), analogous to that used in the generative model (see Fig 2). We chose numbers because prior exposure to number sequences means that participants are ‘over-trained’, and do not need to learn new sequences. To increase the volatility of this stimulus, we decreased the precision of the transition matrix used to generate the sequence. This introduced violations of the 1–8 sequence. Every sequence, irrespective of the number of violations, comprised 8 numbers. Each number was presented for 250ms, followed by a 250ms with no stimulus. In other words, each number was always followed by the absence of a number, such that each of these pairs lasted for 500ms. After 8 pairs, there was a delay for 1 second. The small 250ms spaces between numbers were required to make transitions distinct (for example, without these interludes a transition from a 2 back to a 2 would simply look like an extended presentation of the number 2). The 1 second breaks between sequences helped prevent transitions such as 8 to 1 as one sequence ends and another began immediately after—as this could equally well be seen as a an unexpected transition. Numbers were presented in a dark font on a grey screen; a truncated example of a sequence is shown in Fig 4e.
To account for the 250ms number absences between the numbers, and a final 1s break between states (comprising 4 consecutive 250ms number absences), we required a 20x20 probability transition matrix. We ensured that all numbers invariably transitioned to a number absence state, but number absence states could stochastically transition back to the previous number, or to the next number. Specifically, we specified the probability of the ‘correct’ transition (i.e. to the next number numerically) as 0.99, and then split the remaining probability mass among the remaining possible state transitions. We then selected 4 levels of volatility that would give, on average, 0, 1, 2 or 3 aberrant transitions within a sequence of 8 numbers. These volatilities were used to create the matrices shown in Fig 3a–3d. To generate the sequences shown in Fig 4, we iterated using the respective matrix 20 times, and—if the resultant sequence was suitable (i.e. composed of 8 numbers, to ensure all sequences are the same length)–the sequence was accepted; otherwise we generated a new sequence.
A single block comprised 25 sequences in total and each participant performed 16 blocks. The first of these sequences was simply 1–8. The remaining 24 sequences were divided into 8 sets of 3 sequences. Each set was assigned a level of precision, and the appropriate B matrix was used to generate each of the 3 sequences. The precisions of the sequences throughout the experiment, and the sequences themselves, are shown in Fig 4. These sequences and their ordering were kept constant throughout all 16 blocks. To ensure participants maintained focus, we asked them to perform an incidental task: they were asked to tap on a tap-counter every time they saw a specific number (this number was different in each block). They were explicitly asked not to count how many times they saw the target number, but rather to focus on the next number (which they were told should always be 1 greater than the number they just saw).
Above, we noted the possibility of an interaction between beliefs about likelihood mappings and state transitions. In the context of these sequences, we looked for this interaction by creating combination sequences for 8 of the 16 blocks that participants completed. In these sequences, we took the basic sequences given in Fig 4, randomly selected 2 numbers and switched them for different numbers. Examples of these switches and the resultant sequences are given in Fig 4.
In summary, each participant completed 16 blocks. 8 are composed of the basic sequences detailed in Fig 4, while 8 are combination sequences, constructed in the manner shown in Fig 4. Importantly, in each set of 8 blocks, the sequence of numbers was exactly the same. In a post-hoc debrief, participants were asked what they noticed about the sequences. None reported that the sequences were the same (either within the sets of 8 blocks or between the sets of 8 blocks), and all commented that they simply ignored the random numbers in the combination sequences, which has a heuristic similarity to the results shown by Parr and Friston (random stimuli that have no informative value are down sampled/ignored) [47]. Following pre-processing (detailed in the next section), we compared the time series generated from the basic sequences to those generated by the combination sequences. We selected random sections of the time series and used their mean and variance to look for a statistically significant difference in the time series and found none. We ran all analyses (see sections below) on both data sets separately, yielding similar results (though with somewhat larger error due to increased noise from reducing the sample size—see 9a insert for an example of these results). We were therefore able to pool the blocks, and do not make any further distinction between the basic and combination sequences.
Following acquisition, all data were processed with the same protocols, which are well established in the literature. First, blinks were removed by identifying data for which the pupil diameter is 0. These time points are padded by 150ms either side, removed, then replaced by linear interpolation [6,48]. We then regressed out the effect of a temporal drift, the presence of a violation (of both types in the combination series) and the presence of a target number (those requested for the counting task) [6,49]. The data were then mean centered, low-pass filtered below 10Hz, and down-sampled to 10Hz. Since the analysis we performed later was in the time domain, we had no need to respect the Nyquist frequency. Down sampling to 10Hz from 1000Hz was required since the predictions we generated (see next section) were generated at 4Hz.
Finally, the data were normalised by their standard deviation, such that the final time series represents the number of standard deviations from the mean diameter. This ensured that we could average the data over subjects, while allowing for the fact that some participants’ responses may have overall smaller pupillary responses due to differential sensitivity to the luminance of the screen. At this point, the data from all 16 blocks for a given participant were averaged together; such that our data-space now comprises 9 time-series (one for each participant) performing exactly the same tasks. This allowed us to construct an ‘event related average’, analogous to the approach used to find evoked responses in EEG research [50]. The grand mean of this average, over subjects, is shown in Fig 5.
To test our hypothesis (that the central adrenergic system mediates Bayes optimal updating of beliefs over volatility), we simulated belief updating in response to our stimuli. The simulations were performed by iterating Eqs 1, 2, 3 and 4 in a Matlab script customised from spm_MDP_VB_X.m (details in supporting materials). This scheme inverts a generative model based on an MDP to provide free energy minimizing solutions to the underlying active inference problem (that entails the solution to Eq 2). Inference about precision is assumed to proceed over a longer time-scale than state inference (inference about the precision requires beliefs about the current state, which must be inferred from the observed outcomes over a number of time-steps). From Eq 2 it is clear that the inferred volatility is dependent on the prior beliefs over precision (β-1). We therefore generated simulations of ω with a range of β-1 from 0.3–20. The inverse of the ensuing precision is then taken to be the inferred volatility. Different prior precisions have a profound effect on the shape and scale of belief updating, as can be seen in Fig 6. These simulations are generated at a 4Hz frequency (notice this is the frequency of the stimulus), and we therefore need to up-sample this to 10Hz (by linear interpolation) for comparison with our empirical data. While it may appear as if the time-course in Figs 5 and 6 depends only upon whether numerical sequences are violated, it is actually a little subtler than this. The nature of the response is highly dependent on the prior precision of the participant and the participant’s current inferences about precision. Furthermore, in the short periods of differing volatility that our experiment affords, participants with particularly high beliefs over the environmental volatility are less likely to track these small changes, since they are already expected, whereas participants with strong beliefs over the precision of the environment experience a greater prediction error and will track these changes. Notably, even as the pupil responds to individual aberrations that allow updating of beliefs about precision, the tonic state will change to reflect these beliefs.
To test for the hypothesized effects of the inferred precision on pupil diameter, we took the grand mean of the data to generate a pupillometry time series, averaged over all blocks for all participants. Noting the shape of the data (Fig 5), we wanted to consider the balance between optical effects (changes in luminance due to the numbers presenting on the screen) and the updating of inferred environmental precision. With this in mind, we generated 4 plausible models, summarised in Table 1. The explanatory variables detailed in this table accommodate the photic stimulation (effect of numbers on the screen) in each model, and then build on this to consider more comprehensive models of pupillary responses. Model 1 comprises photic stimulation only. This represents a null model. Model 2 contains the photic stimulation and an interaction effect, where the inferred precision acts in concert (i.e. non-additively) with the optical effects. Model 3 contains the photic stimulation and the inferred precision, suggesting that the inferred precision acts independently from optical effects to drive pupillary dilation. Finally, model 4 contains all three effects (optical, interaction and precision). These models are summarised in Table 1. We included a further two interesting models: Model 5 supposes that there are no tonic effects beyond a slow return to baseline following an unexpected event (note this supposes that the tonic effects are simply a due to slow dynamics of the pupil in response the phasic effects). Model 6 supposes that the participant immediately knows the current environmental volatility, rather than having to infer it from the observed data.
Taking inspiration from the field of neuroimaging, we analysed the pupillometry data using a general linear convolution model [44,49,51,52], comparing the evidence for each model using Bayesian general linear regression [53]. The prior expectation of regression parameters were set to 0 within uninformative prior variance. The results presented below were robust to changes in this prior variance. This Bayesian general linear model (GLM) allows us to balance the increase in accuracy from additional regressors in models 1–6 with the accompanying increase in complexity. We convolved our stimuli with 5 gamma functions (with associated parameters), which can be reasonably expected to model the pupillary response to our stimuli—in the spirit of a pupillary response function; i.e., the pupillary response to neuronal afference (modelled by inferred precision). While 5 gamma functions are not required to model pupillary dilation (indeed, analysis of the parameters of each gamma function suggest only the widest gamma function is necessary), we did not want to make any prior assumptions about the pupillary response function, and therefore began with a range of possible functions. We retain this full range to allow for the simulations shown in the results section.
Photic stimulation was modelled as a boxcar function encoding the presence of a stimulus, the interaction term is the mean centred product of the simulated precision and the optical effects, while the precision terms are generated as described above. These explanatory variables are then convolved with a basis set comprising (five) gamma functions, such that the design matrix for model 3 has 10 columns. We added a constant term to account for the z-scoring performed in the pre-processing. Examples of these design matrices for a β−1 of 1.75 are shown in Fig 7. Model comparisons are performed with flat priors over each model, to avoid favouring one model over another.
Treating model 1 as a null hypothesis, we test the alternate hypotheses to see if they provide a better explanation of the data. The results are shown in Fig 8, with the log model evidence relative to that of the null model (and the R2 value to quantify the model fits to the grand mean data) provided for the other models. Model evidence can be read as the probability that a given model would generate the data at hand. The relative (log) model evidence between two models indicates how much better an explanation one model (or hypothesis) is for some given data set compared to another. Crucially, this considers both the accuracy and the complexity of the model, such that larger model evidence indicates that the model either accounts for the data with greater accuracy, or is a simpler explanation with comparable accuracy. This is closely related to other approaches for comparing models, including the Akaike information criterion (AIC) and the Bayesian information criterion (BIC).
Technically, model evidence is the log likelihood (accuracy) minus a (complexity) penalty for the effective number of model parameters. The AIC and BIC may be regarded as approximations to log model evidence, while statistics such as R2 reflect accuracy. The key difference between accuracy and log evidence is that log evidence (and its free energy bound) penalises models whose parameters must be moved from their default (prior) values to explain the data (as measured by the Kullback-Leibler divergence between posterior and prior). The AIC and BIC approximate this complexity, but do not take account of whether or not these parameters are used to explain the data. In most model comparison settings, the variational free energy is a better approximation to (log) model evidence than the AIC or BIC [54].
In the current application of model comparison, we compare models for a range of prior beliefs from 0.3 to 20. Effectively, this is a line search for the optimal prior belief, relative to the null model (which has no dependence on the prior beliefs). Fig 8 shows that at very low prior beliefs, the more complicated model (model 4) is superior, but from a prior precision of β-1 > 2.25 the simpler model (model 3) is sufficient. Furthermore, the results confirm that pupillary responses are highly dependent on the β-1 (as suggested in Eq 2), and importantly that our model comparison can detect this dependency. With this in mind, we proceeded with a more delicate analysis, using model 3 (since it is superior to the other models over a larger range of precisions, and in particular around the most interesting range of precisions, see Fig 9) to examine the differences among our 9 participants.
We may perform the same analysis used above (a Bayesian linear regression with uniform priors) to optimise the model of data generated by each of the 9 participants. This allows one to identify the optimal prior precision for each participant. To do this, following the regression analysis, we can pass the log model evidence of model 3 –for different prior precisions—through a softmax function [53,55] to obtain the posterior probability over prior precision. These results are shown in Fig 9, which shows the posterior probabilities for each of the 9 participants.
In Fig 9, we also report a confusion matrix, constructed by simulating data with different prior precisions (rows), and then computing the posterior probability afforded to each level of prior precision (columns) by each simulated dataset. This is to demonstrate that—if we simulate data—we can easily recover the parameter used to generate the simulations (demonstrating the sensitivity of our measure). To illustrate the face validity of this approach, we simulated data using the parameter values inferred for 4 of the participants. The correspondence between these and the measured data are shown Fig 9c–9f.
Fig 9a shows that our participants displayed a range of β-1, with three below the average, four close to the average and two with slightly higher β-1. Importantly, we are able to identify an optimal value for all participants. This conclusion is reinforced by Fig 9b, where we show that if we generate simulated time series for 5 similar values of β-1, we can accurately assign the precisions to the correct simulation; through model comparison of models with different prior precisions.
This characterisation of model identifiability is reflected in the fact that the highest probabilities lie on the (confusion) matrix diagonal. In other words, we can recover the correct model that generated pupillometry data based on, and only on, the data themselves. To characterise the simulations that are used to find the optimal prior precision, in Fig 9c–9f we overlay the simulations on the recorded data for 4 participants, showing that the regression parameters can be used to generate plausible data. Note in Fig 9c there is a large deviation between the simulation and the recorded data around 400ms; this might explain the double peak seen in the thick blue trace in Fig 9a.
To establish the role of inferred precision (inverse volatility) in explaining pupillometry data, we considered 6 models, each of which has a unique physiological interpretation. Model 1 proposed that optical factors alone are sufficient to explain the data. Had Model 1 won over the others, this would have represented evidence against our hypothesis; namely, that pupil dilation tracks inferred volatility. However, we show that models 2–4—all of which include the inferred precision in some form—are superior to the optical model. Furthermore, the results suggest that the model with inferred precision acting directly on the pupil diameter (Model 3) is the most effective over the largest range of prior precisions; including for the prior precisions shared by most of the participants. Interactions between inferred precision and optical effects have little explanatory value for these data. This suggests that the effect of precision on pupil diameter is distinct and separable from the optical impact (within the bounds of maximum and minimum pupil diameter). Finally, we are able to estimate, in a Bayes optimal fashion, the prior precision for our participants, demonstrating the sensitivity of this estimation in relation to intersubject variability.
While the results presented here are highly complementary to those in previous work taking a Bayesian perspective on pupillary dynamics (most notably by Nassar et al, 2012 and Krishnamurthy et al 2017, 6,31), our approach offers two additional benefits. First, our focus is on inference, as opposed to learning. Intuitively, this generalises previous approaches that focus on the optimisation of parameters of a generative model (learning) to accommodate beliefs about current states of the world (inference), and their changeability. Second, we have formulated our generative model to be consistent with a Markov decision process formulation of Active Inference (16,18). The importance of this is threefold. This sort of model is equipped with a process theory that has been used to account for a range of behavioural and electrophysiological observations, affording it a face validity. In addition to this face validity, the capacity to use exactly the same model to generate pupillary responses and choice behaviour (or evoked EEG responses) provides an opportunity to test the predictive validity of our model. In future work, we hope to be able to use the estimated parameters from pupillary data for individual subjects (or groups of subjects) to predict what one might measure using (for example) electroencephalography. Finally, given established associations between other neurotransmitter systems and parameters of these generative models [2,22,56], we are now in a position to investigate the interactions between these systems (e.g. how does my uncertainty about the changeability of my environment influence my uncertainty in how I am going to act?).
While pupillary dilatation is typically associated with central noradrenergic signalling, it is notable that other neurotransmitter systems have also been correlated with these responses in both humans [57] and animals [58]. As such, the link between pupillary dilatation and the precision of transitions demonstrated here could be a manifestation of other transmitter systems in addition to (or in place of) noradrenaline, as well as different sources of noradrenergic stimulation [59–62]. For these reasons, we can only conclude that neuronal processes upstream of fibres projecting to the pupillary muscles are engaged in estimation of precision (or volatility), and that noradrenaline is the likely substrate of this. However, to implicate noradrenaline with greater confidence, it will be necessary to dissociate this from alternative transmitters. This could be through fMRI, comparing activity in the locus coeruleus, dopaminergic midbrain, and basal forebrain nuclei. Alternatively, it could be through the use of pharmacological intervention, exploring whether a central noradrenergic blockade abolishes the responses observed here. These experiments, when paired with a suitable paradigm to probe changes in likelihood mappings, could be used to further explore the theories of Yu and Dayan and others [2,56,63,64] in probing the neurotransmitter systems that underlie different forms of uncertainty.
Given recent work on the role of aberrant prior beliefs in autism and anxiety disorders, we also suggest that the techniques introduced above could be used to quantify group differences between neurotypical persons and people with autism. In practical terms, this would provide clinicians with a tool to quantitatively phenotype patients and provide a diagnostic aid for autism. Recent findings suggest that these measures may correlate with the severity of symptoms [14]. This suggests there may be utility in this type of phenotyping in quantifying the effects of therapeutic interventions. However, if this was to be used as a diagnostic tool, a change in experimental paradigm would be needed; autism spectrum disorders are often diagnosed very early in life (around 3–4 years old) [65], an age at which children are often not yet able to count. The first step to such a tool would be to use the current paradigm to examine differences between a small group of neurotypical people and patients with autism, within the age range examined in this work (18–35). If these experiments were successful in finding differences between the two sets of participants, subsequent studies could examine the efficacy of the paradigm in younger age groups, adjusting the paradigm to suit those not yet able to count—and those who find it difficult to focus on the stimulus.
Finally, we refer to the introduction and our argument that belief updating over the precision of state transitions is essential for intelligent life. While this work simply shows that humans do appear to use Bayes optimal updating for beliefs regarding volatile state transitions, it provides a solid framework from which to launch further exploration of the subtleties of this precision updating, including its interaction with belief updating for precision of likelihood mappings and for actions. With a solid theoretical and practical understanding of these concepts, the leap to a general artificial intelligence would be less of a jump, and almost a trivial consequence of (variational) optimality principles.
|
10.1371/journal.pgen.1000463 | Large-Scale Selective Sweep among Segregation Distorter Chromosomes in African Populations of Drosophila melanogaster | Segregation Distorter (SD) is a selfish, coadapted gene complex on chromosome 2 of Drosophila melanogaster that strongly distorts Mendelian transmission; heterozygous SD/SD+ males sire almost exclusively SD-bearing progeny. Fifty years of genetic, molecular, and theory work have made SD one of the best-characterized meiotic drive systems, but surprisingly the details of its evolutionary origins and population dynamics remain unclear. Earlier analyses suggested that the SD system arose recently in the Mediterranean basin and then spread to a low, stable equilibrium frequency (1–5%) in most natural populations worldwide. In this report, we show, first, that SD chromosomes occur in populations in sub-Saharan Africa, the ancestral range of D. melanogaster, at a similarly low frequency (∼2%), providing evidence for the robustness of its equilibrium frequency but raising doubts about the Mediterranean-origins hypothesis. Second, our genetic analyses reveal two kinds of SD chromosomes in Africa: inversion-free SD chromosomes with little or no transmission advantage; and an African-endemic inversion-bearing SD chromosome, SD-Mal, with a perfect transmission advantage. Third, our population genetic analyses show that SD-Mal chromosomes swept across the African continent very recently, causing linkage disequilibrium and an absence of variability over 39% of the length of the second chromosome. Thus, despite a seemingly stable equilibrium frequency, SD chromosomes continue to evolve, to compete with one another, or evade suppressors in the genome.
| Mendel's first law of segregation holds that a heterozygous parent will transmit alternative alleles to offspring equally. Segregation Distorter (SD) is a naturally occurring selfish gene complex in D. melanogaster that subverts Mendel's first law. During spermatogenesis in heterozygous SD/SD+ males, SD effectively kills SD+-bearing sperm, monopolizing fertilization. SD chromosomes carry a distorter gene and a complement of genetically linked enhancers, often held together by inversions. Thus, SD chromosomes are selfish, co-adapted gene complexes. Although SD is one of our best-characterized selfish gene systems, we still have a poor understanding of its evolutionary history and population dynamics. We therefore performed a large screen for SD chromosomes in African populations of D. melanogaster and studied their genetic properties and history. We found a new SD chromosome type, SD-Mal (endemic to Africa), that has a perfect transmission advantage and lacks recombination over much of the chromosome. This new SD chromosome rapidly swept across sub-Saharan Africa sometime within the last ∼3,000 years. These findings show that selfish gene complexes evolve continuously to evade suppression by other genes in the genome and to compete with one another for a place in the population.
| The Segregation Distorter (SD) system of the fruitfly, Drosophila melanogaster, is a naturally occurring meiotic drive complex—instead of fair Mendelian transmission, heterozygous SD/SD+ males transmit SD chromosomes to most, if not all, progeny [1]–[8]. Full strength distortion is caused by three interacting loci clustered around the centromere of chromosome 2 (an autosome): the trans-acting Segregation distorter (Sd) locus; an upward modifier, Enhancer of SD (E(SD)); and a cis-acting distortion-insensitive allele at the target locus, Responder (Rspi). (By convention, Sd refers to the locus whereas SD refers to chromosomes assumed to carry the full complex of loci.) SD chromosomes are thus Sd E(SD) Rspi, whereas SD+ chromosomes, which lack the distorting Sd locus and usually carry sensitive alleles of Rsp, are Sd+ E(SD)+ Rsps (Figure 1A). During spermiogenesis in heterozygous SD/SD+ males, the sperm-specific histone transition required for proper chromatin packaging is disrupted in Rsps-bearing SD+ sperm, leaving functional Rspi-bearing SD sperm to monopolize fertilization [9]–[12]. For decades, the SD system has been a model in evolutionary genetics, not only for being selfish, propagating at the expense of its bearers, but as a coadapted gene complex whose fitness is determined by multiple epistatic interactors [5]–[7], [13]–[15].
The evolution and persistence of the SD complex depend critically on genetic linkage. Multilocus drive systems can only invade a population when recombination is restricted among loci, as the transmission advantage of distorter chromosomes (Sd Rspi) must not be offset by the formation of so-called ‘suicide’ chromosomes (Sd Rsps) that distort against themselves [16]. The clustering of SD loci around the centromere of chromosome 2, where crossing over is reduced, is therefore unsurprising [15]. Epistatic selection further favors the evolution of secondary suppressors of recombination [15],[17],[18]. Many SD chromosomes, for instance, have recruited a pericentric inversion, In(2LR)39D-42A, that further reduces crossing over in the centromeric region, while some have recruited paracentric inversions on 2R (reviewed in [2],[5],[6],[17]). The paracentric inversions are thought to reduce crossing over between the centromeric SD elements and modifiers of distortion distributed across 2R, such as Modifier of SD (M(SD)), Stabilizer of SD (St(SD)), and possibly others [19]–[22]. Thus, SD chromosomes have evolved a complex of multiple, epistatically interacting loci with coadapted alleles whose linkage relationships are usually further tightened by one or more chromosomal inversions.
The geographic distribution of inversions on different SD chromosomes may shed light on the origins, and possibly the age, of the complex. SD can be found in nearly all populations of D. melanogaster at a frequency of ∼1–5% [23] (but see ref. [24]). In North America, Hawaii, Japan, and Australia, SD chromosomes invariably carry inversions (though not necessarily the same ones). In Italy and Spain, however, both inversion-bearing and presumably ancestral, inversion-free SD chromosomes occur. The presence of both derived and ancestral types has been taken as evidence that SD chromosomes originated in the Mediterranean basin [3],[4]. An origin in Mediterranean Europe further implies that the SD complex evolved recently, as D. melanogaster is a sub-Saharan African species whose range expanded to Europe only ∼15,000 years ago, probably via a single major out-of-Africa founder event [25]–[28]. The first population genetic analysis of SD found little divergence between four loci on SD versus SD+ chromosomes, consistent with a recent origin for the complex [14],[29].
Much about the evolutionary history and population dynamics of SD in natural populations remains unclear. For one, a recent, Mediterranean origin for SD in the D. melanogaster lineage has important implications, explaining its absence from closely related species and suggesting that the multiple genetic components of the complex evolved very quickly. But the Mediterranean origins hypothesis hinges on few data—the presence of inversion-free SD chromosomes from collections in Italy and Spain and nowhere else. For another, what little is known about the population dynamics of SD comes from laborious, large-scale phenotypic assays to determine the frequency of Sd and Rsp in natural populations (e.g., [30],[31]). These have revealed that in natural populations of D. melanogaster worldwide, the frequency of SD is remarkably similar (1–5%) and thus presumably stable. The stability of SD occurs because its intrinsic transmission advantage is balanced by several forces: the sterility of many SD/SD males [32]; the reduced sperm numbers in SD-bearing flies [33]; the presence of suppressors of distortion [24],[30],[31],[34]; and selection against linked deleterious mutations that accumulate in the large non-recombining regions of SD chromosomes [1]. This apparent stability may however mask an underlying evolutionary turnover among competing SD chromosomes predicted by theory [15]. Using realistic parameters, Charlesworth and Hartl [15] showed that an inversion-free SD chromosome will invade a SD+ population and spread to a low-frequency equilibrium; this mixed population, however, is susceptible to invasion by an inversion-bearing SD chromosome that will displace the inversion-free SD and spread to the same low-frequency equilibrium. It appears, then, that SD chromosomes may evolve continuously, as a small subpopulation of second chromosomes in D. melanogaster, competing with one another and evading suppressors.
Fifty years of SD work has produced a rich body of genetics and theory [2], [5]–[8], and recently the molecular basis of distortion has begun to emerge [8], [35]–[38]. The two main SD loci have been identified: Sd is a partial, tandem duplication of the gene RanGAP, called Sd-RanGAP [37]; and Rsp is a large array of 120-bp AT-rich repeats in the centric heterochromatin of 2R [38], where alleles with ≤300 repeats are “insensitive” (Rspi), 700–1100 are “sensitive” (Rsps), and ≥1100 are “super-sensitive” (Rspss; [3],[4],[38]). Sd-RanGAP encodes truncated RanGAP (Ran-GTPase Activating Protein; [37]), a protein with essential and evolutionarily conserved functions in nuclear transport, mitosis, and chromatin processing [39]–[41]. The truncated Sd-RanGAP protein is enzymatically active but mislocalizes to the nucleus (normally RanGAP is cytoplasmic) where, for reasons not yet clear, it causes segregation distortion in SD/SD+ males [8],[35],[36].
Surprisingly, in the decade since its discovery [37], there have been no direct evolutionary analyses of Sd-RanGAP, the gene that actually causes distortion. In this paper, we study the molecular population genetics of the SD complex to investigate its evolutionary history and recent population dynamics. First, we perform the first screen for SD in populations from Africa, the ancestral range of D. melanogaster. Second, we study patterns of DNA sequence variation at the distorter, Sd-RanGAP, as well as its parent gene, RanGAP, and eight noncoding loci on chromosome 2. Finally, we characterize the strength of distortion, inversion status, and mutational load of SD chromosomes. We show that Sd-RanGAP is present in Africa and that a new SD chromosome type has spread very recently across the African continent, causing a large-scale selective sweep among SD chromosomes. These results call into question our current understanding of the timing and location of SD's origins and suggest that, despite its remarkably stable population frequency, SD evolution is not at equilibrium.
We used a three-primer PCR assay to screen 452 isofemale lines collected from 13 localities in Africa for the Sd-RanGAP duplication (Figure 1B; Table 1). We found 12 SD chromosomes from across the continent, including west (e.g., Benin, Gabon, Cameroon) and east Africa (e.g., Zimbabwe, Kenya; Table 2). Assuming that all isofemale lines are homozygous, the population frequency of SD is 12/452 = 0.027; and assuming that all isofemale lines are heterozygous, the population frequency of SD is 12/904 = 0.013. These estimates suggest that SD chromosomes occur in Africa at a frequency of 1.3–2.7%, similar to its frequency in other natural populations [23].
After genetically extracting the SD chromosomes, we sequenced the ∼4.5-kb Sd-RanGAP sequence from all 12 as well as the homologous region of the parent gene, RanGAP, from 10 wildtype (non-SD) chromosomes sampled from Zimbabwe (see Methods). RanGAP and Sd-RanGAP show typical levels of silent divergence per site from the RanGAP homolog in the outgroup species, D. simulans, with Ksil = 0.0471 and 0.0478, respectively. Silent divergence between the duplicate genes, RanGAP and Sd-RanGAP, within D. melanogaster is more than an order of magnitude lower, Ksil = 0.0027 (see also ref. [37]). These findings confirm that Sd-RanGAP arose in D. melanogaster well after the split from D. simulans [29]. Using D. simulans RanGAP as an outgroup sequence, we polarized the substitutions between D. melanogaster RanGAP and Sd-RanGAP. Of five fixed differences between RanGAP and Sd-RanGAP, all were fixed in the common ancestor of the Sd-RanGAP sequences: three noncoding changes, one fixed 6-bp deletion, and a single nonsynonymous change (Figure 2). The first intron of RanGAP contains the gene, Hs2st, raising the possibility that some “silent” changes in one gene are not silent in the other. However, of the five fixed substitutions occurring in Sd-RanGAP, only four affect Hs2st: two are noncoding and two are synonymous.
The amount and distribution of DNA sequence variability at RanGAP is not unusual for an autosomal locus sampled from African populations of D. melanogaster. First, among the 10 wildtype RanGAP sequences, we detect 29 segregating sites (Figure 2), with two measures of DNA sequence polymorphism per site, π = 0.0020 and θ = 0.0023. These values show that RanGAP harbors less variability than the average autosomal locus in African populations (π = 0.0104 and θ = 0.0114; ref. [27]), but this is not unexpected as RanGAP resides in a centromere-proximal region (37E) with a relatively low rate of crossing over and is thus especially susceptible to background selection and hitchhiking effects [42],[43]. Three polymorphisms are synonymous, two are nonsynonymous, and 23 are noncoding, with 66% falling in the large first intron (3.2 kb). The site frequency spectrum at RanGAP does not deviate significantly from standard neutral expectations (Tajima's D = −0.606, P = 0.297; Fay and Wu's H = −4.711, P = 0.127 [44],[45]), where significance was evaluated from 10,000 coalescent simulations conditioning on the observed θ and assuming no recombination). The moderately negative Tajima's D is consistent with recent expansion in the African D. melanogaster populations as inferred from other autosomal loci [27].
Sd-RanGAP is less variable than RanGAP: among the 12 Sd-RanGAP sequences, we detect only five segregating sites (Figure 2), with π = 0.0003 and θ = 0.0004, and a site frequency spectrum skewed towards a moderate excess of rare variants, although not significantly (Tajima's D = −0.313, P = 0.395; Fay and Wu's H = −0.242, P = 0.250). Of the five polymorphisms, two are synonymous and three are noncoding. The lower variability at Sd-RanGAP relative to RanGAP is, of course, expected as Sd-RanGAP is present on only ∼2% of second chromosomes. There are no shared polymorphisms between RanGAP and Sd-RanGAP and hence no evidence for recent gene conversion due to ectopic recombination [46]. The lack of recombination between the two loci implies that Sd-RanGAP evolves as an isolated subpopulation of sequences with a distinct genealogical history.
We found six haplotypes among the ten wildtype RanGAP sequences, with levels of linkage disequilibrium (LD) typical of an autosomal locus in Africa (ZnS = 0.247; [27],[47]). In contrast, the spatial distribution of polymorphic sites in the 4.5-kb Sd-RanGAP sequences is unusual: mutations at five segregating sites are in perfect linkage disequilibrium (ZnS = 1.0), forming just two haplotypes (K = 2). The major haplotype occurs ten times in the sample (M = 10) and the minor haplotype twice. We used coalescent-based haplotype configuration tests to estimate the probability of observing such unusual haplotype structure under standard neutral model assumptions [48]. We performed 100,000 coalescent simulations without recombination (a conservative assumption), assuming a sample size of n = 12 and five segregating sites (S = 5). The cumulative probability that the observed haplotype configuration, or one more extreme, occurs by chance is P = 0.0378. Two features of the haplotype configuration, in particular, differ significantly from the expectations of a neutral genealogical process: the major haplotype is too common in the sample, P(M≥10|n = 12, S = 5) = 0.0313; and there are too few kinds of haplotypes, P(K≤2|n = 12, S = 5) = 0.0285. Both features of the data are consistent with an incomplete selective sweep in which the major haplotype has quickly and recently risen to high frequency, but not fixation, among SD chromosomes [49].
If the major African SD haplotype has indeed risen to high frequency due to positive selection or superior segregation distortion, then the haplotype structure may extend beyond the Sd-RanGAP region. To test this possibility, we sequenced eight noncoding regions across chromosome 2 from all 12 African SD chromosomes and from 10 wildtype chromosomes (Figure 3; Table 3). The amount and distribution of DNA sequence variation among wildtype chromosomes was typical for African D. melanogaster populations [27], with θ ranging from 0.0053 to 0.0137 and Tajima's D ranging from −1.630 to 0.445 (P = 0.045 for locus G, but ≥0.05 for other loci; Table 3). In addition, there is ample evidence for recombination in all but one of the loci (region E has an unusual lack of variability and a small number of haplotypes, though not significantly so; Table 3; Figure 3).
The distribution of variation among SD chromosomes differs strikingly from wildtype chromosomes in two ways. First, the frequency spectra at several loci show patterns consistent with a recent selective sweep. Of the eight noncoding regions surveyed, three loci (J, K, and F) show significant excesses of rare variants (Tajima's D; Table 3), four contiguous loci (J, K, E and F) show significant excesses of high-frequency derived variants (Fay and Wu's H; Table 3), and a fifth contiguous locus (G) possesses no variability at all. The three loci whose frequency spectra do not deviate from neutral expectations include the most distal locus on 2L (M at 37B) and the two most distal loci on 2R (H and I at 58E and 59E, respectively). Second, and more striking, the haplotype structure seen at Sd-RanGAP extends across most of chromosome arm 2R: the 10 major Sd-RanGAP chromosomes possess a single identical haplotype that extends from cytological region 37E on 2L (Sd-RanGAP) to region 55B on 2R (locus G; Table 3; Figure 3). The long distance LD does not extend to the most distal locus on 2L (locus M) or the two most distal loci on 2R (H and I; Table 3; Figure 3). Among all 12 SD chromosomes, forty-five segregating sites occur at Sd-RanGAP and the six regions extending to cytological subdivision 55B. Remarkably, all are differences between the major and minor SD chromosomes (17) or between the two minor SD chromosomes (28). The 10 major SD haplotypes are identical—there is not a single polymorphism in >8.1 kb of sequence. A haplotype configuration test assuming n = 12, S = 45, and no recombination confirms that this haplotype configuration (M = 10, K = 3) is highly unusual under a standard neutral genealogical process (P = 0.00002): the major haplotype is too common in the sample (P[M≥10|n = 12, S = 45]≤0.00001) and there are too few kinds of haplotypes (P[K≤3|n = 12, S = 34] = 0.00002). The chromosomal region between Sd-RanGAP and region 55B (G) spans ≥14 Mb and ∼30 cM, comprising more than 39% of the euchromatic length of chromosome 2. Taken together, the significantly skewed frequency spectra and the existence of an extraordinarily long, high-frequency, mutation-free haplotype suggest a large-scale selective sweep in progress among SD chromosomes [49].
The magnitude of a selective sweep is determined by two key parameters: the local rate of recombination and the strength of selection driving the major haplotype to high frequency. The SD chromosomes carrying the major haplotype are unique in both respects. First, recombination is suppressed along much of the major SD haplotype. We cytogenetically characterized the 12 African SD chromosomes by crossing SD/SD or SD/CyO males to virgin cn bw females (which are homozygous for standard-arrangement second chromosomes) and examined polytene chromosome squashes from larval salivary glands. None of the African SD chromosomes possess the In(2R)NS inversion (52A2–52B1;56F9–56F13) found on most non-African SD chromosomes [2]. Indeed, SD-BN19 and SD-MD31 are inversion-free chromosomes (Table 2). The other ten SD chromosomes, however, possess a complex chromosomal arrangement on 2R (Table 2). The cytological order (40–44F|54E–55E|51BC–44F|54E–51BC|55E–60) shows that these SD chromosomes have recruited two overlapping inversions: In(2R)51BC;55E first, followed by In(2R)44F;54E (Figure 4). These inversion breakpoints match a previously identified, but rare, African endemic chromosomal arrangement found in Malawi [50], hereafter called In(2R)Mal. In addition to In(2R)Mal, four SD chromosomes (SD-KN20, SD-KY87, SD-ZK178, SD-ZK216) carry the cosmopolitan In(2L)t inversion (22D3–D6;34A8–A9; Table 2).
The association between major haplotype and the In(2R)Mal arrangement is perfect: all major haplotype SD chromosomes carry In(2R)Mal, whereas both minor haplotype SD chromosomes (SD-BN19 and SD-MD31) lack In(2R)Mal (Fisher's Exact P = 0.015). Hereafter, we refer to this new class of In(2R)Mal-bearing, major haplotype SD chromosomes as SD-Mal. To test the effect of the In(2R)Mal arrangement on crossing over, we crossed heterozygous SD-NK04/cn bw females (SD-NK04 is a SD-Mal chromosome) to cn bw males and recorded the frequency of recombination between cn (43E16) and bw (59E2). As negative controls, we crossed heterozygous SD-MD31/cn bw females to cn bw males (SD-MD31 is inversion free). Among progeny from the SD-MD31 control crosses with a standard arrangement second chromosome, 34.8% carried recombinant chromosomes (n = 2,155 progeny). In contrast, the In(2R)Mal arrangement almost entirely eliminates crossing over between cn and bw: among progeny from crosses with the In(2R)Mal-bearing SD-NK04 chromosome, only 0.2% carried recombinant chromosomes (n = 1,564). By restricting recombination with wildtype chromosomes, In(2R)Mal sequesters a large piece of chromosome arm 2R as an effectively non-recombining region. The lack of recombination helps to explain the long-range LD produced by the selective sweep (Figure 3) as well as the strong population differentiation at loci between SD-Mal and inversion free chromosomes (Snn = 1.0, P≤0.0001 [51], for the Sd-RanGAP to 55B regions concatenated; Table 4).
We next assayed the strength of segregation distortion by estimating k, the proportion of progeny inheriting SD chromosomes from heterozygous SD/Rsps males. In preliminary work, we found that the dominantly marked balancer chromosome, In(2LR)Gla (hereafter, Gla), carries a sensitive Responder (Rsps). We therefore measured transmission from heterozygous SD/Gla flies (see Methods). Surprisingly, the two SD chromosomes bearing the minor haplotype showed no detectable distortion: k* = 0.538±0.025 for SD-BN19 and k* = 0.415±0.012 for SD-MD31 (mean k*±s.e. are corrected for viability; Table 5). SD-BN19 and SD-MD31 chromosomes also failed to cause distortion when heterozygous against the super-sensitive Rspss allele of the lt pk cn bw chromosome (not shown). In contrast, males heterozygous for SD-Mal chromosomes collectively sired 10,664 progeny and failed to produce a single Rsps-bearing offspring (k* = 1.0; Table 5). The genetic and phenotypic data on recombination and distortion thus provide a clear explanation for the rise of the major haplotype-bearing SD-Mal chromosomes in Africa: they recombine less and distort more.
The complete absence of even low frequency polymorphisms in ∼8.1 kb of sequence distributed from Sd-RanGAP on 2L to cytological subdivision 55B on 2R (G) suggests that SD-Mal rose to high frequency among SD chromosomes quickly and recently. To obtain estimates of the upper 95% confidence limit for the age of the sweep, we assumed that the genealogy relating SD-Mal haplotypes is star-shaped, as expected for a selective sweep, and then estimated the time back to their most recent common ancestor [52]. The expected number of segregating sites in such a sample is E(S) = ntu, where n = number of lineages, t = time in the past when the lineages coalesce into a single common ancestor, and u = the total mutation rate of the sequenced regions. Assuming that the number of mutations on the ten lineages is Poisson distributed, we numerically solved for the probability of observing zero polymorphisms, P(S = 0) = e−ntu, for different times to the common ancestor, t. We used two different estimates of the sequence-specific mutation rate. First, we estimated the mutation rate per generation from θ, which equals 4Neu under standard neutral assumptions, estimated from the wildtype sequences and assuming that Ne = 106 for D. melanogaster. Second, we estimated the mutation rate per year based on the number of fixed differences between D. melanogaster and D. simulans, assuming a divergence time of 3 Mya [53]. The two mutation rates yield qualitatively similar limits for the age of the sweep. Using the polymorphism-based estimate of u, the 95% upper confidence limit for the age of the sweep is 1,875 years. Using the divergence-based estimate of u, the 95% upper confidence limit for the age of the sweep is 3,360 years. Both estimates suggest that the major SD-Mal haplotype expanded across Africa very recently, within the last few thousand years.
We performed complementation tests among all pairwise combinations of the 12 SD chromosomes, producing 12 SDi/SDi and 66 SDi/SDj genotypes. Both minor SD chromosomes (SD-BN19 and SD-MD31) are homozygous viable, but all ten SD-Mal chromosomes are homozygous lethal (Table 6). Crosses among SD-Mal chromosomes, however, show that all ten fall into unique complementation groups—none of the lethal mutations is shared among major SD-Mal chromosomes (Table 6). This distribution of lethal mutations supports a star-shaped genealogy: all of the lethal mutations must have arisen on the external branches of the genealogical history of the SD-Mal chromosomes in our sample. These complementation data also reveal that lethal mutations are significantly over-represented on SD-Mal chromosomes relative to wildtype chromosomes: 29% of wildtype second chromosomes are lethal or semi-lethal [54] versus 100% of SD-Mal chromosomes (Fisher's exact P = 0.0015). The large In(2R)Mal rearrangement on SD-Mal chromosomes provides a large non-recombining target for lethal mutations that can persist by hitchhiking with the SD system.
For the 66 viable SDi/SDj and 2 viable SDi/SDi genotypes, we tested the fertility of both sexes. None of the 68 genotypes were female-sterile, but 10 were male-sterile (Table 6). SD-ZK178 is male-sterile in combination with five other SD chromosomes; SD-KN20 is male sterile in combination with four others; and SD-NK04 is male-sterile in combination with SD-ZK216. The patterns of complementation for male fertility are complex. For instance, SD-KY38 and SD-MD21 complement one another and yet both fail to complement SD-ZK178. Similarly, SD-ZK178 and SD-NK04 complement one another and yet both fail to complement SD-ZK216. Assuming that male sterility results from male-sterile mutations on chromosome 2, the data in Table 6 require a circular complementation map with at least 10 unique lesions. A more plausible hypothesis, however, is that male sterility results not from linked male-sterile mutations but from interactions among different alleles at SD complex loci [32],[55]. Indeed, previous work has shown that deletion of one copy of Sd rescues sterility in otherwise male-sterile SDi/SDj combinations, supporting a connection between distortion and sterility [55]. The complex patterns of fertility complementation in SDi/SDj males cannot, however, be explained by intragenic complementation at the Sd locus, as the Sd-RanGAP sequences among SD-Mal chromosomes are identical, suggesting that interactions involving other SD loci must be involved.
Two major findings emerge from our analysis of the SD system. First, SD occurs in ancestral, African populations of D. melanogaster at a frequency similar to that of other populations worldwide. This discovery raises doubts about the Mediterranean-origins hypothesis. Second, the evolution and rapid spread of a newer, stronger SD chromosome has left a dramatic population genetic signature: a remarkably long haplotype, spanning more than 39% of chromosome 2—roughly 30 cM—that is both free of polymorphisms (Table 3, Figure 3) and differentiated from other chromosomes in the population (Table 4). These findings suggest that a new SD chromosome type endemic to Africa, SD-Mal, has swept across the continent sometime within the last few thousand years.
The Mediterranean-origins hypothesis is based on the geographic distribution of inversions on SD chromosomes: inversion-bearing SD chromosomes occur throughout the world; but both inversion-bearing and inversion-free, presumably ancestral, SD chromosomes occur in Spain and Italy [3],[4]. The presence of ancestral SD chromosomes suggests that the complex may have arisen in Spain or Italy or nearby. Our discovery of SD chromosomes in African populations of D. melanogaster raises questions about the Mediterranean-origins hypothesis. Did SD originate in the Mediterranean and subsequently invade sub-Saharan Africa via back-migration? Or did SD originate in Africa and then make its way to Europe (and the rest of the world) as part of the D. melanogaster out-of-Africa event, ∼15,000 years ago [25]–[28]? The presence of inversion-free SD chromosomes in Benin and Cameroon (SD-BN19 and SD-MD31, respectively) would seem to make a sub-Saharan African origin as likely as a Mediterranean one. In either case, the fact that inversion-free SD chromosomes occur in both Africa and the Mediterranean suggests that Sd-RanGAP dispersed from one location to the other shortly after it originated and then subsequently acquired different inversions on different continents.
The relative youth of the Sd-RanGAP duplication makes distinguishing between sub-Saharan African and Mediterranean origins with the present data difficult. We cannot, for instance, precisely date the origin of Sd-RanGAP from RanGAP based on the five fixed differences (1 indel, 4 nucleotide changes) by assuming a simple neutral molecular clock for two reasons. First, we cannot exclude the rapid, non-neutral fixation of changes in Sd-RanGAP. Second, some (or all) of the five fixed differences may have been segregating as the ancestral RanGAP sequence that ultimately gave rise to Sd-RanGAP. This putative ancestral RanGAP haplotype may be missing from our population sample by chance, or because it was lost from the population, or because it does not occur in African populations. Determining the time and place of origin for the SD system will therefore require deeper resequencing of Sd-RanGAP and RanGAP from both Europe and Africa.
The population genetic analyses revealed six striking patterns among SD chromosomes (Table 3; Figure 3): significant excesses of rare variants; significant excesses of high frequency derived variants; an unusual distribution of haplotype frequencies (10+2 or 10+1+1; Figure 3); exceedingly long-range LD; a complete absence of polymorphism in >8.1 kb spanning >39% of the length of SD-Mal chromosomes; and significant population genetic differentiation between SD-Mal and other chromosomes (Table 4). Together these observations suggest that SD-Mal has spread to high frequency among SD chromosomes in Africa sometime within the last 3,000 years. Why might one type of SD chromosome rise in frequency so quickly, apparently displacing other SD chromosomes? The answer seems straightforward: SD-Mal chromosomes distort more than SD-BN19 and SD-MD31 and recombine less over the length of 2R, perhaps preserving a favorable distortion-enhancing combination of alleles in the In(2R)Mal region. Similar displacement of one SD type (SD-5) by another (SD-72) appears to have occurred during a 30-year period in populations in Wisconsin [31]. Thus, the apparently stable equilibrium frequency of SD chromosomes in D. melanogaster populations worldwide (1–5%) appears to mask a dynamic turnover among competing SD chromosome types.
There are at least two, non-exclusive explanations for the turnover of SD chromosomes. First, the SD system may be sufficiently new that it has not yet reached a stable evolutionary equilibrium: older Sd-RanGAP bearing chromosomes are still being displaced by new ones, like SD-Mal in Africa or SD-72 in North America [31], as predicted by theory [15]. Second, an ultimately stable evolutionary equilibrium for SD chromosomes may not exist: SD may be engaged in a perpetual coevolutionary conflict with the rest of the genome [17]. Indeed, there is considerable variation among populations in the frequency of insensitive Rspi alleles [30],[31] and other unlinked genetic variants that affect distortion (e.g., [24],[34]). Under this scenario, the rise of SD-Mal and decline of SD-BN19 and SD-MD31 could reflect a transitional phase in the genetic conflict in Africa: SD-BN19 and SD-MD31 may no longer cause distortion because they have come under the effective control of unlinked suppressors in the genome, whereas adaptive changes specific to SD-Mal chromosomes allow them to escape suppression.
The discovery of two Sd-RanGAP bearing chromosomes that fail to cause distortion is surprising—indeed, classical phenotypic screens for segregation distortion undoubtedly would have misclassified SD-BN19 and SD-MD31 as wildtype chromosomes. While these chromosomes may now be suppressed, there are four other possibilities. One is that SD-BN19 and SD-MD31 have experienced mutations causing a loss of distortion. Mutational disruption of the Sd-RanGAP sequence seems unlikely, however, as all five differences that distinguish SD-BN19 and SD-MD31 from SD-Mal are silent. A second possibility is that recombination has stripped SD-BN19 and SD-MD31 chromosomes of essential modifiers required for distortion. Wildtype chromosomes that carry Sd-RanGAP transgenes but lack upward modifiers cause either very weak or even no distortion [37]. However, both of these scenarios—disruption by mutation or recombination—require that we explain the seemingly improbable coincidental loss of distortion by two identical, and relatively rare, Sd-RanGAP haplotypes. A third possibility is that SD-BN19 and SD-MD31 are not “SD chromosomes” but rather ancestral Sd-RanGAP-bearing chromosomes that never caused drive. This scenario would imply that SD chromosomes evolved from a neutral, non-driving ancestral haplotype: Sd-RanGAP arose as new duplication, drifted to sufficiently high frequency to become established via migration in Europe and in Africa, and then subsequently recruited genetic modifiers that conferred distortion. This history, if true, implies that African and non-African SD chromosomes independently acquired convergent distorting gene complexes. A final possibility is that SD-BN19 and SD-MD31 may cause distortion but not in the particular genetic backgrounds used in our assay. Further genetic analyses are required to distinguish these possibilities.
The long SD-Mal haplotype spans Sd-RanGAP, region 43E (locus J), and the In(2R)Mal inversions (K, E, F, and G; Figure 3) but does not extend distal to Sd-RanGAP on 2L or distal to In(2R)Mal on 2R. The structure of the SD-Mal haplotype probably reflects the hitchhiking effects of epistatic selection. First, consistent with the lack of loci known to affect distortion distal to Sd-RanGAP, SD and SD+ chromosomes are free to recombine without consequence on the distal part of 2L, preventing LD there [4],[29]. Second, although In(2R)Mal suppresses recombination within the inverted regions, there is opportunity for crossing over in the interval between the SD complex loci (Sd, E(SD), and Rsp) and the proximal breakpoint of the In(2R)Mal. The perfect LD across this interval suggests that strong epistatic selection maintains the association between the SD loci and the In(2R)Mal inversions. In principle, double-recombinants in the interval between centromeric SD loci and In(2R)Mal could preserve their association, but these may be rare events relative to the strength of epistatic selection favoring SD-Mal. Thus, positive epistatic selection on the SD-In(2R)Mal genotype may have caused hitchhiking effects to dominate the intervening sequence between them, explaining the skewed frequency spectrum, LD and lack of variability on SD-Mal chromosomes in region 43E (locus J). It is also possible that epistatic selection directly preserves an association with a M(SD) allele in the SD-In(2R)Mal interval [20], but we do not yet know if SD-Mal carries M(SD). Third, inversions on 2R have been interpreted as tightening the association between SD and St(SD), a modifier (or region of polygenic modifiers; ref. [56]) that increases the strength of distortion, putatively located near the tip of 2R [21],[22]. The fact that we fail to detect LD between SD and loci in cytological regions 58–59 (H and I; Figure 3) suggests that either no St(SD) loci reside in (or distal to) regions 58–59 as previously reported [22] or that no such St(SD) loci enhance distortion on SD-Mal chromosomes. It is important to note that St(SD), like M(SD), was characterized from non-African SD chromosomes; African SD chromosomes may carry a distinct set of linked modifiers.
Although there appears to be competition among SD chromosomes, the overall frequency of SD in populations throughout the world is remarkably similar (1–5%; but see ref. [24]). Considering that different populations have experienced different environments, genetic backgrounds, and demographic histories, the seemingly stable frequency of SD suggests that its equilibrium is the result of strong deterministic forces. What prevents SD from reaching higher frequencies or even fixation? Three factors limit the spread of SD. First, as SD frequency increases, so does selection for insensitive Rspi alleles and other genetic suppressors. Second, as SD frequency increases, intrinsically male-sterile SDi/SDj genotypes become more common, placing an upper-limit on the spread of SD (Table 6; ref. [32]). Third, SD/SD+ males have been shown to suffer reduced male fertility, as might be expected when 50% of sperm are destroyed [9]. Finally, many SD chromosomes worldwide, including the new SD-Mal chromosomes, carry linked recessive lethal and other deleterious mutations (Table 6). The large non-recombining, inverted blocks of chromosome that become associated with SD present a large mutational target. Without recombination, linked recessive lethal and other deleterious mutations are able to persist by hitchhiking with SD. It remains unclear if these factors are sufficient to explain the distortion-selection balance that causes the frequency of SD to settle at 1–5% in D. melanogaster populations worldwide.
The hitchhiking effects of selfish meiotic drive gene complexes have shaped patterns of DNA sequence variability in at least five other cases: four selfish X chromosome systems (one in Drosophila pseudoobscura [57], two in Drosophila simulans [58],[59], and one in Drosophila recens [60]) that drive in the male germline and a selfish autosomal centromere that drives in the female germline of the monkeyflower, Mimulus guttatus [61]. Like SD, all five of these drive systems are associated with haplotypes of reduced variability and three show long-range LD—the signatures of partial selective sweeps. Notably, all five are balanced polymorphisms in which the drive elements are prevented from going to fixation by modifiers or countervailing selection. It is important to note that these well characterized drive systems may not be representative, as there is a clear detection bias: to be discovered and characterized, drive systems must be conspicuous (e.g., causing strong drive or distorting sex ratios) and segregate within populations (i.e., balanced) [7]. But what about those drive elements that are not balanced and thus able to spread to fixation? These would also invade when concentrated in the centromeric regions of autosomes or on sex chromosomes (little or no crossing over occurs between the X and Y) and then sweep through populations, causing complete rather than partial selective sweeps. The extent to which hitchhiking effects of selfish meiotic drive systems contribute to overall patterns of DNA sequence variation, reducing variability around centromeres and on sex chromosomes (e.g., ref. [62]), remains to be determined.
We used a molecular assay to screen for SD chromosomes in a collection of 452 isofemale lines from across sub-Saharan Africa, kindly provided by Drs. John Pool, Charles Aquadro and Andy Clark (Cornell University). We used a single-reaction PCR assay involving three primers, a forward primer (F) and two reverse primers (R1 and R2): F = TTTGGAGACTGCCTGATCAAAACTAATG; R1 = CAACGTCGCGGAGGAGACTGCCTATGT; R2 = CGTGTTCTGAGCGTTTCGCACAGTGTAT. One primer pair (F-R1) amplifies a 463-bp fragment from the parent gene, RanGAP (a positive control), and the other (F-R2) amplifies a 353-bp SD-specific fragment that spans the breakpoint of the Sd-RanGAP-RanGAP junction (Figure 1B). Only one amplicon results from flies that lack SD chromosomes and two result from flies that carry SD (Figure 1B).
Isofemale lines found to be SD-positive by PCR assay could be homozygous SD/SD or heterozygous SD/SD+. We therefore extracted SD chromosomes onto a common genetic background, then maintained homozygous viable SD chromosomes as homozygous stocks, and maintained homozygous lethal SD chromosomes over the CyO balancer chromosome. To extract SD chromosomes, we crossed 3–5 w118; In(2LR)Gla, wgGla-1 Bc1/CyO (hereafter, w118; Gla/CyO) virgin females to 3–5 males from the SD-positive isofemale lines. We then collected 5 white-eyed CyO sons and individually backcrossed them to 5–10 w118; Gla/CyO females. Once larvae appeared in the backcross vials, we PCR-tested the 5 white-eyed CyO sons for SD (see above) and retained progeny from a single SD-positive male. We then crossed w118/w118; SD/CyO virgin daughters to w118; SD/CyO sons. If the SD chromosome was homozygous viable, we used the progeny to establish a w118; SD/SD stock; if the SD chromosome was homozygous lethal, we maintained a w118; SD/CyO stock. Last, we confirmed that all of the final stocks carried the SD chromosome by PCR assay.
Many SD chromosomes possess one or more inversions on chromosome 2 (reviewed in ref. [2]). To determine the inversion types of SD chromosomes, we examined polytene chromosomes from larval salivary gland squashes. We crossed virgin cn bw females to SD males to generate larvae; cn bw chromosomes have standard arrangement second chromosomes. Salivary glands were dissected from F1 larvae in 1% Na-citrate hypotonic solution on siliconized slides and then transferred and fixed for 10–15 seconds in 45% acetic acid. The dissections were stained with 1% lacto-aceto-orcein for 25–35 minutes. We determined inversion breakpoints by comparing photographs with the standard maps of chromosome 2.
We performed complementation tests between all pairwise combinations of SD chromosomes. For all homozygous lethal SD chromosomes, we tested the viability of all SDi/SDj combinations by crossing five SDi/CyO virgin females to 3–5 SDj/CyO males. If CyO+ progeny appear, then the lethality of SDi and SDj chromosomes must map to different complementation groups. We also tested the male and female fertility of viable SDi/SDj combinations. At least two replicates each of 3–5 SDi/SDj males and 3–5 virgin SDi/SDj females were crossed to OreR virgin females and males, respectively. SDi/SDj flies that produced larvae were considered fertile, whereas those that failed to produce any progeny over multiple replicates were considered sterile.
We estimated the strength of distortion for each SD chromosome by measuring the rate of transmission, k, of the SD chromosome through heterozygous SD/Gla males. In preliminary work, we screened a series of balancer chromosomes (Bal) for sensitivity to distortion by assaying transmission from SD-5/Bal males. SD-5 is a well-characterized, non-African SD chromosome. These crosses revealed that the In(2LR)Gla chromosome (hereafter, Gla) carries a sensitive Rsps allele. Gla is an effective balancer of most of the second chromosome and carries a dominant eye-phenotype marker. We estimated k by individually crossing five SD/Gla males of each SD chromosome to five 3–5 day old cn bw virgin females each. After four days, each cross was transferred to a fresh food vial every fourth day. We then scored all progeny emerging until 20 days after the parents were removed from each of the four vials.
The rate of transmission of SD to progeny depends both on the strength of distortion and on the relative viability of the SD chromosome. Therefore, to distinguish the strength of distortion from relative viability, we measured the rate of transmission of SD chromosomes through heterozygous SD/Gla females. As distortion is male-specific, the rate of transmission of SD through females allows estimation of SD relative viability. By using the Gla balancer to minimize recombination on the second chromosome in females, we could estimate the viability of intact SD chromosomes like those transmitted through males (which lack recombination in D. melanogaster). For each SD chromosome we set up three replicate crosses of five 3–5 day old SD/Gla virgin females with three 3–5 day old cn bw males. After four days, each cross was transferred to fresh vial every fourth day. We used our estimates of relatively viability to estimate a corrected strength of distortion, k*, following ref. [63].
To sequence the new Sd-RanGAP duplicate gene, we first isolated SD chromosomes in heterozygous state over a chromosomal deficiency, Df(2L)Sd77, which deletes the 37D1–37D2;38C1–38C2 region including the RanGAP locus. After isolating genomic DNA from SD/Df(2L)Sd77 flies, we PCR amplified two fragments from the Sd-RanGAP region with two sets of primers. All PCR products therefore come from the SD chromosome. The first set amplifies a 2,994-bp fragment from the 5′-half of Sd-RanGAP. The forward primer (F4) binds the distal intergenic region between Sd-RanGAP and the neighboring gene CG10237; the reverse primer (R4) binds in intron 1 of Sd-RanGAP (which, on the reverse strand, is exon 2 of Hs2st). The second primer set amplifies a 2,410-bp fragment from the 3′-half of Sd-RanGAP with a 280-bp overlap with the first fragment. The forward primer (F6) binds in the first intron of Sd-RanGAP (which, on the reverse strand, is intron 2 of Hs2st); the reverse primer (R6) binds the intergenic region between Sd-RanGAP and RanGAP. Both the R4 and F6 primers bind two genomic locations in flies with SD chromosomes. First, R4 binds the first intron of Sd-RanGAP and the homologous sequence of the parent gene RanGAP. However, when the F4-R4 primer pair is used and PCR extension times are constrained, only product from the first R4 binding location results. Second, F6 binds the first intron of Sd-RanGAP and the homologous sequence of RanGAP. However, when the F6-R6 primer pair is used, only the 3′-half of Sd-RanGAP is amplified. We used Exo-SAP to clean PCR products and then sequenced both strands of the PCR products using internal sequencing primers (Table S1), BigDye Terminator chemistry, and standard cycle sequencing protocols. All sequences were manually edited using Sequencher v. 4.5 (Gene Codes). We obtained outgroup sequences via BLAST searches of the D. simulans genome [62].
In addition to Sd-RanGAP, we sequenced the parent gene and eight non-coding regions across chromosome 2 from a collection of SD chromosomes and from 10 wildtype chromosomes from Zimbabwe. As many SD chromosomes, and some wildtype ones, are homozygous lethal (see RESULTS), we could not make homozygous lines for sequencing for all stocks. Instead, for homozygous lethal lines, we used deficiencies to produce flies hemizygous for the focal chromosomal regions. The eight regions ranged in size from 567–874 bp long (Table 3). We sequenced fragments from the proximal intergenic region of tup (cytological position = 37B; deficiency used for hemizygous flies = Df(2L)Exel7073); a large intron from CG30947 (43E; Df(2R)Exel6054); the distal intergenic region of Myd88 (45C; Df(2R)Np3); the proximal intergenic region of off-track (48D6; Df(2R)BSC39); the proximal intergenic region of scab (51E; Df(2R)Jp1); the proximal intergenic region of staufen (55B; Df(2R)Pcl7B); a large intron of plexus (58E4–8; Df(2R)Exel7173); a large intron of CG34372 (59E1; Df(2R)bw-S46). To sequence the parent gene, RanGAP, from the 10 wildtype chromosomes, we used the Df(2L)Sd77.
We performed most population genetic analyses using DnaSP [64]. Probability values for Tajima's D and Fay and Wu's H were obtained from 10,000 coalescent simulations with no recombination, conditioning on the observed θ. For coalescent-based haplotype configuration tests we used the haploconfig software [48].
|
10.1371/journal.pcbi.0040001 | Chemotaxis in Escherichia coli: A Molecular Model for Robust Precise Adaptation | The chemotaxis system in the bacterium Escherichia coli is remarkably sensitive to small relative changes in the concentrations of multiple chemical signals over a broad range of ambient concentrations. Interactions among receptors are crucial to this sensitivity as is precise adaptation, the return of chemoreceptor activity to prestimulus levels in a constant chemoeffector environment. Precise adaptation relies on methylation and demethylation of chemoreceptors by the enzymes CheR and CheB, respectively. Experiments indicate that when transiently bound to one receptor, these enzymes act on small assistance neighborhoods (AN) of five to seven receptor homodimers. In this paper, we model a strongly coupled complex of receptors including dynamic CheR and CheB acting on ANs. The model yields sensitive response and precise adaptation over several orders of magnitude of attractant concentrations and accounts for different responses to aspartate and serine. Within the model, we explore how the precision of adaptation is limited by small AN size as well as by CheR and CheB kinetics (including dwell times, saturation, and kinetic differences among modification sites) and how these kinetics contribute to noise in complex activity. The robustness of our dynamic model for precise adaptation is demonstrated by randomly varying biochemical parameters.
| Bacteria swim in relatively straight lines and change directions through tumbling. In the process of chemotaxis, a network of receptors and other proteins controls the tumbling frequency to direct an otherwise random walk toward nutrients and away from repellents. Receptor clustering and adaptation to persistent stimuli through covalent modification allow chemotaxis to be sensitive over a large range of ambient concentrations. The individual components of the chemotaxis network are well characterized, and signaling measurements by fluorescence microscopy quantify the network's response, making the system well suited for modeling and analysis. In this paper, we expand upon a previous model based on experiments indicating that the covalent modifications required for adaptation occur through the action of enzymes on groups of neighboring receptors, referred to as assistance neighborhoods. Simulations show that our proposed molecular model of a strongly coupled complex of receptors produces accurate responses to different stimuli and is robust to parameter variation. Within this model, the correct adaptation response is limited by small assistance-neighborhood size as well as enzyme kinetics. We also explore how these kinetics contribute to noise in the chemotactic response.
| Through the process of chemotaxis, the bacterium Escherichia coli swims up the concentration gradients of attractants (nutrients) and down the concentration gradients of repellents. E. coli moves via the rotation of multiple flagella. When the flagella rotate counterclockwise, they bundle and propel the bacterium forward. Rotation in a clockwise direction causes the flagella to fly apart, and the organism tumbles to change direction. Swimming up a gradient of attractant causes a decrease in the probability of tumbling, whereas swimming up a gradient of chemorepellent causes an increase in the probability of tumbling. The result is that E. coli performs a biased random walk toward chemoattractants and away from chemorepellents [1].
The signaling pathway that governs E. coli chemotaxis is well characterized [2–4]. Out of five different membrane-bound chemotaxis receptors, Tar and Tsr are expressed at high levels, whereas Tap, Trg, and Aer are expressed at lower levels. The receptors form homodimers that can each bind one molecule of ligand [5]. The homodimers in turn form trimers of dimers, and associate with CheW and CheA. CheW is a linker protein, and CheA is a histidine kinase [6,7]. Receptor signaling activates CheA autophosphorylation, and the phosphoryl group is transferred to the response regulator, CheY. Phosphorylated CheY diffuses and binds to the flagellar motors, favoring clockwise rotation and tumbling. CheY is dephosphorylated by the phosphatase CheZ.
E. coli are able to react to small relative changes in concentration over a range of several orders of magnitude. In experiments done by Mao et al. [8], bacteria responded to changes in concentration from 10 mM to as low as 3.2 nM of the attractant aspartate. Two properties of the network that underlie the broad range of responsiveness are interactions among receptors and precise adaptation [9]. In vivo fluorescence resonance energy transfer (FRET) measurements [10,11] suggest that signaling is mediated by strongly coupled complexes of 10–20 receptor homodimers that are all active or inactive together. FRET also reveals that levels of phosphorylated CheY adapt precisely following a transient response to steps of chemoeffector concentration. Precise adaptation occurs though the methylation by CheR and demethylation by CheB of eight sites on each homodimer receptor [12,13]. Methylation at each site increases the activity level of receptors, whereas demethylation decreases activity. Each Tar or Tsr receptor has a tether at its C terminus, with a pentapeptide site that can bind one CheR or CheB [14]. Experiments indicate that when transiently bound to one receptor, each CheR or CheB can act on five to seven adjacent receptor homodimers, defining an “assistance neighborhood” (AN) [15].
The dynamics of receptor modification in complexes is not well understood. A two-state single-receptor model was proposed by Barkai and Leibler in which the modification activities of CheR and CheB depend only on receptor activity, not ligand concentration or methylation level [16]. This simple model naturally leads to precise adaptation, but does not include interactions among receptors. More recent approaches have incorporated receptor interactions using a Monod–Wyman–Changeux (MWC) model [17] in which a complex of receptors is either active (on) or inactive (off) as a whole [11,18–20]. The free-energy difference between the on and off states of a complex dictates the probability of its being in the active state. To preserve precise adaptation within the MWC model, the Barka–Leibler (BL) model was extended [18,19] to include the action of CheR/CheB on ANs of receptors. Static nonoverlapping ANs of size 6 were utilized. Here, we build on this earlier model by incorporating the binding and unbinding of CheR and CheB, creating dynamic ANs. This extension allows us to consider limits to precise adaptation from small AN size as well as from CheR and CheB kinetics, including dwell times, saturation, and kinetic differences among modification sites.
Models for the E. coli chemotaxis network are complex and depend on numerous parameters, bringing into question how well the essential property of precise adaptation is preserved when parameters are altered. Barkai and Leibler showed that their simple two-state single-receptor model of the chemotaxis network was robust to parameter variation [16]. We find a similar robustness of our dynamic AN model.
In this paper, we explore a MWC model of mixed and strongly coupled Tar and Tsr chemoreceptor homodimers. Within the MWC model, a complex of receptors is either on or off as a whole (Figure 1). The average complex activity is the probability that the complex is in the on state and is determined by the free-energy difference between the on and off states [19]. Here, the MWC model is used to calculate the thermal equilibrium complex activity from the instantaneous attractant concentration and receptor methylation state.
We assume that each receptor homodimer is a two-state system, being either on or off. Each receptor homodimer can bind a ligand molecule in either state, albeit with different affinities. Therefore, the four possible configurations for each homodimer and their free energies are (1) on with no ligand bound,
, (2) on with ligand bound,
, (3) off with no ligand bound,
, and (4) off with ligand bound,
. Here
and
are the binding constants in the on and off states for a specific type of receptor r, and m is the methylation level (m = 0,...,8). Based on experimental data, these binding constants are assumed to be independent of ligand concentration or methylation level [21–23]. For the two on states, the sum of the equilibrium Boltzmann factors is
therefore, the combined free energy of the two on states is
. Similarly, the combined free energy of the two off states is
. All energy units are expressed in units of the thermal energy, kBT. Since binding of attractant favors the off state, Kon > Koff. The opposite applies to repellents: Kon < Koff.
Adaptation to attractant occurs through methylation, which favors the on state. Therefore, the offset energy
decreases as m increases. The free-energy difference between the on and off states of a single receptor is
The free-energy difference F of a complex of receptors is the summation of the individual fr(m) of all of the receptors in the complex,
The average activity A of the complex of receptors is its probability of being in the on state and is given according to equilibrium statistical mechanics by
Within our model, receptors dynamically bind and unbind the adaptation enzymes CheR and CheB (Figure 1A). We assume that a receptor-bound CheR only methylates receptors when the complex is off, whereas a receptor-bound CheB only demethylates receptors when the complex is on (Figure 1B). For precise adaptation to occur, the rates of methylation and demethylation by CheR and CheB must depend only on the activity of the complex A. We assumed each bound CheR adds methyl groups at a rate kR(1 − A), and each bound CheB protein removes methyl groups at a rate kBA. In most simulations, we assumed saturated kinetics for CheR and CheB. Specifically, we assumed each receptor-bound CheR or CheB acts on available sites in the AN with equal probability, independent of the number of available modification sites. However, we also explored the effects of nonsaturated kinetics by introducing a factor N/(N + Msat) into the methylation/demethylation rates, where N is the total number of available sites for methylation/demethylation and Msat is a constant. Smaller values of Msat mean more nearly saturated kinetics, with full saturation corresponding to Msat = 0. For all of our simulations, the methylation/demethylation rate is zero if there are no available modification sites (N = 0).
Unlike the previous AN model [18,19], we include dynamical CheR/CheB binding and unbinding to receptors. Free receptors bind CheR/CheB molecules at a rate
, and receptor-bound CheR/CheB molecules unbind at a rate
. In addition, each receptor can bind at most one CheR or one CheB. At steady state, this gives for the average proportion of receptors bound by CheR (with a similar expression for CheB):
The CheR/CheB binding rates
are assumed to increase linearly with the concentration of free CheR/CheB, whereas the CheR/CheB unbinding rates
are assumed to be independent of concentration. As the concentration of CheR, and therefore
, rises, the proportion of CheR-bound receptors
also rises. Since each receptor can bind at most one CheR or one CheB, the proportion of CheR-bound receptors
decreases with an increase in the CheB binding rate
. The CheR/CheB unbinding rates also define an average dwell time each CheR or CheB is bound to a receptor to be
.
A fixed hexagonal arrangement of 19-receptor homodimers (Figure 1A) was used for every simulation. For simplicity, from this point, we refer to each homodimer as a receptor. The ANs consist of a receptor and its nearest neighbors. This creates 19 possible ANs, each centered on one receptor: seven ANs of size 7, six ANs of size 5, and six ANs of size 4. The average AN size is 5.4 receptors. Six receptors were Tar, and 13 receptors were Tsr, consistent with the wild-type ratio [24]. We also explored the effect of AN size through the use of size-one AN complexes and half AN complexes. For size-one AN clusters, each CheR or CheB can only modify the bound receptor. For half AN complexes, we randomly chose half of each receptor's nearest neighbors to be in the AN and used the same configuration for all simulations. For the six receptors with three nearest neighbors, three have half ANs including two adjacent neighbors, and the other three have half ANs including only one adjacent neighbor.
By considering the average net methylation rate of a complex, we derived a simple mean field theory for complex activity. The average methylation rate of a complex is the rate of methylation by a single CheR times the number of bound CheR with available methylation sites. Similarly, the average demethylation rate of the complex is the rate of demethylation by a single CheB times the number of bound CheB with available demethylation sites. Therefore, the average net methylation rate for a single receptor is
where
and
are the average proportions of fully methylated and fully demethylated ANs, respectively. The factors
and
account for the fact that CheR cannot methylate an already fully methylated AN, and CheB cannot demethylate an already fully demethylated AN. The condition that the average net methylation rate is zero (
) determines the average steady-state activity,
As long as
, the activity will always adapt precisely to
which is independent of ligand concentration or methylation level, since
and
depend only on the concentrations and binding rates of CheR and CheB.
Fluctuations in methylation and demethylation will lead to finite values of
and
. However, for large enough ANs, the probability that a neighborhood will become fully methylated or fully demethylated by chance will be very small (as long as the average receptor methylation level is not close to
or
), so
will be a good approximation. Therefore, we expect the activity in large-AN models to adapt to A* over a broad range of ligand concentrations.
As the average receptor-methylation level reaches
, the methylation level cannot increase to compensate for the increased free-energy difference due to attractant binding. Therefore, activity drops to zero beyond the limiting ligand concentration at which receptors become fully methylated. At full methylation of the complex,
where ns is the number of Tsr receptors in the complex and na is the number of Tar receptors. Therefore, failure of precise adaptation begins near the ligand concentration [L] for which Fm=8 = F*, i.e., the value of F at precise adaptation. Further increase in attractant concentration causes a rapid decay in activity:
Here, we derive an analytical expression for the fluctuations of complex activity due to discrete methylation and demethylation events by receptor-bound CheR and CheB. First, we calculate the variance of the total complex methylation level within the Langevin approximation [25]. If the free-energy difference between the active and inactive state of a receptor depends linearly on the methylation level m, then for a single receptor
and for a complex of ns Tsr receptors and na Tar receptors (with total methylation level M),
Since CheR/CheB methylation/demethylation rates depend on complex activity, fluctuations in the free-energy difference F translate into fluctuations in complex activity A. Linearization of Equation 5 with
yields for
,
where NR/NB are the number of bound CheR/CheB enzymes, ηR(i)/B(i) are independent Langevin noise terms for each bound CheR/CheB, and
. After a Fourier transform and integration of the power spectrum, we obtain (with
and
single CheB demethylation rate =
)
The variance in methylation level depends only (inversely) on δε, the step in free-energy difference per added methyl group. The variance in activity is therefore
Figure 2 shows simulated response curves of complexes of 19 chemoreceptor dimers to step increases in concentration of alpha-methyl aspartate (MeAsp), an attractant. The results shown include the dynamics of CheR and CheB (see Model) and are similar to those obtained with static ANs [18]. Precise adaptation occurs over four orders of magnitude of MeAsp concentration, with methylation levels increasing to compensate for drops in activity due to increases in attractant concentration.
In Figure 2, the Tar-only complexes exemplify two different limits of precise adaptation at high attractant concentrations, as in the static AN model [18]. For the Tar-only complex with higher
(dot-dashed curve), the activity continues to adapt precisely, but the activity stops responding to increases of MeAsp. In this case, the receptors become saturated, and further increases in MeAsp do not produce changes in the free-energy difference between the on and off states of the complex. The average methylation of the complex reaches a constant value, below full methylation (Figure 2B). In contrast, for the complex with lower
(dashed curve), the activity approaches zero at high concentrations. In this case, full methylation occurs before saturation of the receptors with MeAsp. When MeAsp concentrations increase further, the resulting increase in the free-energy difference between the on and off state of the complex cannot be compensated by additional methylation, so the activity drops without recovering.
Compared to the Tar-only complexes, the heterogeneous receptor complex with six Tar and 13 Tsr receptors (solid curve) continues to respond to MeAsp increases and to adapt precisely over an extended range. The Tar receptors in this complex have
, as in the second case considered above (dashed curve); these six Tar receptors allow for a sensitive response at low concentrations of MeAsp. In contrast, the 13 Tsr receptors in the complex have low affinity for MeAsp. Therefore, at low MeAsp concentrations the Tsr receptors act as extra methylation sites, increasing the range of precise adaptation. As MeAsp concentrations increase, the Tar receptors become fully saturated, but the Tsr receptors begin to bind MeAsp. This increases the upper limit of response to well over 100 mM MeAsp.
For homogeneous complexes, the limit of adaptation at high attractant concentration depends on which occurs first, saturation of receptors by attractant or full methylation of receptors. Which of these occurs first depends on the ratio
. The crossover ratio between the two limits of adaptation can be estimated in mean field theory (Equation 2). At high ligand concentrations:
where n is the number of receptors in the complex. Loss of activity occurs if the offset energy at full methylation εr(8) cannot compensate for the free-energy difference per receptor due to saturating attractant,
. Therefore, for F* denoting the precisely adapted free-energy difference, if
or equivalently, if
, loss of activity will occur at high concentrations of attractant (dashed curve in Figure 2A). In contrast, if
, loss of response will occur at high concentrations of attractant (dot-dashed curve in Figure 2A). For fixed εr(m), F*, and complex size n, the limit of adaptation depends only on the ratio
, not on the individual magnitudes of
and
. For our simulation, εr(8) = −30, n = 19, and A* = 1/3, so F* = log2 = 0.693. Therefore, the expected crossover ratio is exp(F*/n − εr(8)) = 20.8. For Tar-only complexes with
,
, so adaptation fails through loss of activity as observed in Figure 2A (dashed curve). For Tar-only complexes with
,
, and adaptation fails through loss of response, as also observed in Figure 2A (dot-dashed curve).
Experiments indicate that the adapted tumbling rate, and therefore, also the adapted receptor activity, increases with the concentration of CheR [26]. At low levels of CheR, the binding rate
is proportional to the concentration of CheR. In Figure 3, we show the adapted activity as a function of the CheR binding rate
. Adapted activity is the average activity calculated according to Equation 3, after allowing the complex to reach equilibrium (see Methods). As the binding rate of CheR increases, the proportion of CheR-bound receptors
also increases (Figure 3, inset). The increase in
causes the rate of methylation for the whole complex to rise, therefore increasing activity. The complex with full ANs (including all nearest neighbors) closely follows the expected mean-field-theory result (Equation 7), whereas the complex with ANs of size one deviates to higher activity over a wide range of CheR binding rates. In these simulations, no attractant is present, and therefore, the average methylation level of the receptors is low. Consequently, complete demethylation of individual receptors is likely to occur in the AN = 1 model (
), leading to missed demethylation attempts, and therefore, to an increase in adapted activity according to Equation 6. In effect, for the AN = 1 model, the demethylation rate is lower than it “should be” because by chance, some individual receptors are already fully demethylated, and therefore, CheB fails to act sometimes when it “should.”
In Figure 4, we explore precision of adaptation over a broad range of MeAsp concentrations for several variants of our model. In general, deviations from precise adaptation occur if and only if the rates of methylation or demethylation cease to depend exclusively on complex activity (Equation 6). We find that large AN sizes, saturated kinetics of CheR/CheB, and short CheR/CheB dwell times favor precise adaptation. In all cases, we consider the same complex of 19 receptors (Figure 1A) composed of six Tar receptors and 13 Tsr receptors. For comparison, we also show the mean field theory result (see Model).
In Figure 4A, we show the effect of AN size on precise adaptation. Within each AN, there is a “ladder” of possible methylation levels. Fluctuations cause the methylation level to move up and down the ladder, deviating from the average. For small ANs, the ladder is shorter, and fluctuations are more likely to produce fully methylated or fully demethylated neighborhoods. At low levels of MeAsp and low average methylation, fluctuations are likely to produce fully demethylated neighborhoods, lowering the rate of demethylation and increasing activity according to Equation 6. Similarly, at high levels of MeAsp and high average methylation, neighborhoods may become fully methylated, lowering the rate of methylation by CheR and decreasing activity.
As shown in Figure 4A, complexes with ANs of size one have a drastically reduced precision of adaptation, but half neighborhood complexes have a precision of adaptation close to that of full AN complexes. Beyond a certain AN size, the methylation ladder is already long enough to effectively prevent fluctuations from causing full methylation or full demethylation of neighborhoods. Therefore, increasing AN size improves precision of adaptation only up to a point, beyond which AN size only affects activity near the concentration at which all receptors become fully methylated. For our parameters, receptors do not become fully demethylated even at zero attractant concentration, but full demethylation could be induced by addition of repellent.
We also performed simulations with varying degrees of saturation of CheR and CheB (Figure 4B). Specifically, we introduced a factor of N/(N + Msat) into the rates of CheR and CheB action, where N is the total number of available sites for methylation/demethylation and Msat is a constant (see Model). For all other simulations, CheR and CheB were assumed to work at saturation, independent of methylation level (Msat = 0). Increasing Msat makes the rate of action of CheR and CheB more dependent on the number of available modification sites. For finite Msat, in low concentrations of MeAsp, and therefore, low average methylation levels, the rate of demethylation by CheB is significantly lower than the saturated (maximal) rate. Conversely, there are many available sites for methylation, so the rate of methylation is near maximal. Therefore NB/(NB + Msat) < NR/(NR + Msat) ≈ 1. A relative decrease in the rate of demethylation by CheB compared to the rate of methylation by CheR causes an increase in the activity of the complex as seen below 0.1 mM MeAsp in Figure 4B. As the average methylation level increases with increasing MeAsp concentration, the inequality is reversed so that NR/(NR + Msat) < NB/(NB + Msat) ≈ 1 results in a relative decrease in the rate of methylation by CheR compared to the rate of demethylation by CheB. Therefore, at high MeAsp concentration, above 0.1 mM, the adapted activity decreases below the expected value for precise adaptation.
Within mean field theory for Msat > 0, we can approximate the crossover concentration, i.e., the concentration of attractant at which the activity of complexes is equal to the expected precisely adapted activity. The crossover occurs when the saturation factors of CheR and CheB are equal, NB/(NB + Msat) = NR/(NR + Msat). This occurs when the average methylation level
is 4, which occurs at 0.36 mM MeAsp. This is close to the crossover concentration observed in our simulations (Figure 4B).
Simulations were also performed in which the average dwell time of CheR and CheB was varied (Figure 4C). The average dwell time is equal to
, whereas the average number of enzymes bound to the complex depends on the ratios
(Equation 4). Therefore, in order to change the average dwell time while conserving the average number of CheR and CheB enzymes bound to the complex, we altered both
and
by the same factor. In the model, when a CheR or CheB is bound for a long time, the enzyme catalyzes the same reaction numerous times before unbinding. The methylation level in the neighborhood will therefore move along the ladder in one direction, possibly reaching the end, i.e., full methylation or full demethylation. As for the AN = 1 model in Figure 4A, the result in Figure 4C for long CheR and CheB dwell times is higher activity at low MeAsp concentrations (where full demethylation is more likely) and lower activity at high MeAsp (where full methylation is more likely).
Deviations from mean field theory occur if the average dwell time of CheR or CheB is long enough to allow full methylation or demethylation of neighborhoods. Below 0.001 mM MeAsp, the average adapted methylation level per receptor homodimer is ≈ 2.2. Since there are on average 5.4 receptors per AN, the average distance to the bottom of the methylation ladder is 2.2 × 5.4 ≈ 12. Therefore, precise adaptation is expected to fail when the demethylation rate is ≈12 times the CheB unbinding rate (
). For our parameters, A* = 1/3 and kB = 0.2 s−1, we expect precise adaptation to fail for
. Consistent with this calculation, our simulations show that deviations from precise adaptation begin to occur at low MeAsp concentrations for
around 0.01 s−1.
The fact that most receptors are either fully methylated or fully demethylated for long dwell times of CheR and CheB is clearly shown by the distribution of methylation levels for different average dwell times (Figure 4C, inset). As dwell time increases, the single-peaked methylation distribution flattens out and becomes bimodal, i.e., most receptors become fully methylated or fully demethylated. Addition of ligand causes a shift in the amplitudes of the two peaks, but the peak positions, at m = 0 and m = 8, do not change. We can exploit this fact along with mean field theory to estimate the crossover attractant concentration where the activity of the complex crosses A*. The average methylation (demethylation) rate has a correction factor equal to the proportion of not fully methylated (not fully demethylated) ANs (Equation 6). The crossover attractant concentration will occur where these two correction factors are equal, namely where
, which implies
. For our mean field–adapted activity of A* = 1/3, and requiring
, we obtain a crossover concentration of 30 mM MeAsp, consistent with the simulation results shown in Figure 4C.
In all our simulations, the methylation levels of receptors fluctuate, translating into fluctuations in complex activity. Figure 5 shows the distribution of activities due to fluctuating methylation levels at 0 mM, 1 mM, and 100 mM of MeAsp. Within the MWC model, complex activity is strictly either zero or one. However, we assume that switching between these two states is rapid, so we consider the distribution of thermally averaged complex activities given by Equation 3. Even for the full AN model, for which adaptation is precise, there is a broad range of complex activities. Note though that for the observed variation in activity of ≈50% for a single complex and assuming ≈500 independent receptor complexes per cell, the resulting variation in total activity would be only ≈2.5%. As shown in Figure 5, for size-one ANs at 0 mM and 100 mM MeAsp, the activity distributions are shifted relative to the activity distributions for full ANs because adaptation is not precise when CheR and CheB act only on single bound receptors (cf. Figure 4A). Also shown in Figure 5, long dwell times of CheR and CheB cause a bimodal distribution of complex activities, corresponding to the bimodal distribution of receptor methylation levels (cf. Figure 4C, inset).
Within our model, noise is caused by fluctuations in both binding/unbinding of CheR and CheB and methylation/demethylation by CheR/CheB. For short average dwell times, fluctuations in the number of bound CheR and CheB enzymes are rapidly averaged out, and the dominant source of noise is the discrete methylation/demethylation events by receptor-bound CheR/CheB. We have estimated the resulting variance in complex methylation
and activity
with the linear noise approximation (see Model and Figure 6). In this limit, the only factor that affects the variance
is the free-energy difference δε per methyl group, with
. In the opposite limit of long average dwell times, fluctuations in the number of CheR and CheB enzymes bound to the complex add to the variance in methylation levels and thus activity. As seen in Figure 6A and 6C, low binding and unbinding rates
cause an increase in noise over the calculated theoretical noise limit due to the discreteness of methylation and demethylation events. Increasing complex size can decrease the noise due to CheR and CheB binding/unbinding, but not the noise due to CheR/CheB methylation/demethylation. Therefore, increasing complex size only decreases noise for long average dwell times of CheR and CheB, but has no effect in the case of short dwell times, where noise is near the theoretical limit (Figure 6B and 6D).
It was observed experimentally by Chalah and Weis [27] that CheR and CheB methylate/demethylate the four different methyl-attachment sites on each receptor monomer at different rates. These observations suggest two possible scenarios: either CheR and CheB have different rates of action on different modification sites, or CheR and CheB divide their time unequally among the sites (or some combination of these two). To test the first scenario, we extended our model to include variation in the rates of action of CheR and CheB, with the results shown in Figure 7. Specifically, we assumed that when a CheR or CheB is tethered to a receptor, it divides its time equally among all available modification sites in the AN. The total rate of action by a bound CheR or CheB is therefore the average over the rates for all available modification sites in the AN. The catalytic rates for a methylation/demethylation reaction were assumed to vary in the ratio 1:2:4:8 for the four different sites [27]. We studied two cases. In the first case, the ratios of methylation and demethylation matched for each site (i.e., for sites 1–4, the ratios for both kB and kR were 1:2:4:8). In the second case, the ratios for methylation and demethylation were inverted relative to each other (i.e., for sites 1–4, the ratios for kB are 1:2:4:8 and for kR 8:4:2:1).
As shown in Figure 7, when the ratios of methylation and demethylation match for each site, precise adaptation is preserved. In this case, since every site has the same ratio of kB/kR as every other site, the average methylation levels of all sites remain the same, as shown in the inset. The average methylation and demethylation rates over sites is therefore constant, independent of ligand concentration, preserving precise adaptation. In contrast, inverted ratios of methylation and demethylation rates among the sites fail to produce precise adaptation. In this case, the ratio kB/kR varies among the four methylation sites, causing varying equilibrium methylation levels (Figure 7, inset). The sites with a low kB/kR ratio are the first to become methylated at low concentrations of MeAsp, leading to low average rates of demethylation compared to methylation, and therefore to high adapted activity. As average methylation levels rise with increasing MeAsp, these low kB/kR sites “fill up,” leading to high average rates of demethylation compared to methylation, and therefore to low adapted activity.
The second scenario suggested by the Chalah and Weis data [27], namely different dwell times for CheR and CheB among the modification sites, leads more robustly to precise adaptation. As long as CheR and CheB work near saturation, differences in dwell times between sites will not affect total rates of methylation and demethylation, and precise adaptation will be preserved, according to Equation 6. Indeed, as shown in Figure 7, even if the relative dwell times for each site are inverted for CheR and CheB, precise adaptation is preserved.
Experiments by Berg and Brown [9] on wild-type E. coli indicate that whereas adaptation to aspartate is precise over a large concentration range, precise adaptation to serine fails at relatively low concentrations. In Figure 8, we compare our model to these experiments. In both cases, adaptation to aspartate (or MeAsp) is precise over four orders of magnitude. However, adaptation to serine fails at approximately 0.1 mM. Within our model, this difference with respect to attractants reflects the presence of more Tsr receptors (13) in the complex than Tar receptors (six). More Tsr receptors amplify the change in complex free energy due to serine, which results in an increased sensitivity at low concentrations, but also results in full methylation of the complex and loss of activity beginning at 0.1 mM serine.
We tested robustness of our theoretical model by randomly varying parameters as described in the Model section. The results shown in Figure 9 demonstrate that precise adaptation is a robust property of our model. Almost ideal adaptation occurs for all parameter sets up to a total parameter variation of K ≈ 1018. For larger parameter variations, in the range of K = 104–105, 77% of the altered models still have a precision of adaptation within 10%. These results are similar to those obtained from the simple single-receptor model of Barkai and Leibler [16]. However, in one regard, our MWC model with ANs is more robust than the single-receptor model. In the single-receptor model, precise adaptation requires that the activity of the receptor is zero at full demethylation and one at full methylation. Our model has the property of precise adaptation without this assumption.
The chemotaxis system in the bacterium E. coli is remarkably sensitive to small relative changes in the concentrations of multiple chemical signals over a broad range of ambient concentrations. We have presented a model of complexes of strongly coupled chemoreceptors to account for precise adaptation, as well as other properties of the chemotaxis network. Similarly to the BBL model of precise adaptation for a single two-state receptor, CheR only methylates inactive receptors, and CheB only demethylates active receptors [16]. A previous MWC model [18,19] extended the BL model to fit observations of in vivo receptor clustering [10,11] and of the action by CheR and CheB on ANs of five to seven adjacent receptor homodimers [15]. Our model builds upon this earlier MWC model, but includes dynamic ANs, created by the transient binding of CheR and CheB to tethering sequences at the C termini of receptors [14]. The importance of tethering of enzymes has recently attracted considerable theoretical interest [28–31]. Transient CheR and CheB binding is of particular relevance because the experimentally observed ratio of receptor homodimers to the enzymes CheR and CheB is approximately 50:1 and 30:1, respectively [4]. Therefore, receptors are not likely to be continuously in the AN of an CheR or CheB enzyme. Here, we have shown that a model with dynamical CheR/CheB binding and unbinding to receptors can reproduce precise adaptation as in the previous AN model [18]. However, CheR/CheB dynamics can both limit precise adaptation and increase noise in complex activity. In addition, we have expanded our results to show robust adaptation, to explain the experimentally observed difference between the responses to aspartate and serine, and to account for the persistence of precise adaptation despite the experimentally observed kinetic variation among methylation sites.
Although both MWC models [11,18–20,32] and Ising-type lattice models [33–35] have been used to represent interactions among receptors, analysis of FRET data provides evidence for strongly coupled MWC complexes [36]. Mello and Tu [20,32] successfully fit the Sourjik and Berg FRET data [10] using an identical MWC model to ours [18,19], but did not include CheR/CheB kinetics. Although Mello and Tu considered methylation-dependent ligand-binding constants Kon/off, fitting results do not require variable Kon/off. CheR and CheB dynamics have been explored in a mixed-receptor Ising-type lattice model by assuming Michaelis-Menten kinetics, with each CheR or CheB only able to methylate or demethylate the bound receptor once before detaching [34,35]. In principle, combining catalysis with unbinding increases precision of adaptation by decreasing the likelihood of fully methylating or fully demethylating a receptor, but enzyme tethering suggests each bound enzyme may catalyze multiple methyl transfers before unbinding.
Within our model, deviations from precise adaptation occur only if CheR/CheB methylation/demethylation rates become dependent on the receptor-methylation level. Small AN size and long dwell times of CheR and CheB cause full methylation or demethylation of neighborhoods, resulting in methylation-dependent rates and failure of precise adaptation. Although the average dwell time of CheR or CheB has not been experimentally determined, the diffusion-limited association rate of protein–protein interactions is on the order of 105–107 M−1s−1 [28,37]. Multiplying by the experimentally determined dissociation constant of 11 μM for CheR [38] yields unbinding rates in the range of 1–100 s−1, well above the CheR unbinding rate
required for precise adaptation (Figure 4). Precise adaptation also requires saturated enzyme kinetics, meaning that each bound CheR or CheB acts at a rate independent of the number of available modification sites. (Unsaturated kinetics would imply decreased demethylation rates at low average methylation levels, and decreased methylation rates at high average methylation levels.) In our model, precise adaptation is robust to differences in dwell times of CheR or CheB on different modification sites, but not, in general, to different rates of CheR or CheB action on these sites, pointing to different dwell times as the explanation for the site-dependent methylation/demethylation rates observed by Chalah and Weis [27].
As with the BL model for a single receptor, our model for a receptor complex robustly yields precise adaptation over a wide range of parameters (Figure 9). One improvement is that our model based on ANs does not require fully methylated receptors to be fully active, or fully demethylated receptors to be fully inactive. Experiments indicate that adaptation time and adapted activity level vary even among genetically identical cells [39]. Consistent with observations by Alon et. al. [26], our model predicts that varying CheR and/or CheB concentrations will lead to different adapted activities (Figure 3) while preserving precise adaptation. The robustness of the essential properties of the network (e.g., sensitivity and precise adaptation) presumably also allows for genetic polymorphisms in the binding and reaction rates of network proteins, making the network robust to evolutionary change.
Within our model, we assume that the rates of modification by CheR and CheB depend directly on complex activity. In fact for precise adaptation, only one enzyme needs to respond directly to complex activity. This is consistent with experiment as CheR rates are affected directly by activity [40], whereas CheB is phosphorylated to an active form by the receptor-regulated kinase CheA [41,42], implying a global feedback mechanism. If, hypothetically, all feedback were global, there would be no direct “return force” on the activity of individual complexes, only an indirect return force on the average complex activity. As a result, sensitivity would be lost as most complexes would drift to nonresponsive methylation levels, becoming either fully active or fully inactive. However, within our model, precise adaptation still occurs if the CheB feedback mechanism is disabled without destroying CheB's demethylating ability as long as direct feedback from complex activity to CheR is maintained. Indeed, experiments mutating the phosphorylation site of CheB demonstrate that CheB phosphorylation is not required for precise adaptation [26], but is important to keep adapted CheY-P levels in the range of motor sensitivity [43].
Our model helps explain the advantage of multiple methylation sites per receptor. First, the number of receptors that a tethered CheR or CheB can modify is constrained by the physical length of the tether. Therefore, to provide enough steps in the ladder of methylation levels to prevent full demethylation or full methylation of neighborhoods (and therefore loss of precise adaptation), the number of modification sites per receptor must be sufficiently large. Second, if the number of methylation sites per receptor were small, then to allow precise adaptation over a large concentration range would require a large change in free energy per methyl group. However, large free-energy steps per methyl group increase the noise in activity (Figure 6), and prevent complexes from operating in the regime of maximal sensitivity.
One longstanding puzzle has been the observed difference in E. coli's chemotactic response to serine and aspartate [9]. Our model explains both the observed broad range of precise adaptation to aspartate/MeAsp and the failure of adaptation at relatively low serine concentrations (Figure 8). Based on receptor in vivo expression levels, complexes contain more Tsr receptors then Tar receptors, so the Tsr receptors act as extra methylation sites and increase the range of precise adaptation to aspartate/MeAsp. As the Tar receptors become fully saturated, the Tsr receptors bind aspartate/MeAsp, thereby also broadening the range of response. In contrast, the high proportion of Tsr receptors amplifies the complex free-energy change due to serine and leads to full methylation of receptors and, therefore, loss of activity, beginning at 0.1 mM serine. The prevalence of Tsr receptors suggests that chemotaxis to low concentrations of serine is biologically important.
For stimulation of low-abundance (minor) receptors, our model predicts a limited range of response. With approximately one minor receptor of each type per complex, there is no amplification of free-energy change, so sensitivity is limited to the off-state ligand affinity
. The range of response is then constrained by the on-state ligand affinity
, unless the range of response is extended though weak binding of ligand to other receptors.
The effect of AN size may be testable experimentally through shortening the flexible tether at the C terminus of receptors while preserving the pentapeptide binding site for CheR and CheB. Decreasing neighborhood size should produce deviations from precise adaptation as the ends of the methylation ladders for ANs are reached (Figure 4A). In addition, the consequence of nonsaturated kinetics may be testable through mutations in CheR/CheB that reduce their affinities for the methyl-modification sites on receptors. Our model predicts global failure of precise adaptation for large deviations from fully saturated kinetics, but even small deviations from full saturation have noticeable consequences near full methylation (Figure 4B).
Experiments demonstrate that precise adaptation is a robust property of the E. coli chemotaxis network [26]. The elegant BL model exhibits robust adaptation through integral feedback control [44], but does not include interactions among receptors. Our model provides a molecular mechanism illustrating how integral feedback control is implemented in the presence of receptor clustering, and highlights the importance of ANs to effectively increase the ladder of methylation levels.
In calculating complex activity, we used the same offset energies for both Tar and Tsr receptors: εr(0) = 1.0; εr(1) = 0.5; εr(2) = 0.0; εr(3) = −0.3; εr(4) = −0.6; εr(5) = −0.85; εr(6) = −1.1; εr(7) = −2.0; and εr(8) = −3.0. Both MeAsp and serine were considered as attractants. For MeAsp (Tar = a and Tsr = s):
,
,
,
. For serine:
,
,
,
. The constants chosen for Tar-MeAsp binding and Tsr-serine binding are approximately consistent with experimental data [10,45]. The high values of
and
for serine indicate that Tar does not bind serine at the concentrations considered (≤2 M). On the other hand, Tsr binds MeAsp at lower concentrations since
. Both MeAsp and serine are attractants, so Kon > Koff. For the demethylation rate, we used the (rounded-off) observed value kB = 0.2 s−1 [46,47], and for the methylation rate we set kR = 0.1 s−1. Since kB = 2kR, this sets an adapted activity of 1/3, assuming that the bound levels of CheR and CheB are the same. Unless otherwise specified, we used the same rates of binding/unbinding for CheR and CheB:
and
, yielding
.
To simulate the dynamics of an MWC complex of receptors, we used an exact stochastic Gillespie algorithm [48]. We assumed that the rates of ligand binding/unbinding and on/off switching of complexes are much faster than the rates of receptor modification and the rates of CheR/CheB binding and unbinding. Therefore, methylation/demethylation and CheR/CheB dynamics were modeled explicitly, whereas the average activity of the complex was calculated using Equation 3. The Gillespie algorithm involves three different steps for the generation of each data point. First, the reaction that occurs is picked randomly, with weighting directly proportional to the individual rates of each event. The possible events are methylation, demethylation, binding of CheR or CheB, and unbinding of CheR or CheB. A receptor cannot have both CheR and CheB bound at the same time. Next, the site of the event is randomly chosen. The time is then incremented by τ = −(log r)/Γ, where r is a random variable picked from a uniform distribution over [0,1], and Γ is the sum of the rates of all possible events.
For each attractant concentration, simulations to determine adapted activity and distributions (Figures 3–8) were averaged over 200 runs of 10,000 Gillespie steps, each following 10,000 steps to allow time for equilibration.
In order to test the robustness of our dynamical model, we randomly varied the parameters and tested the precision of adaptation. Altered systems were obtained by modifying eight parameters (kR, kB,
,
,
,
,
, and
) by factors of kn=1,...,8. Total parameter variation is expressed by
. Each K was randomly chosen as K = 105r, where r is a random variable picked from a uniform distribution over [0,1]. Values of the |log kn| were randomly chosen over [0,1], and were then normalized to yield the correct sum for log K. The sign of each log kn was then chosen with equal probability to be negative or positive. These systems were then subject to a concentration change from 0 mM of ligand to 1 mM. Precision of adaptation was calculated by dividing the adapted activity at 1 mM by the adapted activity at 0 mM.
Simulations to test the effect of the free-energy difference δε per methyl group on the variances in methylation and activity levels (Figure 6) were performed at 10 mM of MeAsp, with the constant free-energy offset ε0 set to yield an average adapted receptor methylation level of 〈m〉 = 4 (i.e., ε0 = −1.0 + 4δε). δε = 0.5 was chosen to approximate the free-energy difference per methyl group used in all other simulations. For increased complex sizes, we used two or three strongly coupled 19-receptor complexes (i.e., all receptors on or off together) to produce complexes of size 38 or 57, respectively, preserving the original AN pattern.
The primary protein accession numbers (in parentheses) from the Swiss-Prot databank (http://www.ebi.ac.uk/swissprot) for the proteins mentioned in the text are as follows: CheA E. coli CHEA_ECOLI (P07363), CheB E. coli CHEB_ECOLI (P07330), CheR E. coli CHER_ECOLI (P07364), CheW E. coli O157 CHEW_ECO57 (P0A966), CheY E. coli O157 CHEY_ECOLI (P0AE67), Tar E. coli MCP2_ECOLI (P07017), and Tsr E. coli MCP1_ECOLI (P02942).
|
10.1371/journal.pntd.0006321 | Complement C1q expression in Erythema nodosum leprosum | Complement C1q is a soluble protein capable of initiating components of the classical pathway in host defence system. In earlier qualitative studies, C1q has been implicated in the pathogenesis of Erythema Nodosum Leprosum (ENL). However, little is known about the role of this complement in ENL reaction. In the present study we described the protein level of C1q production and its gene expression in the peripheral blood and skin biopsies in patients with ENL reaction and lepromatous leprosy (LL) patient controls before and after treatment. Thirty untreated patients with ENL reaction and 30 non-reactional LL patient controls were recruited at ALERT Hospital, Ethiopia. Peripheral blood and skin biopsies were obtained from each patient before and after treatment. The level of circulating C1q in the plasma was determined by enzyme-linked immunosorbent assay. The mRNA expression of the three C1q components, C1qA, C1qB, and C1qC in the peripheral blood and skin biopsies was determined by qPCR. Circulating C1q in the peripheral blood of untreated ENL patients was significantly decreased compared to LL patient controls. Untreated ENL patients had increased C1q gene expression in the peripheral blood compared to LL controls. Similarly, C1qA and C1qC gene expression were substantially increased in the skin biopsies of untreated ENL patients compared to LL controls. However, after treatment none of these genes show significant difference in both groups. In conclusion, while circulating C1q is inversely correlated with active ENL reactions, its gene expression is directly correlated with ENL. The decreased circulating C1q may suggest the utilization of C1q in immune-complex formation in these patients. Therefore, C1q could be a potential diagnostic marker for active ENL reactions as well as for monitoring ENL treatment.
| Erythema nodosum leprosum (ENL) is a painful inflammatory reaction which occurs in patients with lepromatous and borderline lepromatous leprosy. The diagnosis of ENL is mainly based on clinical signs and symptoms and there is no definitive laboratory diagnosis. The causes of ENL are not fully understood but immune-complexes and T-cells activation are among the main factors described by researches. Complement C1q is a protein which initiates the classical complement pathway to form antigen-antibody complex (immune-complexes) which is normally removed before deposition. However, we do not have clear information on the role of C1q in patients with ENL reaction. Therefore, we investigated the production and gene expression of C1q in patients with ENL reaction and in non-reactional LL patient controls before and after corticosteroid treatment of ENL cases. We found that the amount of circulating C-q protein was significantly reduced in untreated ENL patients compared to non-reactional LL patients and healthy controls. On the other hand, the gene expression of C1q in peripheral blood and skin biopsies had been considerably higher in untreated ENL patients compared to LL patients controls. Both C1q protein and gene expression were not significantly different between the two groups after corticosteroid treatment of ENL cases. The decreased C1q protein in untreated ENL may indicate the consumption of C1q in immune-complex formation in these patients and suggests that the involvement of immune-complex in the pathogenesis of ENL. Hence, C1q could be used as a potential candidate marker for ENL diagnose in the future.
| Leprosy is a chronic infectious disease caused by Mycobacterium leprae which infects mainly skin and peripheral nerves [1]. The disease manifests with a spectrum of clinical pictures ranging from the localized tuberculoid leprosy (TT) through borderline forms to the generalized lepromatous leprosy (LL) of the Ridley-Jopling (RJ) classification [2]. Erythema nodosum leprosum (ENL) is an immune-mediated inflammatory complication affecting about 50% of patients with lepromatous leprosy (LL) and 10% of borderline lepromatous (BL) patients [3–5]. ENL first occurs as acute but may pass in to a chronic phase or can be a recurrent [6]. It involves multiple organs and manifests as systemic illness [7]. The occurrence of crops of tender erythematous skin lesion is the clinical diagnostic of ENL [8]. Histologically, the infiltrations of neutrophils throughout the dermis and subcutis is the defining characteristic of ENL[9]. However, not all clinically confirmed ENL cases have neutrophilic infiltration in the lesions.
ENL is mainly diagnosed clinically and if the facility is available supported by histological findings. Nevertheless, no clinical or laboratory tests accurately predict who is most likely to develop ENL reaction or when it might occur.
C1q is one of the several candidate markers proposed for ENL diagnosis in earlier studies [10, 11]. However, little is known about the possible role of complement in ENL reaction. Recently the complement system have been perceived as a central constituent of innate immunity, defending the host against pathogens, coordinating various events during inflammation, and bridging innate and adaptive immune responses [12]. Complement system not only protect the host against infection but also contribute to the amplification of inflammation if activated in excess or inappropriately controlled [13].
C1q is a 460-kDa protein made up of 18-subunit glycoprotein consisting of 3 subunits: A, B and C which are coded by C1qA, C1qB and C1qC genes respectively. It has been shown that C1q is assembled in a 1:1:1 ratio from these three different subunits. The three human C1q genes are closely located on chromosome 1 and arranged C1qA-C1qC-C1qB orders [14]. It is believed that C1q is mainly produced by macrophages [15]. It has been reported that IFN-γ and IL-6 promote C1q production by macrophages while IL-1 inhibits its production by macrophages [16]. C1q initiate the classical complement pathway. However, recently it has been shown that C1q is also involved in non-complement activation such as modulating dendritic cell maturation, pro-inflammatory cytokine production and T- and B-cell responses [17].
Increased C1q production in patients with active pulmonary tuberculosis (TB) compared to latent TB has recently been reported indicating the potential use of C1q as biomarker to discriminate between patients with active TB cases from latent TB [16]. C1q has also been implicated in several immune-complex disorders such as acute glomerulonephritis [18] and acute systemic lupus erythematosus (SLE) [19, 20] and rheumatoid arthritis [21].
Previous qualitative studies looked at C1q in the immunopathology of ENL and reported that serum samples from leprosy with or without reaction confer similar results[10, 11, 22]. However, serum samples from leprosy with or without reaction showed positive test. A recent microarrays study in the PBMCs of 3 ENL and 3 T1R patients has reported increased expression of the classic complement pathway particularly complement C1qA, B, C and the complement receptor C3AR1 and C5AR1 [23]. Increased fluorescent intensity of C1q in skin lesions of these patients has also been reported in the same study. Therefore, the association of C1q with the pathophysiology of ENL need to be explored in large number of ENL patient cohort. In the present study, the circulating C1q in the peripheral blood of patients with ENL and LL controls before and after prednisolone treatment was quantified using a special C1q-ELISA. The gene expression of C1q (C1qA, C1qB and C1qC) was quantified in peripheral blood and skin biopsies of these patients by quantitative polymerase chain reaction (RT-qPCR) before and after prednisolone treatment of ENL patients.
Informed written consent for blood and skin biopsies were obtained from patients following approval of the study by the Institutional Ethical Committee of London School of Hygiene and Tropical Medicine, UK, (#6391), AHRI/ALERT Ethics Review Committee, Ethiopia (P032/12) and the National Research Ethics Review Committee, Ethiopia (#310/450/06). Children under 18 years old and minor or vulnerable groups have been excluded from the study. All patient data analyzed and reported anonymously.
We recruited 30 untreated patients with ENL reaction and 30 matched non reactional LL patient controls. All patients recruited into this study were attending the ALERT Hospital, Addis Ababa, Ethiopia. The patients were classified on the leprosy spectrum based on the Ridley-Jopling (RJ) classification schemes [2]. ENL was clinically diagnosed when a patient with BL or LL leprosy had painful crops of tender cutaneous erythematous skin lesions [4]. New ENL was defined as the occurrence of ENL for the first time in a patient with LL or BL. Lepromatous leprosy was clinically diagnosed when a patient had widely disseminated nodular lesions with ill-defined borders and BI above 2 [24]. Patients with ENL were treated according to the World Health Organization (WHO) treatment guideline with steroids that initially consisted of 40mg oral prednisolone daily and the dose was tapered by 5mg every fortnight for 24 weeks. All patients were received WHO-recommended leprosy multidrug treatment (MDT). We also included 15 apparently healthy volunteers and used for C1q ELISA result comparison only.
Ten micro-litter of venous blood was collected into sterile BD heparinised vacutainer tubes (BD, Franklin, Lakes, NJ, USA) before and after prednisolone treatment of ENL cases on week 24 from each patient and healthy controls. Plasma was separated and used for ELISA. In addition, 2mL of blood was collected into PAXgene Blood RNA Tubes (PreAnalytix, GmbH, Switzerland) before, and after prednisolone treatment for mRNA isolation and stored at -80°C. Six-mm punch biopsy was taken from each patient before and after prednisolone treatment into a to a Nunc tube containing 1mL RNAlater solution (Thermo-Fisher Scientific) and was kept at -20°C for 48 hours and then transferred to -80 oC freezer. The ENL and LL lesions for biopsy sample were identified and marked by a dermatologist and then biopsy samples were taken from the marked area by trained research nurses under supervision. For ethical reason we didn’t collect biopsy samples from healthy controls.
For quantitative detection of human C1q in the plasma samples of patients with ENL, non reactional LL and healthy controls, we used human C1q platinum ELISA (ready-to-use sandwich ELISA) with a sensitivity of 0.08ng/mL purchased from eBioscience (Affymetrix, eBioscience, UK). The procedure is briefly described as follow as: The microwell plate coated with monoclonal antibody to human C1q was aspirated and washed twice with wash buffer. Then 100μl of serially diluted C1q standard was pipetted to the first two columns of the microwell strips. To the remaining strips pre-diluted to 1:1000 plasma samples (100μl/well) were added and to the last two wells a blank (assay buffer) was added as a negative control. The plate was sealed and incubated at 22°C for 2hrs on a microplate shaker set at 400 revolutions per minute (rpm). After 2 hours incubation, the plate was washed 6 times with wash buffer and tapped on the absorbent pad. To each well, 100μl of biotin-conjugated anti-human C1q antibody was added and incubated at 22°C for 1hr on a microplate shaker set at 400rpm. Then the microplate was washed as described above and followed by the addition of 100μl Streptavidin-HRP to all wells. The plate was sealed and incubated at 22°C for 1hr on a microplate shaker as described. After six washes 100μl of TMB substrate solution was added to all wells and incubated at 22°C for 30 minutes in the dark. The colour development on the plate was monitored and the substrate reaction was stopped by pipetting 100μl of stop solution (1N H3PO4). The optical density (OD) at 450nm was measured using an ELISA plate reader (Microplate reader; Bio-Rad, Richmond, CA). A curve fit was applied to the standard curve according to the manufacturer’s manual using Microplate Manager 6 Software (Bio-Rad, Richmond, CA) and the unknown concentration of C1q in each sample was extrapolated from these standard curves.
Isolation of RNA from peripheral blood and skin biopsies stored in RNAlater (Ambion, Austin, Texas) was performed using PAXgene Blood RNA Kit and RNeasy Fibrous Tissue Kit (QIAGEN Crawley, West Sussex, United Kingdom) respectively according to the manufacturer’s protocol. DNase I (QIAGEN) was included for all RNA preparations for DNA digestion. RNA yield was determined using a NanoDrop 2000, spectrophotometer (Thermo Scientific, Epsom, UK) and integrity was checked by agarose gel electrophoresis. For all samples Complementary DNA (cDNA was synthesized on the same day to avoid the risk of RNA degrades during storage. cDNA was synthesised from RNA (200 ng/reaction mixture) using High Capacity cDNA Reverse Transcriptase Kit (AB Applied Biosystems, UK). Reactions were incubated in an ABI9700 Programmable Thermal Cycler (Applied Biosystems, Foster City, California) for 10 minutes at 25°C followed by 120 minutes at 37°C and 5 minutes at 85°C and then cooling to 4°C.
Primers between 20–24 nucleotides in length were designed across intron/exon boundaries on mRNA sequence obtained from the Nation Centre for Biotechnology Information database (NCBI) to give a product between 100-500bp. All primer sequences were blasted on the NCBI data bank to confirm their specificity. Custom synthesis of oligonucleotide primers was performed by Sigma-life science and provided in desalted form. The nucleotide sequences of the forward and reverse primers, respectively, used in this study were as follows: for C1-qA, 5'-ATGGTGACCGAGGACTTGTG-3' and 5'-GTCCTTGATGTTTCCTGGGC-3'; for C1-qB, 5'-CAGGTTGAAATCAGCATTGCC-3' and 5'-CTGTGTCAGACGCCTCCTTTC-3'; for C1-qC, 5'-AAGGATGGGTACGACGGACTG-3' and 5'- TTTCTGCTTGTATCTGCCCTC-3' and for human acidic ribosomal protein (HuPO) house-keeping gene: 5'-GGACTCGTTTGTACCCGTTG-3' and 5'-GGACTCGTTTGTA CC CG TTG-3'.
Real-time quantitative PCR for all genes was performed on the Rotor-Gene 3000 programmable thermal cycler (Corbett Life Science, Qiagen, Crawley, UK) using Roter-gene SYBR Green PCR Kit (Qiagen, Crawley, UK). The Rotor-Gene conditions were set as follows: Initial activation step (polymerase activation) was achieved by incubating at 95°C for 15 minutes, 40 cycles of denaturation at 95°C for 5 seconds, annealing at 60°C for 10 seconds, extension at 72°C for 20 seconds and fluorescence acquisition for 5 seconds at 72°C. The primer-dimer formation was checked by melting curve analysis. Melting point data were obtained by increasing the temperature from 50°C to 99°C by 1°C on each step. The interval between increases in temperature was 30 seconds for the first step and then 5 seconds for subsequent steps. An assay control was included from mRNA extraction to the amplification steps. For mRNA extraction, one assay control per batch was used. The assay control included all buffers except the sample and was processed under identical conditions with the samples. The same assay control was used during cDNA synthesis and real-time quantitative PCR.
The relative gene expression (fold change) was analyzed by using the 2-ΔΔ CT method (Cikos et al., 2007). The difference in threshold number for the amplification of the target gene (ΔCT) was obtained by subtracting the CT of the target gene from the CT of the control gene. To compare the target gene expression in patients with ENL and LL controls, ΔΔCT was obtained by subtracting the ΔCT of LL patient control from the ΔCT of the patient with ENL. Then, the fold change (FC) was obtained by using the formula FC = 2-ΔΔ CT. Similarly, for the comparison of the relative target gene expression within ENL group before and after treatment, ΔΔCT was obtained by ΔCT (after) minus ΔCT (before). Then the fold change for the target gene expression in untreated ENL was given by 2-ΔΔ CT.
Thirty patients with ENL reaction and 30 LL patient controls without ENL reaction were recruited between December 2013 and October 2015. The male to female ratio was 2:1 with a median age of 27.5 [range: 18–56] years in patients with ENL and 3:1 with a median age of 25.0 [range: 18–60] years in patients with non-reactional LL controls. All ENL patients were untreated with corticosteroid before recruitment. At time of recruitment, 20 ENL patients were previously untreated with MDT, 21 were on MDT and 5 were completed MDT treatment. Twenty non-reactional LL patients were about to start MDT, 7 were on MDT and 4 were completed MDT at recruitment. From 15 apparently healthy controls, 9 were male and 6 were female with median age of 26.5 [range: 20–57] years.
Patients with ENL reactions had significantly lower circulating C1q (11698pg/mL ± 618.3) compared to LL patient controls (21059pg/mL ± 2382.0) and health controls (18448pg/mL ± 1161) before treatment (P≤ 0.0001) (Fig 1A). However, the amount of circulating C1q considerably increased in ENL patient to 22287pg/mL ± 2154 after treatment while it was only slightly increased to 23721pg/mL ± 1886 in LL patient controls. Circulating C1-q production was similar in patients with ENL and LL controls after treatment (Fig 1B). Similar, C1q production was similar in patients with ENL reactions (after treatment) and healthy controls (P≤ 0.0001). Although, C1q production was slightly higher (but not significantly different) in LL patients controls compared to health controls before and after treatment, (Fig 1B).
Comparison within patients with ENL reactions before and after treatment has shown that circulating C1q production in the peripheral blood of ENL patients was considerably increased after prednisolone treatment and the difference significantly different (P≤ 0.0001) (Fig 1C). This indicates that patients with ENL reactions do not have C1q genetic defect since the level of C1q production has increased after the ENL reactions subside.
We quantified the mRNA gene expressions for the three genes (C1qA, C1qB, C1qC) in the blood and skin biopsies obtained from patients with ENL and LL controls before and after treatment.
The gene expression for C1qA (FC = 2.56), C1qB (FC = 2.63) and C1qC (FC = 4.55) in the peripheral blood of untreated patients with ENL reaction were significantly increased compared to the corresponding non-reactional LL patient controls (P≤ 0.05). In skin biopsies, increased expression of C1qA (FC = 6.45) and C1qC (FC = 7.46) was associated with untreated active patients with ENL reactions compared to LL patient controls (P≤ 0.05). However, unlike in the peripheral blood, the gene expression for C1qB in untreated patients with ENL was not statistically significantly different compared to LL patient controls. None of these genes show significant difference after treatment in both groups (Table 1).
C1q gene expression in the peripheral blood and skin biopsies were also compared within ENL group before and after prednisolone treatment. The gene expression for all the three genes in the peripheral blood was not statistically significantly different before and after treatment within the ENL group. The gene expression in the skin biopsies for C1qA and C1qB did not change before and after treatment within ENL group. Interestingly, the C1qC gene expression in the skin biopsies remarkably decreased (FC = 0.13) after treatment within ENL group (P < 0.0001) (Table 2).
The amount of circulating C1q in the plasma of patients with ENL and LL was quantified using a C1q ELISA. The gene expression of C1q was also quantified in blood and skin biopsy samples of these patients before and after treatment. Patients with ENL had significantly lower circulating concentrations of C1q than LL controls before treatment. However, after treatment, the amount of circulating C1q was not significantly different in both groups. This is the first work to determine circulating C1q production using ELISA in leprosy patients in leprosy patients with and without ENL. Therefore, comparison with previous studies would not have been possible. one earlier qualitative study had reported that lower C1q binding activity in the sera of untreated ENL patients compared to treated ENL patients[25].
The decreased frequency of circulating C1q in the sera of untreated patients with ENL compared to healthy controls as well as non reactional LL patient controls could be due to its utilisation by the antigen-antibody complex formation in ENL reaction. Similarly, a decreased serum complement C1q levels have been observed in other immune complex disorders such as acute glomerulonephritis (Lange et al., 1960, Lewis et al., 1971, Ohi and Tamano, 2001) and acute systemic lupus erythematosus (SLE) (Baatrup et al., 1984, Grevink et al., 2005). However, since the pathophysiology of autoimmune disorders and infectious diseases like ENL reactions are different, further study is required.
The gene expression level of C1q A, B and C in the blood and C1q A, and C in the skin biopsies from patients with ENL reactions were significantly higher than in LL controls before treatment. However, after treatment, none of these genes were found to be significantly different between the two groups. It is interesting to note that while the amount of circulating C1q protein is decreased in active ENL, its gene expression is increased in untreated ENL patients compared to the corresponding LL patient controls. This phenomenon may be explained by the utilization of C1q in immune-complex formation in active ENL patients. Although the initiation of ENL is not precisely known, several studies have reported the involvement of immune-complexes in the pathophysiology of ENL reaction. Immune-complexes activate the classical complement pathway that opsonise or coat antigen-antibody complexes with complement molecules which facilitates the clearance of immune-complexes by the macrophages [26]. Hence, by maintaining immune-complexes in solution, the complements allow clearance of the complexes from their site of formation, minimizing local inflammatory consequences [26]. However, in the event of immune-complexes deposition complements are consumed until [27].
In the comparison within ENL group, it has been found that C1qC gene expression was significantly increased in the skin biopsies of untreated ENL and subsequently decreased after ENL Treatment. The increased C1qC gene expiration in untreated active ENL may be associated with the immune activation in ENL reactions. Previously we have shown that increased T-cell activation in untreated active ENL patient compared to non-reactional LL controls [27]. It has also been reported that the inflammatory cytokines such as IFN-γ and IL-6 increase C1q production by macrophages [28]. Several studies indicated that C1q can modulate dendritic cell maturation, pro-inflammatory cytokine production, and T- and B-cell responses in addition to its classical function to initiate complement activation [29–31]. Recently, increased circulating C1q and C1qC gene expression in patients with active tuberculosis compared to healthy controls and individuals with latent TB infection has been reported [16] indicating its potential use as a biomarker to discriminate active TB from latent TB cases. Therefore, in the same scenario, C1qC gene expression could be used as potential diagnostic marker for ENL, which necessities further investigation particularly in the lesions.
|
10.1371/journal.pcbi.1003061 | Estimation of Vaccine Efficacy and Critical Vaccination Coverage in Partially Observed Outbreaks | Classical approaches to estimate vaccine efficacy are based on the assumption that a person's risk of infection does not depend on the infection status of others. This assumption is untenable for infectious disease data where such dependencies abound. We present a novel approach to estimating vaccine efficacy in a Bayesian framework using disease transmission models. The methodology is applied to outbreaks of mumps in primary schools in the Netherlands. The total study population consisted of 2,493 children in ten primary schools, of which 510 (20%) were known to have been infected, and 832 (33%) had unknown infection status. The apparent vaccination coverage ranged from 12% to 93%, and the apparent infection attack rate varied from 1% to 76%. Our analyses show that vaccination reduces the probability of infection per contact substantially but not perfectly ( = 0.933; 95CrI: 0.908–0.954). Mumps virus appears to be moderately transmissible in the school setting, with each case yielding an estimated 2.5 secondary cases in an unvaccinated population ( = 2.49; 95%CrI: 2.36–2.63), resulting in moderate estimates of the critical vaccination coverage (64.2%; 95%CrI: 61.7–66.7%). The indirect benefits of vaccination are highest in populations with vaccination coverage just below the critical vaccination coverage. In these populations, it is estimated that almost two infections can be prevented per vaccination. We discuss the implications for the optimal control of mumps in heterogeneously vaccinated populations.
| Less than two decades ago, it was generally believed that in developed countries infectious diseases such as measles, mumps, and pertussis were under firm control via vaccination. Nowadays, it is increasingly recognized that this picture has been overly optimistic. A central question is whether recurrent disease outbreaks are caused by vaccination coverage having dropped below safe levels, or by vaccines having become less effective. To answer this question, the authors study outbreaks of mumps in primary schools in the Netherlands. Using disease transmission models, the authors estimate vaccine efficacy and the critical vaccination coverage needed to prevent large outbreaks. The analyses show that the vaccine has been highly effective in preventing infection, but that vaccination coverage has been insufficient in some schools. The authors argue that catch-up vaccination campaigns aimed at populations with intermediate vaccination coverage will be most efficient, as these would maximize the (direct and indirect) benefits of vaccination.
| Mass vaccination programs for childhood diseases have been highly successful in reducing the incidence and public health impact of the targeted diseases. Nevertheless, with the exception of smallpox, eradication has not been achieved, and outbreaks continue to occur even in highly vaccinated populations [1]–[4]. A prominent example is that of mumps, which has re-emerged in the past decade in highly vaccinated populations throughout the world [5]–[7]. The question arises as to whether this re-emergence is due to current vaccines becoming less effective, or to reduced vaccine coverage which allows the virus to spread in partially vaccinated populations [8].
In the Netherlands, large outbreaks of mumps genotypes D and G have occurred in recent years [9]–[11]. Since 1987, a combined MMR (measles-mumps-rubella) vaccine containing live attenuated virus is routinely given at 14 months and 9 years of age. Vaccination coverage has been high ever since introduction of the vaccine in 1987 (90–95%). Nevertheless, there are municipalities in which vaccination coverage is substantially lower [12], [13].
To determine whether the outbreaks of mumps are the result of low vaccination coverage or insufficient protection conferred by the vaccine, we estimate vaccine efficacy using outbreak data from ten primary schools in the Netherlands [9], [11]. The total number of children included in our study is 2,493, of whom 510 had a reported mumps infection. Vaccination coverage in these schools ranged from 12%–93%, and infection attack rates ranged from 4% to 76%, with highest attack rates occurring in schools with the lowest vaccination coverage and lowest attack rates in schools with high vaccination coverage (Table 1). Notably, the attack rates in unvaccinated individuals varied from more than 80% in schools with low vaccination coverage (<15%) to lower than 25% in schools with high vaccination coverage (≥75%), indicating substantial differences in the infection pressure between schools.
Classical methods to estimate vaccine efficacy from outbreak data compare the infection attack rates in the vaccinated versus unvaccinated groups (i.e. the cohort method) [14], [15]. This method, however, has significant drawbacks. First, it is not straightforward to take account of missing data on vaccination and infection status. This is unfortunate as outbreak data are almost never complete, and judicious choices will have to be made to avoid introducing systematic bias in the parameter estimates. Even more importantly, the cohort method fails to acknowledge that the probability of infection of an individual is dependent on the number of infections in the population, i.e. on the infection status of others.
To take account of the dependencies between individuals that arise naturally in infectious disease outbreaks we base the statistical analyses on a Bayesian inferential framework using infectious disease transmission models. In this framework, missing vaccination and infection information is imputed in a consistent manner, thereby making efficient use of the available information, and enabling precise estimation of vaccine efficacy and the critical vaccination coverage needed to prevent epidemic outbreaks [16], [17]. The basis of our statistical analyses is the contact process that specifies how often and with which person-types each person makes infectious contacts, i.e. contacts that are sufficient for transmission if the sender is infected and the receiver as yet uninfected [18]–[20]. The contact process specifies a directed graph, of which the connected component with the initial infective as the root determines which individuals are ultimately infected. Estimation of the epidemiological parameters (basic reproduction number, vaccine efficacy) is based on the likelihood of directed graphs that are compatible with the data.
The analyses reveal that mumps vaccine effectively prevents infection, and that herd immunity against mumps is achieved with moderate vaccination coverages. We argue that resource-limited catch-up vaccination efforts should be focused at communities with intermediate vaccination coverages, thereby maximizing both the direct and indirect benefits of vaccination.
Our baseline scenario assumes a common transmissibility and vaccine efficacy across schools. The analysis indicates that mumps is moderately transmissible ( = 2.49; 95%CrI: 2.36–2.63), and that the vaccine reduces the probability of transmission by more than 90% per contact that would have resulted in transmission to an unvaccinated person ( = 0.933; 95CrI: 0.908–0.954)(Figure 1). The differences between the apparent and estimated vaccination coverages and attack rates are small (<2% and <5%, respectively; Tables 1–2).
We use estimates of transmissibility and vaccine efficacy to obtain estimates of the critical vaccination coverage. The analyses yield an estimated critical vaccination coverage of 0.642 (95%CrI: 0.617–0.666), indicating that herd immunity in the school setting can be obtained with moderate vaccination coverages. Estimates of transmissibility and vaccine efficacy are used to obtain an estimate of the number of infections prevented per vaccination. This number is highest for vaccination coverages just below the critical vaccination coverage, as at these values the slope of attack rate versus vaccination coverage is steepest (Figure 2). The number of infections prevented per vaccination near the threshold coverage is well approximated (using a Taylor series expansion) by. Hence, it is expected that the (direct and indirect) benefits of vaccination are such that infections can be prevented per vaccination if the initial vaccination coverage is just below the threshold value, which is estimated by .
Schools in our study population span a large range of possible vaccination coverages, and it is of interest to evaluate the consistency of the estimates of vaccine efficacy and pathogen transmissibility. Figure 2 shows the relation between vaccination coverage and infection attack rate in the ten schools, together with the theoretical relation between vaccination coverage and attack rate in a large population, and simulations of a finite population. Overall, the correspondence between the observed and simulated data is excellent for schools with low vaccination coverage and high attack rates, while there is a tendency for higher attack rates than expected in schools with high vaccination coverage and a small number of infections.
To investigate the information contained in the data by school we perform analyses in which each school is equipped with its own transmissibility and vaccine efficacy. It appears that precise estimates of transmissibility and vaccine efficacy can be obtained in schools with high attack rates (schools 1–4), but not in schools with only a handful of infections (schools 7–10). In fact, in schools with less than 10 confirmed infections credible intervals of the reproduction number range from well below 1 to more than 3, while vaccine efficacy estimates can range from less than 0.20 (schools 8–10) to almost 1 (schools 7–10; Table 3, Figure 3). Further, the analyses show that in schools with high attack rates (schools 1–4) the parameter estimates are quite close to those of the baseline scenario, indicating that estimates of transmissibility and vaccine efficacy in the baseline scenario are dominated by schools with large numbers of infections and low vaccination coverages.
In comparison with our estimates of vaccine efficacy as the reduction in the probability of infection (Table 3), estimates of vaccine efficacy by the cohort method tend to be somewhat lower in schools with low vaccination coverage and high infection attack rates (schools 1–4; Table 4, Table S3). Moreover, in these schools credible intervals tend to be slightly broader when using the cohort method. The most conspicuous difference, however, is that in populations with high vaccination coverage (schools 7–10), vaccine efficacy is sometimes estimated with fair precision when using the cohort method, even though the number of infections is very small (≤6).
Our analyses have shown that mumps is moderately transmissible in the setting of primary schools, and that the vaccine used in these populations is highly effective in preventing infection. These results are largely in line with earlier studies [5], [6], but contrast with a recent study that suggested that outbreaks of mumps in populations with large-scale vaccination programs may be due to the vaccine having become less effective in preventing infection [4]. The younger average age of our study population and the fact that these outbreaks have been caused by viruses of different genotypes (genotype D versus genotype G) may help explain these contrasting findings. Since genotype D viruses are genetically distant from the current vaccine virus (Jeryl Lynn strain, genotype A) our results indicate that the Jeryl Lynn-based vaccine is highly effective in curbing transmission to vaccinated persons, even if genetic differences between the vaccine and outbreaks strains are substantial [8].
Estimates of the transmissbility of mumps are most precise in schools 1–4, i.e. in schools with low vaccination coverage and large numbers of infections. In these schools, the basic reproduction number is estimated at 2.5, 2.3, 2.8, and 2.5, with credible intervals ranging from 1.9 to 3.2. Vaccine efficacy, on the other hand, is estimated most precisely in schools 1, 3, 4, and 5 (Table 3, Figure 3). In these schools, estimates of vaccine efficacy are 0.97, 0.99, 0.94, and 0.98, with credible intervals ranging from 0.84 to 1. These schools have low vaccination coverage and high levels of exposure (i.e. high attack rates) but still more than 30 vaccinated persons. In school 2 the exposure level has been high but the number of vaccinated persons is too small for precise estimation of vaccine efficacy. In schools with high coverage, vaccine efficacy cannot be estimated with any precision, as in these schools it is uncertain whether escape from infection is caused by the vaccine or by a lack of exposure.
The schools included in this study differ greatly with respect to vaccination coverages (range: 12%–93%) and infection attack rates (range: 4%–76%). Nevertheless, estimates of vaccine efficacy are remarkably consistent across schools (Table 3, Figure 3). In fact, only in schools with just a handful of infections (≤6) (schools 8–10) does the estimated vaccine efficacy drop below 0.88. In these schools, credible intervals of vaccine efficacy are wide, and estimates are less determined by the information contained in the data than by the prior distribution of vaccine efficacy. This is also the reason that estimates of vaccine efficacy in the baseline scenario are dominated by schools with low vaccination coverages and high attack rates, as these schools contain much more information than schools with high vaccination coverages and low infection attack rates (Figure 2).
Schools in our study were included based on confirmed mumps infections. It is therefore possible that large outbreaks are more likely to be detected and included than small outbreaks. In other words, it is conceivable that the inclusion process systematically favours inclusion of schools with uncharacteristically high attack rates, thereby leading to selection bias. For schools with low vaccination coverage (and high attack rates) this is arguably not a problem as variation in outbreak sizes is expected to be minor, given the sizes of the schools included (Figure 2). For schools with high vaccination coverage, however, selection bias may well have played a role, and may explain the relatively high attack rates in some of these schools (school 6 and to a lesser extend schools 9–10) (Figure 2). Fortunately, one could argue that our statistical methodology provides a natural weighting of schools, in which schools with small number of infections have lower weight than schools with high number of infections. If specific details were available on the inclusion process, one could envisage extension of the analyses in which the selection process is modelled explicitly. This, however, would introduce more model options, additional parameters to be estimated, and would certainly lead to a more complicated analysis.
We have assumed throughout that infections outside the school played a marginal role. Again, this assumption is probably less problematic in schools with low vaccination coverage and high infection attack rates than in schools with high vaccination coverage and lower attack rates, as variation in the expected number of infections is expected to be small in schools with low vaccination coverage. Moreover, there was no sustained community transmission during the study period, suggesting that the impact of infection outside the schools may have been small. Nevertheless, it would be interesting to extend the current analyses, e.g., along the lines of [21], [22] by inclusion of other major transmission settings.
Classical estimates of mumps transmissibility have been based on the mean age at infection in the pre-vaccination era ([23] and references therein), or on seroprevalence data from the pre-vaccination era [24], [25]. These analyses yielded estimates of the basic reproduction number in fully unvaccinated populations that are substantially higher (∼7–20) than our estimates (∼2–3). It should be noted that these population-based estimates cannot directly be translated to our school-based estimates. Still, should those early estimates be indicative of the current transmissibility of mumps at the population level, then not only are schools an important transmission route but other settings also have the potential to contribute significantly to overall transmission. Again, to assess the contribution of different settings to the overall transmission dynamics, it would be desirable to extend the current studies beyond the school setting, by including household information and, in the specific case of this study, information on the churches attended by the participants [9]. This, however, is only possible if detailed information were available on these settings, not only with respect to their composition but also with respect to vaccination and infection status of a sizeable part of the population.
Vaccine efficacy and transmissibility together determine the critical vaccination coverage needed to prevent epidemic outbreaks. In our study, estimates of the critical vaccination coverage are 64% (95%CrI: 62%–67%) in the baseline scenario, and range from 63% (95%CrI: 58%–68%; school 1) to 76% (95%CrI: 66%–83%; school 6) in schools with more than 10 confirmed infections (schools 1–6). This indicates that the critical vaccination coverage does not need to be as high as suggested by early population-based estimates, which are in the range of 86%–95%.
In none of the analyses presented here have we made a distinction between children who had been vaccinated once and those that had been vaccinated twice. This was done because preliminary analyses and previous results [9], [11] could not find any evidence for differences in vaccine efficacy between the two groups. In view of the data this is not unexpected, as the total number of infections in vaccinated children was small, and as attack rates in the two subpopulations were identical (15 infections among the 582 children who had been vaccinated once; 10 infections among the 370 who had been vaccinated twice). The fact that attack rates were identical is somewhat surprising, as one could have expected more infections in the group that had been vaccinated only once, more than five years ago. For completeness, we have presented the full data in Table S2.
Further, in our analyses we assume that the vaccine works by reducing the probability of transmission (i.e. we assume a leaky vaccine), rather than by providing all-or-nothing immunity. This was done for simplicity, and since the current data do not allow us to distinguish between the different workings of the vaccine. If additional data was available, e.g., on the pre-outbreak antibody titres, one could consider extension of the method by using pre-outbreak antibody titres as an indicator for the ‘level of immunity’, and use this indicator to estimate how the level of pre-existing immunity relates to the probability of infection. In most situations, however, such information will be hard to get, as this would necessitate a large prospective study.
Our definition of vaccine efficacy has a clear-cut biological interpretation (reduction of the probability of infection per contact). This makes it possible to meaningfully average over populations with varying vaccination coverages and exposure levels, and also to extrapolate beyond the study population. This contrasts with traditional estimates of vaccine efficacy that are based on a comparison of attack rates in vaccinated and unvaccinated individuals (the cohort method), or that simply use the vaccination status of the infected individuals together with the population vaccination coverage (the screening method) [14], [26]. Vaccine efficacy estimated by these methods lack a clear biological interpretation, and in essence assumes that a person's risk of infection is independent of whether or not others in the population are infected. This makes interpretation of the estimates problematic, and forbids estimation of the critical vaccination coverage [15], [27]–[29].
Even though our definition of vaccine efficacy differs fundamentally from vaccine efficacy measured by the cohort method, the results are quantitatively in fair agreement with traditional estimates, especially in populations with low vaccination coverage and large number of infections (Table 3 versus Table 4 and Table S3). In schools with high vaccination coverage and small numbers of infections the reverse tends to be true, and estimates of vaccine efficacy generally are both higher and more precise when using the cohort method. For instance, in school 10 there are 6 confirmed infections, and vaccine efficacy is poorly estimated in our analysis (95%CrI: 0.16–0.96) but with fair precision by the cohort method (95%CrI: 0.68–0.98). This is arguably an artefact of the latter method's assumption that all 139 uninfected vaccinated persons have been exposed to an infected person, thereby artificially increasing the precision of the estimates of vaccine efficacy.
Our results point to strategies to efficiently allocate catch-up vaccination efforts in heterogeneously vaccinated populations. No additional vaccination is needed in schools with high vaccination coverage (>75%, say) as these are already protected against epidemic outbreaks affecting a large fraction of students. Similarly, allocating vaccines to schools with low vaccination coverage (<50%, say) is inefficient as it does not markedly reduce the probability of infection for those who are not vaccinated, i.e. the indirect benefits of vaccination are small in these populations. Our analyses suggest that vaccination of populations in the range between these two extremes is most efficient, and that in these populations a single vaccination can potentially prevent almost two infections. Of course, in practice other considerations, for instance on ethical issues, communication, and cost-effectiveness would also come into play.
In the Netherlands, several large outbreaks of mumps virus (genotype D) occurred in 2007–2009. We collected data from children attending primary schools with evidence of mumps virus transmission (report of at least one laboratory confirmed mumps case or more than one clinical mumps case) [9], [11]. Children's parents were asked to fill out a questionnaire asking for information on the child's vaccination status and occurrence of mumps. Individual data on vaccination status were also retrieved from the national Dutch vaccination register. When these were not available, we used the self-reported vaccination status (vaccinated/unvaccinated). Children who were vaccinated more than twice (one case), and who were reported to have had mumps before September 2007 (three cases) were excluded. The study was approved by the medical ethics committee of the University Medical Centre Utrecht and the Radboud University Nijmegen Medical Centre. The data are presented in Table 1 and Tables S1, S2.
To explore the correspondence between the parameter estimates with the data, we simulated outbreaks in schools of size 200 using the digraph construction described above. To prevent early extinction we introduced three infectious persons with random vaccination status in each simulation. For each vaccination composition, we generated 5,000 random digraphs with the values of the basic reproduction number and vaccine efficacy sampled without replacement from the posterior distribution. Subsequently, for each graph we calculated the attack rate among those that were initially susceptible, and present the median and 2.5% and 97.5% percentiles of the resulting distributions (the black line and grey area in Figure 2).
To compare our results with estimates of vaccine efficacy using the cohort method [14], we have calculated vaccine efficacy as 1 minus the relative risk of infection in vaccinated versus unvaccinated persons. In these analyses only information of persons with known vaccination and infection status was taken into account (Table S1). As in the above we employ a Bayesian framework in which the probabilities in the unvaccinated and vaccinated groups are assigned uniform prior distributions, yielding beta-binomial posterior distributions for the infection probabilities. Estimates are obtained using Markov chain Monte Carlo (MCMC) methods, specifically by taking a thinned sample of 10,000 from a converged chain of length 500,000. Table S3 reports classical (frequentist) estimates of vaccine efficacy using the cohort method [14].
|
10.1371/journal.ppat.1006709 | Experience-dependent olfactory behaviors of the parasitic nematode Heligmosomoides polygyrus | Parasitic nematodes of humans and livestock cause extensive disease and economic loss worldwide. Many parasitic nematodes infect hosts as third-stage larvae, called iL3s. iL3s vary in their infection route: some infect by skin penetration, others by passive ingestion. Skin-penetrating iL3s actively search for hosts using host-emitted olfactory cues, but the extent to which passively ingested iL3s respond to olfactory cues was largely unknown. Here, we examined the olfactory behaviors of the passively ingested murine gastrointestinal parasite Heligmosomoides polygyrus. H. polygyrus iL3s were thought to reside primarily on mouse feces, and infect when mice consume feces containing iL3s. However, iL3s can also adhere to mouse fur and infect orally during grooming. Here, we show that H. polygyrus iL3s are highly active and show robust attraction to host feces. Despite their attraction to feces, many iL3s migrate off feces to engage in environmental navigation. In addition, H. polygyrus iL3s are attracted to mammalian skin odorants, suggesting that they migrate toward hosts. The olfactory preferences of H. polygyrus are flexible: some odorants are repulsive for iL3s maintained on feces but attractive for iL3s maintained off feces. Experience-dependent modulation of olfactory behavior occurs over the course of days and is mediated by environmental carbon dioxide (CO2) levels. Similar experience-dependent olfactory plasticity occurs in the passively ingested ruminant-parasitic nematode Haemonchus contortus, a major veterinary parasite. Our results suggest that passively ingested iL3s migrate off their original fecal source and actively navigate toward hosts or new host fecal sources using olfactory cues. Olfactory plasticity may be a mechanism that enables iL3s to switch from dispersal behavior to host-seeking behavior. Together, our results demonstrate that passively ingested nematodes do not remain inactive waiting to be swallowed, but rather display complex sensory-driven behaviors to position themselves for host ingestion. Disrupting these behaviors may be a new avenue for preventing infections.
| Many parasitic nematodes infect by passive ingestion when the host consumes food, water, or feces containing infective third-stage larvae (iL3s). Passively ingested nematodes that infect humans cause severe gastrointestinal distress and death in endemic regions, and those that infect livestock are a major cause of production loss worldwide. Because these parasites do not actively invade hosts but instead rely on being swallowed by hosts, it has been assumed that they show only limited sensory responses and do not engage in host-seeking behaviors. Here, we investigate the olfactory behaviors of the passively ingested murine parasite Heligmosomoides polygyrus and show that this assumption is incorrect; H. polygyrus iL3s show robust attraction to a diverse array of odorants found in mammalian skin, sweat, and feces. Moreover, the olfactory responses of H. polygyrus iL3s are experience-dependent: some odorants are repulsive to iL3s cultured on feces but attractive to iL3s removed from feces. Olfactory plasticity is also observed in the ruminant parasite Haemonchus contortus, and may enable iL3s to disperse in search of new hosts or host fecal sources. Our results suggest that passively ingested nematodes use olfactory cues to navigate their environments and position themselves where they are likely to be swallowed. By providing new insights into the olfactory behaviors of these parasites, our results may enable the development of new strategies for preventing infections.
| Passively ingested gastrointestinal parasitic nematodes of humans and livestock are a significant health and economic problem. Human-infective nodular worms in the genus Oesophagostomum are a growing health concern in endemic regions of Africa, where they can cause abdominal pain, weight loss, diarrhea, and death [1–3]. Passively ingested parasites of livestock result in decreased production and economic loss worldwide. For example, Haemonchus contortus is an important parasite of ruminants that causes gastrointestinal distress, anemia, edema, and death in livestock [4]. In the United States alone, over 2.7 million goats and 2.6 million sheep are infected with H. contortus [5]. Infections with these parasites can be cleared using anthelmintic drugs, but frequent administration has led to increased drug resistance [6–9]. Although the host immune response to infection with passively ingested nematodes is well-studied [10–12], remarkably little is known about the behaviors of the parasites themselves. A better understanding of the behaviors exhibited by the environmental life stages of these parasites could facilitate the development of new strategies for preventing infections of humans and livestock, such as the use of targeted traps or repellents.
Parasitic nematodes that actively invade hosts by skin penetration are known to engage in sensory-driven host seeking [13]. For example, the human hookworms Ancylostoma duodenale and Necator americanus, and the dog hookworm Ancylostoma caninum, are relatively inactive in the absence of sensory stimuli but show increased activity in the presence of heat, CO2, and/or skin extract [14–16]. Hookworms also migrate robustly toward a heat source [14, 17]. The human, non-human primate, and canine threadworm Strongyloides stercoralis, and the rat parasites Strongyloides ratti and Nippostrongylus brasiliensis, also respond robustly to host-emitted sensory cues. They are active in the absence of sensory stimuli [18], and show robust attraction to a wide variety of odorants emitted by human skin and sweat [18–20]. S. ratti is also known to be attracted to blood serum, and S. stercoralis to blood serum, sweat, and heat [19, 21, 22].
The sensory behaviors of passively ingested nematodes are much less understood. Some passively ingested worms are capable of responding to environmental sensory cues such as temperature, humidity, and odorants [13]. For example, H. contortus uses temperature and humidity cues to migrate vertically through grass in response to changes in environmental conditions [23, 24]. Because passively ingested worms do not actively invade hosts, it has often been assumed that they do not host seek and do not respond to host-emitted sensory cues. However, we recently showed that H. contortus is attracted to some host-emitted odorants, raising the possibility that it can use olfactory cues to position itself in the vicinity of potential hosts [18]. Since many hosts develop immunity to passively ingested worms following repeated infection [25, 26], behaviors that expose these parasites to new hosts may be important for parasite propagation.
Here, we use the passively ingested gastrointestinal murine parasite H. polygyrus (also called H. bakeri [27, 28]) as a model system for studying the sensory behaviors of passively ingested gastrointestinal nematodes, and for testing the hypothesis that passively ingested nematodes engage in host seeking. As a mouse parasite, H. polygyrus is one of the only passively ingested nematodes that can be easily maintained in the lab [29, 30]. H. polygyrus is only infective as developmentally arrested iL3s, which are analogous to Caenorhabditis elegans dauers (S1 Fig) [31]. H. polygyrus iL3s were thought to primarily reside in host feces and infect when mice, which are coprophagic, eat infested feces [32, 33]. However, H. polygyrus iL3s can also attach to mouse fur and be ingested during grooming [34]. H. polygyrus iL3s were previously shown to nictate [33, 34], a behavior where the iL3 stands on its tail and waves its head [13], which may increase the probability of being swallowed during coprophagy or of becoming attached to mouse fur [34]. Once inside the host, the nematodes grow to adulthood and reproduce in the host intestine. H. polygyrus eggs then exit the host in feces and develop there into iL3s capable of infecting new hosts.
The fact that H. polygyrus develops on feces and infects mice from feces raises the question of whether H. polygyrus iL3s engage in environmental navigation using either host-emitted or environmental sensory cues, or whether they simply remain on feces and wait to be ingested. While this question had not been investigated thoroughly, H. polygyrus iL3s were previously found to be attracted to mouse urine and skin lipids, suggesting they are capable of responding to at least some host sensory cues [34]. However, the extent to which H. polygyrus iL3s engage in sensory behaviors that increase the likelihood that they will be swallowed by hosts remained unclear.
To address this question, we conducted a large-scale quantitative analysis of the unstimulated and odor-stimulated behaviors of H. polygyrus. We found that H. polygyrus iL3s were active in the absence of odor stimulation. In addition, they were attracted to host fecal odor. While they showed robust attraction to fresh feces, they showed reduced attraction to aged feces and ultimately migrated off their original fecal source to engage in environmental navigation. H. polygyrus iL3s were attracted to skin odorants as well as fecal odorants, suggesting that they are capable of migrating toward hosts as well as new host fecal sources. In addition, H. polygyrus iL3s showed experience-dependent olfactory plasticity, such that some host-emitted odorants were repulsive to iL3s cultured on feces but attractive to iL3s cultured off feces. Olfactory plasticity was also observed in the ruminant parasite H. contortus, and may be a general mechanism that enables passively ingested iL3s to shift from dispersal behavior to host-seeking behavior. Our results suggest that passively ingested nematodes disperse from feces and engage in host seeking to position themselves where they are likely to be ingested by new hosts.
Parasitic nematodes are known to vary in their environmental navigation strategies: some are cruisers that actively navigate toward hosts; some are ambushers that are less active and primarily attach to passing hosts; and some use an intermediate strategy [13]. To gain insight into the movement strategy used by H. polygyrus, we first examined the unstimulated movement of H. polygyrus iL3s, and compared their movement to that of S. stercoralis and S. ratti iL3s, which are known to be cruisers [18]. Using a dispersal assay in which iL3s were allowed to migrate on an agar surface in the absence of applied sensory stimulation for 1 hour, we found that H. polygyrus iL3s and S. ratti iL3s dispersed to a similar extent, whereas S. stercoralis iL3s dispersed more than either rodent parasite (Fig 1A). These results demonstrate that H. polygyrus iL3s are active in the absence of sensory stimulation and are capable of exhibiting a movement strategy resembling that of a cruiser. The increased movement of S. stercoralis iL3s relative to H. polygyrus and S. ratti iL3s may reflect the larger habitats of humans relative to nesting rodents [18]; since nesting rodents spend more time near their fecal deposits than do humans, non-human primates, and dogs, S. stercoralis iL3s may need to disperse farther into the environment to successfully locate a host.
Dispersal behavior reflects both crawling speed and other parameters such as crawling trajectory and tendency to pause during crawling. To gain more insight into the navigational strategy used by H. polygyrus iL3s, we tracked their crawling speed using automated worm tracking [35]. We found that H. polygyrus iL3s crawled more slowly than S. ratti iL3s, while S. stercoralis iL3s crawled much more rapidly than the rodent parasites (Fig 1B). The ability of H. polygyrus iL3s to disperse to the same extent as S. ratti iL3s despite their slower crawling speed suggests that H. polygyrus iL3s exhibit more linear and/or continuous movement than S. ratti iL3s.
We also evaluated the nictation behavior of H. polygyrus. Many skin-penetrating and passively ingested iL3s engage in nictation, a common ambushing behavior, as a means of increasing host contact. By standing up on a surface, nictating iL3s are more likely to touch and then transfer onto a passing host, or to be swallowed by a foraging host [13]. We assayed the nictation behavior of H. polygyrus, and compared it to that of S. ratti and S. stercoralis, using “micro-dirt” agar chips with near-microscopic pillars as an artificial dirt substrate (S2 Fig) [36]. The pillars on the agar surface minimize surface tension, allowing the iL3s to stand. We found that all three of the species showed similarly low nictation frequencies: only ~20–30% of the tested iL3s nictated during the assay period (Fig 1C). The low nictation frequencies of S. ratti and S. stercoralis are consistent with a cruising navigational strategy [18]. The similarly low nictation frequency of H. polygyrus, combined with its active crawling behavior, suggests that it also behaves more like a cruiser than an ambusher. These results demonstrate that passively ingested iL3s do not remain inactive waiting to be swallowed by passing hosts. Rather, like skin-penetrating iL3s, they engage in environmental navigation.
If passively ingested iL3s utilize active strategies to position themselves in optimal locations for host ingestion, one strong prediction is that the species that infect coprophagic hosts (e.g., mice) will be attracted to host feces. We examined the response of H. polygyrus iL3s to fresh fecal odor using a chemotaxis assay in which the iL3s could smell but not make contact with the feces. We found that H. polygyrus iL3s were strongly attracted to fresh mouse feces (Fig 2A and 2B). Moreover, they preferred mouse feces to gerbil or rabbit feces (Fig 2B and 2C), indicating that they can distinguish host from non-host feces. By contrast, S. stercoralis and S. ratti iL3s were neutral to host feces (Fig 2A) [18]. The different responses of H. polygyrus and Strongyloides iL3s to fecal odor are understandable in the context of their different lifestyles. Although the pre-infective larvae of both H. polygyrus and Strongyloides inhabit host feces, H. polygyrus iL3s can infect hosts from feces while skin-penetrating iL3s must migrate off feces and onto host skin [13, 30]. Thus, attraction to host feces would likely be ecologically advantageous for H. polygyrus iL3s but not Strongyloides iL3s. In addition, we found that H. polygyrus iL3 were more attracted to fresh feces than aged feces (Fig 2D), suggesting that the iL3s use olfaction to identify favorable fecal sources. In contrast, they did not show a preference for feces from uninfected versus infected hosts (Fig 2D), suggesting that they are attracted to fresh host feces regardless of the infection status of the host. Attraction of H. polygyrus iL3s to fecal odor may cause some of the iL3s on a fresh fecal source to remain there, and may draw iL3s from fecal sources that have become suboptimal due to age, desiccation, or other conditions.
The robust attraction of H. polygyrus iL3s to fecal odor raised the question of whether the iL3s leave feces under normal conditions. To address this question, we performed two different fecal dispersal assays, the first to assess short-term dispersal over the course of hours and the second to assess long-term dispersal over the course of days. In the short-term dispersal assay, iL3s were placed on fresh feces in the center of an agar surface. The frequency with which the iL3s migrated off the feces and onto the agar was then quantified. We found that on average, 50% of the iL3 population left the fresh feces; in some trials, over 80% of the iL3s left the feces (Fig 2E). These results demonstrate that even for iL3s on fresh feces, which are presumably a favorable fecal source, a substantial portion of the iL3 population migrates off of the feces and engages in environmental navigation.
In the long-term dispersal assay, a fresh fecal pellet from an infected animal was collected, and one-half of the pellet was placed in the center of an agar surface. The frequency with which the nematodes migrated off of the feces and onto the agar was then quantified each day for a period of 10 days. Thus, this assay examined H. polygyrus dispersal in the more natural context of fecal aging. We found that nearly all of the nematodes remained on the feces until day 5. On day 5, by which time the nematodes had developed into iL3s [29], over 80% of the nematodes migrated off the feces (Fig 2F). By day 10, nearly 100% of the nematodes had migrated off of the feces (Fig 2F). In the same assay, we also examined nictation behavior and found that nictation occurs primarily on day 5 (S3 Fig), at the time when the majority of the population migrates off of the feces (Fig 2F). Together, these results argue against the possibility that some members of the iL3 population are ambushers while others are cruisers, and suggest instead that nearly all H. polygyrus iL3s ultimately engage in cruising behavior.
Our results show that H. polygyrus iL3s will eventually leave their original fecal source and migrate toward new fecal sources to position themselves for ingestion during coprophagy. However, H. polygyrus iL3s can infect during grooming [34], raising the question of whether they also migrate toward hosts by detecting host-emitted olfactory cues. To investigate this possibility, we examined the responses of H. polygyrus iL3s to a large panel of odorants that included compounds found in mammalian skin and sweat using a chemotaxis assay (S4 Fig) [18]. We found that H. polygyrus iL3s showed robust attraction to 6 of the 35 odorants tested: 2-butanone; 2,3-butanedione; geranyl acetone; 3-methyl-1-butanol; 2-methyl-1-butanol; and 3-heptanol (Fig 3). In contrast, CO2 was repulsive for H. polygyrus iL3s (Fig 3). All of the attractive odorants are emitted from mammalian skin, feces, and/or urine [18, 37–41]. Notably, 2-methyl-1-butanol, 3-methyl-1-butanol, and geranyl acetone are present in skin microbiota [42, 43] and are known attractants for skin-penetrating nematodes [18]. Attraction to these odorants could drive migration of H. polygyrus iL3s toward hosts.
To gain insight into how the olfactory preferences of H. polygyrus iL3s differ from those of other iL3s that engage in environmental navigation, we compared the odor-driven behaviors of H. polygyrus to those of 7 other nematode species: the skin-penetrating human-parasitic nematode S. stercoralis, the skin-penetrating rat-parasitic nematodes S. ratti and N. brasiliensis, the passively ingested ruminant-parasitic nematode H. contortus, the actively invading entomopathogenic nematodes Heterorhabditis bacteriophora and Steinernema carpocapsae, and the free-living bacterivorous nematode C. elegans. This comparison revealed that H. polygyrus responds differently to the odorant panel than the other species (S5A Fig), consistent with previous studies demonstrating that parasitic nematodes show species-specific olfactory preferences [18, 44, 45]. Moreover, cluster analysis of the 8 species based on their olfactory preferences revealed that parasitic nematodes that infect the same hosts have more similar olfactory preferences than parasitic nematodes that infect different hosts (S5B Fig) [18, 44, 45]. In contrast, parasitic nematodes that infect different hosts but share the same mode of infection do not respond similarly to odorants. In particular, H. polygyrus and H. contortus are both passively ingested but infect different hosts, and their olfactory preferences are dissimilar (S5B Fig). Thus, olfactory preferences appear to be determined primarily by host range rather than infection mode. The fact that distantly related species that target the same host respond similarly to odorants strongly suggests that parasitic nematode olfactory behavior has evolved to mediate specific parasite-host interactions.
iL3s that have migrated off feces likely face a greater ethological drive to search for new hosts or fecal sources than iL3s that have remained on feces. We therefore wondered whether iL3s that have migrated off feces might exhibit different behaviors than iL3s on feces. To test this possibility, we compared the unstimulated migration of iL3s cultivated on feces to those of iL3s that had been removed from feces and maintained in dH2O for 1 week. We found that the off-feces iL3s dispersed to a greater extent than the on-feces iL3s (Fig 4A), demonstrating that the unstimulated activity of H. polygyrus iL3s is subject to experience-dependent modulation. The greater dispersal of off-feces iL3s was not due to changes in crawling speed (Fig 4B); thus, the difference in dispersal reflects a difference in navigational strategy rather than motility. In addition, the nictation rate of on-feces vs. off-feces iL3s was unchanged (Fig 4C), demonstrating that removal from feces results in a specific change in crawling behavior. The increased dispersal of off-feces iL3s likely increases the probability of encountering a new host or fecal source.
To further elucidate the effects of recently experienced environment on H. polygyrus behavior, we compared the olfactory preferences of on-feces vs. off-feces iL3s to a subset of mammalian odorants. The odorant panel was selected to include attractive, neutral, and repulsive odorants. We found that on-feces and off-feces iL3s responded differently to 2 of 8 tested odorants: CO2 and benzaldehyde. Both odorants were repulsive for iL3s on feces but attractive for iL3s off feces (Fig 5A). For both on-feces and off-feces iL3s, CO2-response valence, i.e. whether CO2 was repulsive or attractive, was consistent across concentrations (S6 Fig). CO2 is a critical host cue for many parasites, including many parasitic nematodes [13]; it is present at high concentrations in both exhaled breath and feces. Benzaldehyde is found in skin, breath, urine, and feces [18]. Thus, the olfactory responses of H. polygyrus iL3s to some host-associated odorants are subject to experience-dependent modulation as a result of recently experienced environmental conditions.
We then examined the relationship between cultivation environment and sensory behavior in more detail by investigating the time course of the change in CO2- and benzaldehyde-response valence. We found that CO2-response valence changed gradually over the course of days when iL3s were removed from feces (Fig 5B). Moreover, culturing iL3s under high CO2 conditions prevented the shift in CO2-response valence following removal from feces. While iL3s cultured off feces at ambient CO2 (~0.04% CO2 [46]) were attracted to CO2, iL3s cultured off feces at high CO2 (2.5% CO2) were repelled by CO2 (Fig 5C). Thus, CO2-response valence is regulated by environmental CO2 levels. Benzaldehyde-response valence also changed gradually over the course of days upon removal from feces and was also determined by environmental CO2 levels (Fig 5D and 5E). These results suggest that the level of environmental CO2 acts as a general regulator of olfactory behavior. Given that feces emit high levels of CO2 [39], H. polygyrus iL3s may use environmental CO2 levels to signal the presence or absence of feces, with the result that exposure to high CO2 levels mimics the effects of exposure to feces. Experience-dependent olfactory plasticity may be a mechanism that enables iL3s on feces to disperse from the feces, and iL3s that have been off feces for a prolonged period to instead migrate toward new hosts or fresh fecal sources.
Our finding that H. polygyrus iL3s exhibit experience-dependent olfactory plasticity raised the question of whether this behavior is unique to H. polygyrus or shared with other parasitic nematode species. To distinguish between these possibilities, we examined the CO2-evoked behaviors of H. contortus, S. stercoralis, and the skin-penetrating human-parasitic hookworm Ancylostoma ceylanicum cultured on versus off feces. We found that like H. polygyrus iL3s, H. contortus iL3s show experience-dependent plasticity in their response to CO2. In the case of H. contortus, CO2 is neutral for iL3s cultured on feces but attractive for iL3s cultured off feces (Fig 6A). Since H. contortus iL3s are long-lived [47, 48], sometimes surviving in the environment for up to 8 months [48], we examined the CO2-evoked behavior of off-feces iL3s over the course of 5 weeks. We found that CO2 changed from neutral to attractive after 1 week, and then remained attractive in subsequent weeks (Fig 6A). Thus, CO2 remains a strong attractant for H. contortus iL3s that have been removed from feces for prolonged periods. Our results demonstrate that experience-dependent olfactory plasticity is not unique to H. polygyrus, but also occurs in other passively ingested nematodes. Experience-dependent modulation of CO2 response may enable H. contortus iL3s to first migrate off feces and then navigate toward grazing hosts, which emit high concentrations of CO2 in their exhaled breath.
In contrast to the passively ingested nematodes, the skin-penetrating nematodes tested did not show experience-dependent modulation of their CO2-evoked behavior. Both S. stercoralis iL3s and A. ceylanicum iL3s were repelled by CO2 when cultured both on and off feces (Fig 6B and 6C). The lack of flexibility in their CO2-evoked behavior may reflect the fact that CO2 attraction would likely not facilitate host finding by skin-penetrating worms, since very low levels of CO2 are given off by the skin [49]. CO2 avoidance may function as a dispersal mechanism to drive skin-penetrating iL3s off host feces; attraction to other sensory cues, such as skin and sweat odorants, may then drive the iL3s toward potential hosts [13]. Thus, the ability to exhibit flexible responses to CO2 may be a specific behavioral adaptation of passively ingested but not skin-penetrating nematodes.
Here we conducted the first large-scale quantitative behavioral analysis of H. polygyrus iL3s. We found that H. polygyrus iL3s were active even in the absence of sensory stimulation (Fig 1). These results argue against the classical notion that passively ingested iL3s remain stationary and wait to be swallowed, and suggest instead that these iL3s actively navigate their environments. We previously showed that H. contortus iL3s are less active than S. ratti and S. stercoralis iL3s [18]. However, the similar dispersal behaviors and nictation rates of H. polygyrus and S. ratti (Fig 1) suggest that some passively ingested nematodes are as active as skin-penetrating nematodes despite their passive mode of infection.
Our examination of the olfactory preferences of H. polygyrus iL3s revealed that they are attracted to fecal odor as well as mammalian skin and sweat odorants (Figs 2 and 3). These results suggest that passively ingested iL3s engage in odor-driven host seeking to position themselves near hosts or host feces, where they are likely to be ingested. Consistent with the attraction of H. polygyrus iL3s to both fecal odor and host odorants, H. polygyrus iL3s have been shown to infect hosts either from feces during coprophagy or from fur during grooming [30, 32–34]. Thus, active migration toward new hosts or fecal sources may be a critical but often overlooked aspect of the environmental stage of the H. polygyrus life cycle.
The robust attraction of H. polygyrus iL3s to fecal odor could serve to keep some of the iL3s on favorable fecal sources, or to direct them away from suboptimal fecal sources toward more favorable sources. However, we found that even when iL3s are placed on fresh feces, which is presumably a favorable fecal source, approximately half of the population migrates off of the feces within an hour (Fig 2E). Moreover, we found that nearly all iL3s eventually leave their original fecal source to engage in environmental navigation (Fig 2F). These results suggest that all H. polygyrus iL3s are capable of engaging in environmental navigation, and that if they are not ingested with feces shortly after reaching the iL3 stage, they will leave their original fecal source and disperse into the environment. Once in the environment, they use olfactory cues to migrate toward hosts or new fecal sources (Fig 7).
At the population level, this behavioral flexibility may help to ensure maximal infection rates. Remaining on a known fecal source can in some cases be beneficial: if that fecal source is in or near a nest, the iL3s may encounter hosts by remaining in the nest. However, this behavioral strategy also carries risk: many mice forage and deposit feces far from their nests, in locations where the iL3s are less likely to encounter a mouse using a “sit-and-wait” strategy [34]. Under these circumstances, first dispersing from feces and then using host-emitted sensory cues to migrate toward new hosts or fecal sources is likely to be essential to continue the life cycle. Thus, maximal parasite survival may be achieved when iL3s that do not immediately encounter a host actively disperse in search of hosts. In future studies, it will be interesting to determine whether nematodes that exit the host early in an infection cycle show different dispersal behavior than nematodes that exit the host late in an infection cycle, or whether nematodes that emerge from hosts with a heavier worm burden show different dispersal behavior than nematodes that emerge from hosts with a lighter worm burden.
What is the mechanism that drives some iL3s to migrate off feces, and subsequently toward new hosts or fecal sources? We speculate that olfactory plasticity may function as this mechanism. We have shown that H. polygyrus iL3s display experience-dependent olfactory plasticity: some odorants are repulsive to iL3s that have been cultured on feces but attractive to iL3s that have been cultured off feces for a week (Fig 5). Repulsion of iL3s from odorants such as CO2 and benzaldehyde, which are emitted by host feces [18, 39], may cause the iL3s to migrate off of their original fecal source and disperse into the environment. Once the iL3s have been in the environment for multiple days, these odorants become attractive, likely driving the iL3s toward new hosts or fecal sources.
The shift from repulsion to attraction for both CO2 and benzaldehyde response is mediated by environmental CO2 levels (Fig 5C and 5E). When iL3s are removed from feces but cultured in the presence of high CO2, they remain repelled by both CO2 and benzaldehyde. However, when iL3s are removed from feces and cultured at ambient CO2, they become attracted to CO2 and benzaldehyde. These results suggest that environmental CO2 levels may be used as a proxy for the presence or absence of feces.
We found that like H. polygyrus iL3s, H. contortus iL3s show experience-dependent modulation of their CO2-evoked behavior. H. polygyrus iL3s showed a shift in their CO2 response from repulsive to attractive following removal from feces (Fig 5), while H. contortus iL3s showed a shift in their CO2 response from neutral to attractive (Fig 6A). Thus, in both cases, CO2 attraction is likely to be observed in nature in iL3s that have migrated off of feces and are engaging in environmental navigation. In contrast to the passively ingested nematodes tested, the skin-penetrating nematodes tested did not show experience-dependent modulation of their CO2-evoked behavior (Fig 6B and 6C). Thus, experience-dependent plasticity based on the presence or absence of feces may be specific to passively ingested nematodes. The differences in CO2-evoked behavior between passively ingested iL3s and skin-penetrating iL3s are consistent with their different ecologies. Skin-penetrating iL3s infect primarily via the skin, which emits low levels of CO2 [49], so CO2 attraction may not be beneficial for skin-penetrating iL3s regardless of their cultivation conditions. Passively ingested nematodes infect via the mouth, which emits high levels of CO2 [50]. Thus, in the case of passively ingested nematodes, repulsive or neutral responses to CO2 by iL3s on feces may initially drive them off feces, while subsequent attractive responses to CO2 may drive them toward the mouths of respiring hosts. H. contortus is one of the most economically significant livestock parasites worldwide [5], and drug resistance resulting from repeated use of anthelmintic drugs is already a major challenge in combatting infections [9]. Our finding that the olfactory responses of H. contortus are experience-dependent could facilitate the development of odor-based traps or repellents that could be used in combination with grazing management interventions [51, 52] to prevent nematode infections.
The circuit mechanisms that drive experience-dependent valence changes in passively ingested nematodes remain to be determined. In C. elegans, CO2-response valence is also subject to experience-dependent modulation: adults cultured at ambient CO2 are repelled by CO2, while adults cultured at high CO2 are attracted to CO2 [53]. Both CO2 attraction and CO2 repulsion by C. elegans are mediated by the BAG sensory neurons in the head and a group of downstream interneurons. The CO2-evoked activity of these interneurons is subject to experience-dependent modulation, enabling them to generate opposite behavioral responses to CO2 [53]. Since sensory neuroanatomy is generally conserved across nematode species [13], similar circuit mechanisms may operate in passively ingested parasitic nematodes to regulate CO2-response valence.
The molecular mechanisms that drive experience-dependent valence changes in passively ingested nematodes are also not yet known. In C. elegans, CO2-response valence is regulated by neuropeptide signaling [53]. However, CO2-response valence in C. elegans changes over the course of hours [53], while CO2-response valence in passively ingested parasitic nematodes changes over the course of days (Figs 5B and 6A). Thus, the valence change in parasitic nematodes could involve changes in gene expression and/or neuronal wiring, which occur on a slower timescale than neuropeptide signaling [54–59]. Elucidating the mechanisms that operate in passively ingested nematodes to control olfactory valence will require the development of genetic engineering techniques for these species, which have so far remained intractable to molecular genetic manipulation [60]. Targeted mutagenesis using the CRISPR-Cas9 system has now been achieved in Strongyloides species [60–61], and may be applicable to other types of parasitic nematodes in the future.
Entomopathogenic nematodes and skin-penetrating nematodes also show olfactory plasticity, but in response to changes in their prior cultivation temperature [62]. In addition, the entomopathogenic nematode Steinernema scapterisci shows age-dependent olfactory plasticity in its response to CO2: CO2 changes from a repulsive cue in young iL3s to an attractive cue in older iL3s [62]. Thus, olfactory plasticity may be a general feature of parasitic nematode behavior that enables iL3s to modulate their sensory responses based on internal or external conditions so as to increase their chances of encountering a host.
Passively ingested nematodes comprise a group of human and livestock parasites whose behaviors have remained elusive. Increased drug resistance [6–9] necessitates the development of new strategies for their control. Our results suggest that passively ingested nematodes engage in robust and dynamic odor-driven host-seeking behaviors. A better understanding of these behaviors may lead to new strategies for preventing infections.
H. polygyrus was passaged in mice, S. stercoralis was passaged in gerbils, and A. ceylanicum was passaged in hamsters. All procedures and protocols were approved by the UCLA Office of Animal Research and Oversight (Protocol 2011-060-13B), which adheres to the standards of the AAALAC and the Guide for the Care and Use of Laboratory Animals.
Heligmosomoides polygyrus (also called Heligmosomoides bakeri [27]) was generously provided by Dr. Raffi Aroian (University of Massachusetts Medical School). Strongyloides stercoralis (UPD strain) was generously provided by Dr. James Lok (University of Pennsylvania), Ancylostoma ceylanicum (Indian strain, US National Parasite Collection Number 102954) was generously provided by Dr. John Hawdon (George Washington University), and Haemonchus contortus was generously provided by Dr. Anne Zajac (Virginia-Maryland College of Veterinary Medicine). Male or female C57BL/6 mice for propagation of H. polygyrus were obtained from the UCLA Division of Laboratory Animal Medicine Breeding Colony. Male Mongolian gerbils for propagation of S. stercoralis and male Syrian golden hamsters for propagation of A. ceylanicum were obtained from Envigo. H. contortus was not propagated in our laboratory.
H. polygyrus was serially passaged in C57BL/6 male or female mice as described [30] and maintained on fecal-charcoal plates as described [18]. Briefly, mice were inoculated with ~150 iL3s administered in 100 μL ddH2O by oral gavage. Feces infested with H. polygyrus were collected between days 10 and 60 post-inoculation. Feces were obtained by placing mice overnight on wire cage bottoms above damp cardboard, and collecting the pellets from the cardboard the following morning. Fecal pellets were mixed with dH2O and autoclaved charcoal granules to make fecal-charcoal plates. Plates were stored at room temperature until use. iL3s used for behavioral analysis were collected from fecal-charcoal plates using a Baermann apparatus [63]. iL3s cultured “on feces” were collected from fecal-charcoal plates on day 14 (with day 0 being the day of fecal collection) and tested immediately; iL3s cultured “off feces” were collected from fecal-charcoal plates on day 7, incubated in dH2O for 7 days at room temperature, and tested on day 14. For the odorant chemotaxis assays in Fig 3, iL3s were either collected from fecal-charcoal plates on days 7–14 and tested immediately or collected on days 7–14 and stored for up to 10 days in dH2O at 4°C prior to testing (storage at 4°C in dH2O is a standard cultivation condition for H. polygyrus [30]). In all cases where differences were observed following storage in dH2O at 4°C, the data from iL3s stored at 4°C in dH2O was excluded from the analysis. For the “off feces” time course in Fig 5, iL3s were collected from fecal-charcoal plates on day 7, incubated in dH2O for the indicated number of days, and then tested. For assays involving iL3s cultured at 2.5% CO2 either on or off feces, iL3s were collected from fecal-charcoal plates on day 7. iL3s for the on-feces condition were placed onto new fecal-charcoal plates containing autoclaved feces, stored in a CO2 incubator with 2.5% CO2 for 7 days, and collected from the fecal-charcoal plates using a Baermann apparatus immediately prior to testing. iL3s for the off-feces condition were incubated in dH2O in a CO2 incubator with 2.5% CO2 for the indicated number of days and then tested.
H. contortus was maintained on fecal-charcoal plates as described [18]. Plates were stored in an incubator at 23°C until use. iL3s used to test CO2 response in Fig 6A were either cultured on fecal-charcoal plates for up to 9 weeks and then tested immediately; or removed from feces, stored in dH2O for up to 5 weeks, and then tested. Notably, iL3s maintained on feces and tested immediately showed a neutral response to CO2 regardless of their age, demonstrating that the CO2 attraction of off-feces iL3s was due to their removal from feces and not their age.
S. stercoralis was serially passaged in male Mongolian gerbils and maintained on fecal-charcoal plates as described [18]. Briefly, gerbils were inoculated with ~2,250 iL3s in 200 μL sterile PBS by subcutaneous injection. Feces infested with S. stercoralis were collected between days 14 and 45 post-inoculation. Feces were harvested and mixed with autoclaved charcoal granules to make fecal-charcoal plates as described above. Plates were stored in an incubator at 23°C until use. iL3s used to test CO2 response in Fig 6B were cultured on fecal-charcoal plates until day 7; they were then either tested immediately, stored in BU saline [64] for 1 week and then tested, or stored in BU saline for 2 weeks and then tested.
A. ceylanicum was serially passaged in male Syrian golden hamsters and maintained on fecal-charcoal plates as described [18]. Briefly, hamsters were inoculated with ~100 iL3s in 100 μL sterile ddH2O by oral gavage. Feces infested with A. ceylanicum were collected between days 14 and 45 post-inoculation. Feces were harvested and mixed with autoclaved charcoal granules to make fecal-charcoal plates as described above. Plates were stored in an incubator at 23°C until use. iL3s used to test CO2 response in Fig 6C were cultured on fecal-charcoal plates until day 10; they were then either tested immediately, stored in BU saline [64] for 1 week and then tested, or stored in BU saline for 2 weeks and then tested.
Short-term dispersal assays without feces (Figs 1A and 4A) were performed essentially as described [18]. For each trial, ~50–100 iL3s were placed on a 10-cm chemotaxis plate [65] on a vibration-reducing platform and allowed to disperse for either 1 hour (Fig 1A) or 10 minutes (Fig 4A) in the absence of applied sensory stimuli. The number of iL3s in the outer zone of the plate (the region that excludes a 4-cm-diameter circle at the center of the plate) was then determined. For short-term fecal dispersal assays (Fig 2E), fresh fecal pellets were collected the morning of the assay from uninfected animals. One fecal pellet (~0.03 g) was placed in the center of a 10-cm chemotaxis plate. 15–40 iL3s were pipetted onto the pellet. The plates were then placed on a vibration-reducing platform for 1 hour. The number of iL3s either on the feces, off the feces but within a 4-cm-diameter circle around the feces (zone 1), or outside a 4-cm-diameter circle around the feces (zone 2) was then determined (Fig 2E). iL3s were not visible when they were on the fecal pellet, so the number of iL3s remaining on the feces at the end of the assay was determined by subtracting the number of iL3s in zones 1 and 2 from the total number of iL3s added to the feces. Note that for all dispersal assays, the outermost zone included the walls of the plate, which functioned as a trap such that most of the iL3s that crawled onto the walls of the plate remained there for the duration of the assay.
Long-term fecal dispersal assays (Fig 2F) were performed by first collecting fresh feces from infected animals; feces were collected as described above, but from a 4-hour collection period. Feces were collected from host animals that were each infected with ~75 iL3s. Individual fecal pellets of similar size were cut in half; one-half of a fecal pellet was then placed on each chemotaxis assay plate and incubated at room temperature. Every 24 hours (within a 3-hour window), the number of animals that had migrated out of the feces and onto the chemotaxis plate was quantified. After quantification, fecal pellets were transferred to fresh chemotaxis plates. On day 10, the fecal pellets were dissociated and the number of iL3s remaining in the fecal pellet was quantified. These numbers were then used to calculate the total number of worms that started out on each fecal pellet, and the cumulative percentage of worms that migrated off the fecal pellet each day. Nictation rates were also determined for each day by counting the number of worms observed to be nictating on the fecal pellet at each time of observation. These numbers were used, in combination with the number of worms remaining on the fecal pellet for each day (calculated as described above), to calculate the percentage of worms nictating on the fecal pellets at each time of observation (S3 Fig).
Automated tracking was performed as described [18]. For each recording session, 10–15 iL3s were placed on a chemotaxis plate and allowed to acclimate for 10 minutes. iL3 movement was then captured for 20 seconds using an Olympus E-PM1 digital camera attached to a Leica S6 D microscope. WormTracker and WormAnalyzer [35] were used to quantify crawling speed. WormTracker and WormAnalyzer settings were previously described [18].
The nictation assays shown in Figs 1C and 4C were performed essentially as described (S2 Fig) [18, 62]. Briefly, agar chips for nictation assays were made from polydimethylsiloxane (PDMS) molds [36]. Chips were approximately 3 cm x 3.5 cm and contained near-microscopic posts that allowed the iL3s to stand. Chips were made using 4% agar dissolved in ddH2O. Once the agar had solidified, chips were placed at 37°C for 2 hours followed by room temperature for 1 hour. 10–20 iL3s were transferred to the center of the chip in a 5 μL drop of dH2O and allowed to acclimate for 10 minutes. Individual iL3s were then monitored for 2 minutes, and the number of iL3s that nictated during the observation period was recorded. Nictation was defined as an iL3 raising at least half of its body off the plate for at least 5 seconds (S2 Fig).
Chemotaxis assays were performed on chemotaxis plates as described [18, 44]. For fecal and odorant chemotaxis assays, 2 μL 5% sodium azide was placed into the center of each scoring region. For fecal chemotaxis assays, feces were obtained from an overnight fecal collection. For assays involving feces from uninfected vs. infected animals (Fig 2D, right), feces were obtained from a 4-hour fecal collection. The feces were then incubated for 3 days at room temperature in a 10-cm Petri dish on filter paper moistened with 1 mL ddH2O to prevent desiccation. For assays involving fresh vs. aged feces (Fig 2D, center), feces were obtained from a 4-hour fecal collection and stored in a 10-cm Petri dish without filter paper. “Fresh feces” refers to feces that were used on the day of collection, while “aged feces” refers to feces that were incubated in the Petri dish for 1 day. For all fecal assays, the feces were moistened to a paste with ddH2O. 0.5-cm squares of filter paper were affixed to the lid of a chemotaxis plate using double-stick tape. 0.25 g fecal paste was placed onto one of the filter paper squares, and either 50 μL ddH2O (for normal fecal chemotaxis assays) or 0.25 g of feces (for fecal competition chemotaxis assays) was added to the other square.
For odorant chemotaxis assays (S4 Fig), 5 μL odorant was pipetted into the center of one scoring region and 5 μL control (paraffin oil, ddH2O, or ethanol) was pipetted into the center of the other scoring region. Liquid odorants were tested undiluted. Solid odorants were dissolved to test concentrations as follows: tetradecanoic acid, indole, and 3-methylindole were diluted 0.05 g in 2.5 mL ethanol; octadecanoic acid was diluted 1 g in 80 mL ethanol; L-lactic acid was diluted 0.05 g in 2.5 mL ddH2O; and ammonia was purchased as a 2 M solution in ethanol. ddH2O was used as a control for L-lactic acid; ethanol was used as a control for tetradecanoic acid, octadecanoic acid, indole, 3-methylindole, and ammonia; and paraffin oil was used as a control for all other odorants. For CO2 chemotaxis assays, gases were delivered at a rate of 0.5 mL/min through holes in the plate lids as previously described [18, 44]. Gas stimuli were obtained from Airgas, and consisted of the test concentration of CO2, 21% O2, and the balance N2. Air controls consisted of 21% O2 and 79% N2. The test concentration of CO2 consisted of 15% CO2 for H. contortus and 10% CO2 for all other species, unless otherwise indicated.
For all chemotaxis assays, ~200 iL3s were pipetted onto the center of the chemotaxis plate and allowed to distribute in the stimulus gradient on a vibration-reducing platform for 3 h (for fecal and odorant chemotaxis assays) or 1 h (for CO2 assays). The number of iL3s in each scoring region was then quantified and a chemotaxis index was calculated as: (# iL3s at stimulus–# iL3s at control) / (# iL3s at stimulus + control). At least two identical assays were always performed simultaneously with the stimulus gradient oriented in opposite directions to control for directional bias due to room vibration or other causes; the pair of assays was discarded if the difference in the chemotaxis indices for the pair of plates was ≥0.9 or if either of the plates had <7 iL3s in the scoring regions. For the odorant chemotaxis assays in Fig 3, significance was calculated relative to a paraffin oil control.
Statistical analysis was performed using GraphPad Prism or PAST [66]. For each experiment, the D’Agostino-Pearson omnibus normality test was used to determine whether the data were normally distributed. If the data were normally distributed, parametric tests were used; otherwise, non-parametric tests were used. Graphs show medians and interquartile ranges to accurately depict the distribution and variance in our datasets. The heatmap in S5A Fig was generated using Heatmap Builder [67].
|
10.1371/journal.pbio.1001446 | Reconstruction of Ancestral Metabolic Enzymes Reveals Molecular Mechanisms Underlying Evolutionary Innovation through Gene Duplication | Gene duplications are believed to facilitate evolutionary innovation. However, the mechanisms shaping the fate of duplicated genes remain heavily debated because the molecular processes and evolutionary forces involved are difficult to reconstruct. Here, we study a large family of fungal glucosidase genes that underwent several duplication events. We reconstruct all key ancestral enzymes and show that the very first preduplication enzyme was primarily active on maltose-like substrates, with trace activity for isomaltose-like sugars. Structural analysis and activity measurements on resurrected and present-day enzymes suggest that both activities cannot be fully optimized in a single enzyme. However, gene duplications repeatedly spawned daughter genes in which mutations optimized either isomaltase or maltase activity. Interestingly, similar shifts in enzyme activity were reached multiple times via different evolutionary routes. Together, our results provide a detailed picture of the molecular mechanisms that drove divergence of these duplicated enzymes and show that whereas the classic models of dosage, sub-, and neofunctionalization are helpful to conceptualize the implications of gene duplication, the three mechanisms co-occur and intertwine.
| Darwin's theory of evolution is one of gradual change, yet evolution sometimes takes remarkable leaps. Such evolutionary innovations are often linked to gene duplication through one of three basic scenarios: an extra copy can increase protein levels, different ancestral subfunctions can be split over the copies and evolve distinct regulation, or one of the duplicates can develop a novel function. Although there are numerous examples for all these trajectories, the underlying molecular mechanisms remain obscure, mostly because the preduplication genes and proteins no longer exist. Here, we study a family of fungal metabolic enzymes that hydrolyze disaccharides, and that all originated from the same ancestral gene through repeated duplications. By resurrecting the ancient genes and proteins using high-confidence predictions from many fungal genome sequences available, we show that the very first preduplication enzyme was promiscuous, preferring maltose-like substrates but also showing trace activity towards isomaltose-like sugars. After duplication, specific mutations near the active site of one copy optimized the minor activity at the expense of the major ancestral activity, while the other copy further specialized in maltose and lost the minor activity. Together, our results reveal how the three basic trajectories for gene duplicates cannot be separated easily, but instead intertwine into a complex evolutionary path that leads to innovation.
| In a seminal book, Susumu Ohno argued that gene duplication plays an important role in evolutionary innovation [1]. He outlined three distinct fates of retained duplicates that were later formalized by others (for reviews, see [2],[3]). First, after a duplication event, one paralog may retain the ancestral function, whereas the other allele may be relieved from purifying selection, allowing it to develop a novel function (later called “neofunctionalization”). Second, different functions or regulatory patterns of an ancestral gene might be split over the different paralogs (later called “subfunctionalization” [4],[5]). Third, duplication may preserve the ancestral function in both duplicates, thereby introducing redundancy and/or increasing activity of the gene (“gene dosage effect” [6]).
Recent studies have shown that duplications occur frequently during evolution, and most experts agree that many evolutionary innovations are linked to duplication [7]–[10]. A well-known example are crystallins, structural proteins that make up 60% of the protein in the lenses of vertebrate eyes. Interestingly, paralogs of many crystallins function as molecular chaperones or glycolytic enzymes. Studies suggest that on multiple occasions, an ancestral gene encoding a (structurally very stable) chaperone or enzyme was duplicated, with one paralog retaining the ancestral function and one being tuned as a lens crystallin that played a crucial role in the optimization of eyesight [11],[12].
The molecular mechanisms and evolutionary forces that lead to the retention of duplicates and the development of novel functions are still heavily debated, and many different models leading to Ohno's three basic outcomes have been proposed (reviewed in [2],[3],[13],[14]). Some more recent models blur the distinction between neo- and subfunctionalization [15]. Co-option models, for example, propose that a novel function does not develop entirely de novo but originates from a pre-existing minor function in the ancestor that is co-opted to a primary role in one of the postduplication paralogs [2],[13]. Examples of such co-option models include the “gene sharing” or “Escape from Adaptive Conflict” (EAC) model [5],[16]–[19] and the related “Innovation, Amplification and Divergence” (IAD) model [20]–[22]. The IAD model describes co-option as a neofunctionalization mechanism. A “novel” function arises in the preduplication gene, and increased requirement for this (minor) activity is first met by gene amplification (e.g., through formation of tandem arrays). After this, adaptive mutations lead to divergence and specialization of some of the duplicate copies. The EAC model, on the other hand, describes co-option rather as a subfunctionalization mechanism by which duplication allows a multifunctional gene to independently optimize conflicting subfunctions in different daughter genes.
Another aspect in which various models differ is the role of positive selection. Some models emphasize the importance of neutral drift, while in other models adaptive mutations play an important role. For example, in the Duplication-Degeneration-Complementation (DDC) model of subfunctionalization [4], degenerative mutations (accumulated by neutral drift) lead to complementary loss-of-function mutations in the duplicates, so that both copies become essential to perform all of the functions that were combined in the single preduplication gene. Whereas this type of subfunctionalization only involves genetic drift [4],[8], other subfunctionalization models, such as the EAC model, attribute an important role to positive selection for the further functional optimization of the postduplication paralogs [2],[14].
There is a sharp contrast between the large number of detailed theoretical models of evolution after gene duplication, on the one hand, and the lack of clear experimental evidence for the various predictions made by these theories, on the other [2]. The key problem is the lack of knowledge about the functional properties of the ancestral, preduplication gene. Since these ancient genes and the proteins they encode no longer exist, many details in the chain of events that led from the ancestral gene to the present-day duplicates remain obscure. In most studies, the activities of the preduplication ancestor are inferred from unduplicated present-day outgroup genes that are assumed to have retained similar functional properties, but this is only an approximation. The central hurdle to surpass to obtain accurate experimental data on the evolution of gene duplicates involves rewinding the evolutionary record to obtain the sequence and activity of the ancestral proteins. Recent developments in sequencing and bio-informatics now enable us to reconstruct ancestral genes and proteins and characterize them in detail [23]–[31]. However, most ancestral reconstruction studies to date did not focus on the mechanisms that govern evolution after gene duplication.
In this study, we used the yeast MALS gene family as a model system to gain insight in the molecular mechanisms and evolutionary forces shaping the fate of duplicated genes. The MALS genes encode α-glucosidases that allow yeast to metabolize complex carbohydrates like maltose, isomaltose, and other α-glucosides [32],[33]. Several key features make this family ideal to study duplicate gene evolution. First, it is a large gene family encompassing multiple gene duplication events, some ancient and some more recent. Second, the present-day enzymes have diversified substrate specificities that can easily be measured [32]. Third, the availability of MALS gene sequences from many fungal genomes enabled us to make high-confidence predictions of ancestral gene sequences, resurrect key ancestral proteins, and study the selective forces acting throughout the evolution of the different gene duplicates. Fourth, the crystal structure of one of the present-day enzymes, Ima1, has been determined [34]. Molecular modeling of the enzymes' binding pocket, combined with activity measurements on reconstructed and present-day enzymes, allowed us to investigate how mutations altered enzyme specificity and gave rise to the present-day alleles that allow growth on a broad variety of substrates. Combining these analyses, we were able to study the evolution and divergence of a multigene family to an unprecedented level of detail and show that the evolutionary history of the MALS family exhibits aspects of all three classical models of duplicate gene evolution proposed by Ohno (gene dosage, neo-, and subfunctionalization).
Some yeast species have evolved the capacity to metabolize a broad spectrum of natural disaccharides found in plants and fruits (Figure 1, tree adapted from [35]). The origin of this evolutionary innovation seems to lie in the duplication and functional diversification of genes encoding permeases and hydrolases [32]. The common Saccharomyces cerevisiae laboratory strain S288c, for example, contains seven different MALS genes (MAL12, MAL32, and IMA1–5), which originated from the same ancestral gene but allow growth on different substrates [32],[33].
To understand how duplications led to functionally different MalS enzymes, we reconstructed, synthesized, and measured the activity of key ancestral MalS proteins. We used the amino acid (AA) sequences of 50 maltases from completely sequenced yeast species, ranging from Saccharomyces cerevisiae to Pichia and Candida species, for phylogenetic analysis and ancestral sequence reconstruction (see Materials and Methods and Dataset S1). A consensus amino-acid-based phylogenetic tree was constructed using MrBayes [36] under the LG+I+G model with four rate categories (see Figure S1, and see Materials and Methods for details). Trees constructed using MrBayes under other models of sequence evolution (WAG, JTT) generated largely identical results (unpublished data). To further check the robustness of the AA tree inferred by MrBayes, we inferred a maximum likelihood (ML) tree under the LG+I+G model using PhyML (Figure S2) [37]. With the exception of a few recent splits in the topology, the MrBayes and PhyML trees agree, increasing our confidence in the constructed tree. Codon-based tree reconstruction using MrBayes yielded similar results (see further). Additional tests were performed to control for potential long branch attraction (LBA) artifacts, specifically to check the placement of the K. lactis branch as an outgroup to the Saccharomyces and Lachancea clades (see Text S1 and Figures S3, S4, S5, S6).
Next, we reconstructed the AA sequence of the ancestral maltases under several commonly used models of protein evolution (LG, WAG, JTT; see Materials and Methods). All models support roughly the same ancestral protein sequences, increasing our confidence in the reconstructed ancestral sequences. In particular, all models identified the same residues for variable sites within 10 Å of the active center (based on the crystal structure of the Ima1 protein), which are likely relevant sites with respect to enzymatic activity. The residues for a few other sites located further away from the active pocket vary between different models, but differences generally involve biochemically similar AAs (see Table S1).
Synthesis of the ancestral enzymes was based on the reconstructed ancestral sequences obtained with the JTT model. For ambiguous residues (i.e., sites for which the probability of the second-most likely AA is >0.2) within 7.5 Å of the binding pocket, we constructed proteins containing each possible AA, while for ambiguous residues outside 7.5 Å we considered only the most likely AA. There is one ambiguous residue close to the active center in the ancestral proteins ancMalS and ancMal-Ima, namely residue 279 (based on Saccharomyces cerevisiae S288c Ima1 numbering). We therefore synthesized two alternative versions of these proteins, one having G and one having A at position 279. Whereas these alternative proteins show different activities for some substrates, the relative activities are similar and our conclusions are robust. Sequences for these reconstructed enzymes can be found in Dataset S2. In the main figures, we show the variant with the highest confidence. Enzymatic data for all variants can be found in Table S2.
The activity of all resurrected ancestral enzymes was determined for different substrates (see Materials and Methods, Text S1, and Figure 2). The results indicate that the very first ancestral enzyme, denoted as ancMalS, was functionally promiscuous, being primarily active on maltose-like substrates but also having trace activity on isomaltose-like sugars. The activity data presented in Figure 2 show how this promiscuous ancestral protein with relatively poor activity for several substrates evolved to the seven present-day enzymes that show high activity for a subset of substrates, and little or no activity for others. This confirms the existence of two functional classes of MalS enzymes that originated from ancient duplication events. First, Mal12 and Mal32 show activity against maltose-like disaccharides often encountered in plant exudates, fruits, and cereals, like maltose, maltotriose, maltulose, sucrose, and turanose (a signaling molecule in plants). The five MalS enzymes of the second class (Ima1–5), which in fact result from two independent ancient duplication events giving rise to the Ima1–4 and Ima5 clades, show activity against isomaltose-like sugars including palatinose (found in honey [38]) and isomaltose. Differences in hydrolytic activity between members of the same (sub)class are more subtle or even absent, which is not surprising since some of these recent paralogs are nearly identical (Mal12 and Mal32, for example, are 99.7% identical on the AA level).
The more recent ancestral enzymes also show a similar split in activity, with some enzymes (ancMal) showing activity towards maltose-like substrates, and others (ancIma1–4) towards isomaltose-like substrates. Moreover, activity on isomaltose-like sugars (isomaltose, palatinose, and methyl-α-glucoside) changes in a coordinate fashion when comparing different enzymes, and the maltose-like sugars also group together. Careful statistical analysis reveals that the maltose-like group consists of two subgroups (maltose, maltotriose, maltulose, and turanose, on one hand, and sucrose, on the other) that behave slightly different, showing that the enzymes show quantitative differences in the variation of specificity towards these substrates (two-way ANOVA analysis followed by Games-Howell test on log-transformed kcat/Km values; p values can be found in Table S3).
Interestingly, the most ancient ancestral enzymes do not show a clear split in activity towards either maltose-like or isomaltose-like sugars after duplication, and the transition of ancMalS to ancMal-Ima even shows an increase in activity for all substrates. This suggests that (slight) optimization for all substrate classes simultaneously was still possible starting from ancMalS. A clear divergence of both subfunctions occurred later, after duplication of ancMal-Ima, resulting in ancMal and ancIma1–4. AncMal shows a significant increase in activity on maltose-like sugars accompanied by a significant drop in activity on isomaltose-like sugars compared to ancMal-Ima; and the reverse is true for ancIma1–4 (see also Table S3 for exact p values for each enzyme–enzyme comparison on the different sugars tested). Together, this illustrates how, after duplication, the different copies diverged and specialized in one of the functions present in the preduplication enzyme.
In two separate instances, a major shift in specificity is observed, from maltose-like sugars to isomaltose-like sugars (transition from ancIMA5 to IMA5, and from ancMAL-IMA to ancIMA1–4). The shift in activity from ancMAL-IMA to ancIMA1–4 is particularly pronounced. The ancMAL-IMA enzyme hydrolyzes maltose, sucrose, turanose, maltotriose, and maltulose but has hardly any measurable activity for isomaltose and palatinose, whereas ancIMA1–4 can only hydrolyze isomaltose and palatinose (and also sucrose). For the evolution of the maltase-like activity from the ancestral MalS enzyme to the present-day enzyme Mal12, we see a 2-fold increase in kcat and a 3-fold decrease in Km for maltose, indicating an increase in both catalytic power and substrate affinity for this sugar. For the evolution of isomaltase-like activity in the route leading to Mal12, kcat decreases more than 3-fold for methyl-α -glucoside. kcat for isomaltose and palatinose and the affinity for isomaltose and palatinose are so low that they could not be measured (see Table S2 for the exact values of kcat and Km for each enzyme and each sugar; results of two-way ANOVA analysis followed by Games-Howell test comparing log-transformed kcat/Km values for different enzymes on each of the sugars can be found in Table S3).
To further explore the evolution of MALS genes and consolidate the measured activities of the ancestral enzymes, we expressed and purified additional present-day α-glucosidase alleles from other yeast species and measured their activities (Figure 3). We focused primarily on enzymes that are directly related to one of the ancestral proteins but did not undergo any further duplication events, and therefore have a higher probability of having retained a similar activity as their (sub)class ancestor. Indeed, the only present-day MalS enzyme of the yeast L. elongisporus has a broad but relatively weak activity comparable to the very first ancestral MalS enzyme, providing extra support for the accuracy of our ancestral reconstructions. Also in K. lactis, which contains two Mal alleles, one of the paralogs retains the broad specificity of ancMalS. The other paralog (GI:5441460) has a deletion of five AAs close to the active pocket that likely explains the general lack of activity of this enzyme (see Materials and Methods and Figure S7). In contrast, yeasts that show multiple duplication events, like K. thermotolerans and S. cerevisiae, exhibit specialization, with some enzymes showing only activity for maltose-like substrates and others for isomaltose-like substrates. Moreover, the activities (maltase- or isomaltase-like) of homologs in S. cerevisiae and K. thermotolerans derived from the same intermediate ancestor are often similar, except in the IMA5 clade. Here, the K. thermotolerans and S. cerevisiae homologs have very different substrate specificities, indicating species-specific evolutionary trajectories and/or reciprocal paralog loss in the different species (Figures 3 and 4).
Next, we investigated which mutations underlie the observed functional changes. We used the recently resolved crystal structure of Ima1 (pdb entry 3A4A) [34] as a template to study the molecular structure of the enzymes' substrate binding pocket (see Materials and Methods). All enzymes share a highly conserved molecular fold, suggesting that changes in activity or substrate preference are likely caused by mutations in or around the substrate binding pocket. We identified nine variable AA residues within 10 Å of the center of the binding pocket in the various paralogs (Figure 4, right panel). Site-directed mutagenesis and crystallographic studies by Yamamoto et al. confirmed the importance of several of these residues for substrate specificity in the present-day Ima1 protein [39],[40]. In particular, Yamamoto et al. [40] characterized the influence of residues 216-217-218 (Ima1 numbering), which covary perfectly with each other and with the observed substrate specificity shifts across the phylogeny presented in Figure 4. Sequence co-evolution analysis on 640 MAL12 homologs identified another cluster of three co-evolving residues among these nine residues (positions 218, 278, and 279 in Ima1), which we investigate here in detail.
Together with residues 216 and 217, residues 218, 278, and 279 seem to contribute to the activity shift observed in the evolution of Ima1–4 (see Figures 4–6, Figure S8, and Supplementary Information for details). Molecular modeling of the mutations at 218-278-279 on the branch leading to ancIma1–4 (see Figure 4) suggests that the change from alanine to glutamine at residue 279 shifts the binding preference of the pocket from maltose-like to isomaltose-like sugars (Figure 5B–E). The two co-evolving residues at positions 218 and 278 are spatially close to AA 279 and cause subtle structural adaptations that help to better position the Q residue.
To investigate if changes at all three positions are necessary for the observed shift in substrate specificity from ancMAL-IMA to ancIMA1–4 and to investigate the possible evolutionary paths leading to these three interdependent mutations, we synthesized all possible intermediate ancIMA1–4 enzyme variants with mutations at positions 218, 278, and 279. We subsequently expressed, purified, and measured activity of these enzyme variants. Figure 5F depicts the results of these enzyme assays and shows that these residues indeed affect substrate specificity, with the largest shift depending on the A to Q change at position 279, as expected from structural analysis. For one mutational path (GVA to GVQ to SVQ to SMQ), we observe a gradual increase in activity towards isomaltose and palatinose, demonstrating that there is a mutational path that leads to a consistent increase in isomaltase activity without traversing fitness valleys. Moreover, in keep with the stabilizing role of the mutations at positions 218 and 278, the A to Q change at position 279 along this path takes place before the two other mutations at positions 218 and 278 (Figure 5F).
Besides allowing the development of isomaltase activity in the Ima proteins, duplication also permitted further increase of the major ancestral function (hydrolysis of maltose-like sugars) in Mal12 and Mal32. Structural analysis reveals that this increase in maltase activity, from ancMalS to Mal12/32, is due to mutations D307E and E411D (Figure 6G–J). These mutations increase the fit for maltose-like substrates but also completely block the binding of isomaltose-like substrates (Figure 6). Similar to what is seen for the evolution of AncMal-Ima to AncIma1–4, changes that increase the binding stability of one type of substrate cause steric hindrance that prevents binding of the other class of substrates. These signs of incompatibilities between substrates indicate that it is difficult to fully optimize one enzyme for both maltose-like and isomaltose-like substrates, with the highly suboptimal ancMalS being a notable exception. After partial optimization of ancMalS, duplication of ancMAL-IMA likely enabled further optimization of the conflicting activities in separate copies.
Interestingly, the transition from AncMalS to Ima5 shows a similar shift in substrate specificity as the transition of AncMal-Ima to AncIma1–4. However, the residue at position 279, a key factor in the evolution of AncMal-Ima to AncIma1–4, remains unaltered in the evolution of AncMalS to Ima5. Instead, L219, a residue located proximal to position 279, has changed into M219 in the Ima5 enzyme (Figure 6C–F). How can such seemingly very different mutations yield a similar change in substrate specificity?
Structural analysis shows that the L-to-M mutation at position 219 in Ima5 causes a very similar structural change as the G279Q change in AncIma1–4 (Figure 6), indicating that different evolutionary routes may produce a similar shift in activity. In both cases, the evolution of isomaltase-like activity involved introducing a residue that can stabilize isomaltose-like substrates but causes steric hindrance for maltose-like sugars in the binding pocket. Based on the phylogeny of binding pocket configurations and on our enzyme activity tests, this functional shift in the IMA5 clade most likely occurred after a duplication in the common ancestor of S. kluyveri and S. cerevisiae (Figures 3–4).
Next, we investigated the role of selective pressure during the different evolutionary transitions. We used MrBayes to construct a codon-based phylogeny under a GTR codon model of evolution, including 32 MALS genes that share the same nuclear genetic code. The resulting codon-based phylogeny was the same as the AA-based phylogeny generated using the LG+I+G protein model for all 50 sequences, apart from two exceptions in the ancIMA1–4 clade. First, S. mikitae IFO1815 c789 and S. paradoxus N45 branch off separately from S. kudriavzevii IFO1802 c1888 instead of together. Second, S. kudriavzevii IFO1802 c1565 now branches off separately instead of multifurcating with S. mikitae IFO1815 c633 and the branch leading to the S. cerevisiae IMA2–4 genes. Relative branch lengths between genes were similar to the branch lengths calculated under protein models of evolution. The topology of the codon-based tree is presented in Figure 4.
GA Branch analysis [41] identified a branch class with an elevated ω (dN/dS) rate (ω = 0.66) but did not detect branch classes with ω>1 that would be considered strong proof for positive selection (see Materials and Methods and Figure 4). These results, combined with our activity test results and the observed sequence configurations around the active center, suggest, however, that positive selection might have been operating on specific sites in three specific postduplication branches associated with enzyme activity shifts, namely the ancIMA1–4, ancIMA5b, and ancMAL branches, indicated with arrows on Figure 4. We used the modified branch-site model A implemented in PAML to assess positive selection along these branches (see Materials and Methods) [42]. Results are presented in Table S4. For both the ancIMA1–4 and ancIMA5b branches, p values and parameter estimates suggest that a proportion of sites has strongly elevated ω values, consistent with the GABranch results. On the branch from ancMAL-IMA to ancIMA1–4, four sites show signs of positive selection, with a posterior Bayes Empirical Bayes (BEB) probability >0.95, of which two, 216 and 279, are within 10 Å of the active center and known to be important for substrate specificity. On the ancIMA5b branch, four sites show signs of positive selection (BEB>0.95), including again site 216. For ancMAL, the null model (no positive selection) was not rejected at the 95% significance level. Both the corresponding parameter estimates and results of the GABranch analysis, however, suggest relaxation of purifying constraints on this branch.
To get more support for the PAML branch-site test results, we performed an additional analysis using an alternative branch-site method that was recently implemented in the HyPhy package [43]. This method identified in total seven branches that possibly experienced positive selection: ancIMA1–4 (p<0.0001), ancIMA5b (p = 0.0232), ancMALS (p = 0.0228), S. kluyveri SAKL0A05698g (p<0.0001), K. thermotolerans GI: 255719187 (p<0.0001), the branch leading from ancIMA5 to the ancIMA5b branch (p = 0.0168), and finally the branch leading up to S. cerevisiae IMA2, IMA3, IMA4, and YPS606 within the ancIMA1–4 clade (p = 0.0353). In other words, the ancMALS, ancIMA1–4, and ancIMA5b branches are suggested to have evolved under positive selection, together with four other branches. The branch-site method implemented in HyPhy currently does not allow the identification of specific sites that may have evolved under positive selection on these branches.
Together, our analyses indicate that some residues near the active pocket, in particular the key residues 216 and 279 that determine substrate specificity (see above), may have experienced positive selection in the postduplication lineages leading to isomaltose-specific enzymes. It should be noted, however, that the specificity and sensitivity of the currently available methods for detecting positive selection, in particular branch-site methods, is heavily debated [42],[44]–[47]. Possible pitfalls include fallacies in the assumption that synonymous substitutions are neutral, a reported increase in the number of false positives due to sampling errors when the number of (non)synonymous substitutions and sequences is low, and potential inadequacies in the null and alternative models that are being compared, leading to difficulties with completely ruling out other explanations for perceived positive selection. For these reasons, the positive selection test results reported here should be approached as indications rather than definitive proof.
The previous results show how duplication of a promiscuous ancestral enzyme with limited activity towards two substrate categories allowed the evolution of separate enzyme clades that each show increased activities for a specific subset of substrates. The functional diversification of the different clades ensures their retention. However, why are recent, near-identical duplicates such as MAL12 and MAL32 conserved?
To investigate if selective pressure might protect the MAL12/MAL32 duplicates, we determined the fitness effect of inactivating each of them. The results in Figure S9 show that strains lacking just one of the MAL12 and MAL32 paralogs show a considerable fitness defect compared to a wild-type strain when grown on maltose. These results suggest that gene dosage may play a primary role in preserving these recent paralogs [6]. Dosage effects increasing maltase and/or isomaltase activity may also have played a role after the earliest MALS duplications, before the duplicates were optimized for different activities.
One of the major issues in the field of molecular evolution is the plethora of theoretical models and variants of models concerning the evolution of gene duplicates, with few of the claims supported by solid experimental evidence. On many occasions, inherent properties of the evolutionary process make it extremely hard to find or generate experimental evidence for a given model. However, recent developments in genome sequencing, evolutionary genomics, and DNA synthesis open up exciting possibilities. Using these new opportunities, we were able to resurrect ancient MALS genes and the corresponding enzymes and provide a detailed picture of the evolutionary forces and molecular changes that underlie the evolution of this fungal gene family. The MALS gene family is an ideal model for the study of duplicate gene evolution, since it underwent several duplication events and encodes proteins for which we could accurately measure different activities. The availability of multiple fungal genome sequences provided sufficient data to robustly reconstruct ancestral alleles and study the selective forces that propelled divergent evolution of the paralogs. Additionally, the existence of a high-quality crystal structure of one of the present-day enzymes made it possible to predict the functional effects of mutations and to study the mechanistic basis of suspected adaptive conflicts between the maltase-like and isomaltase-like subfunctions.
Our results paint a complex and dynamic picture of duplicate gene evolution that combines aspects of dosage selection and sub- and neofunctionalization (see Figure 7). The preduplication ancMalS enzyme was multifunctional and already contained the different activities found in the postduplication enzymes (the basic idea of subfunctionalization), albeit at a lower level. However, the isomaltase-like activity was very weak in the preduplication ancestor and only fully developed through mutations after duplication (increase of kcat/Km with one order of magnitude for isomaltase-like substrates from ancMalS to Ima1), which resembles neofunctionalization. The ancestral maltase-like activity also improved substantially but to a lesser extent (factor 6.9 on average from ancMalS to Mal12), which therefore perhaps fits better with the subfunctionalization model. Moreover, our activity tests on Mal12/Mal32 mutants indicate that gene dosage may also have played a role in preserving MALS paralogs, especially right after duplication. This may not only have been the case for the recent MAL12–32 and IMA3–4 duplications but also for more ancient duplications involving multifunctional ancestors. In summary, whereas the classical models of dosage, sub-, and neofunctionalization are helpful to conceptualize the implications of gene duplication, our data indicate that the distinction between sub- and neofunctionalization is blurry at best and that aspects of all three mechanisms may intertwine in the evolution of a multigene family.
Although it is difficult to classify our results decisively under one of the many models of evolution after gene duplication, most of our findings agree with the predictions of the “Escape from Adaptive Conflict” (EAC) model [5],[16],[17],[19], a co-option-type model in which duplication enables an organism to circumvent adaptive constraints on a multifunctional gene by optimizing the subfunctions separately in different paralogs. The EAC model makes three key predictions: (i) the ancestral protein was multifunctional, (ii) the different subfunctions could not be optimized simultaneously in the ancestral protein (or at least not in an evolutionarily easily accessible way), and (iii) after duplication, adaptive changes led to optimization of the different subfunctions in separate paralogs [13],[16],[48]. In general, our findings fit with these predictions: (i) we find that several of the ancestral preduplication maltase enzymes (ancMALS, ancMAL-IMA, and ancIMA5) were multifunctional; (ii) we provide evidence, through molecular modeling and activity tests of present-day enzymes, ancestors, and potential intermediates, that the maltase and isomaltase functions are difficult to optimize within one protein (but see also below); and (iii) we find that duplication resolved this adaptive conflict, and we find indications that positive selection might have driven key changes that optimized the minor isomaltase-like activity of the preduplication enzyme in one paralog, while the major maltase-like activity was further optimized in the other paralog.
Figure 2 and the statistical analysis in Table S3 indicate that the activity of the different enzymes changes significantly at certain points along the evolutionary path. Interestingly, the overall image that emerges suggests that the enzymes developed activity towards either maltose-like or isomaltose-like sugars, but not both. This pattern is most clear in the evolution of ancMal-Ima to ancMal and ancIma1–4. The postduplication improvement of the different activities present in the ancestral allele, with each of the new copies displaying increased activity for one type of substrate and concomitantly decreased activity towards the other substrate class, could be indicative of trade-offs in the evolution of the MALS gene family. However, the word “trade-off” implies that the two incompatible functions are both under selection, which is difficult to prove for the ancient enzymes. Moreover, our results indicate that for the ancient ancMalS enzyme, it is possible to simultaneously increase the activity towards both maltose-like and isomaltose-like substrates. Together, our analyses show that it is possible to optimize (to a certain extent) one function of a multifunctional enzyme without significantly reducing the other (minor) activity. However, analysis of the complete evolutionary path and molecular modeling of the active pockets of the enzymes shows that full optimization of both functions in a single enzyme is difficult to achieve, due to steric hindrance for one substrate class when fully optimizing the active pocket for binding of the other substrate type. This problem can be most easily overcome by duplication of the enzyme, allowing optimization of the different subfunctions in different paralog copies, as can be seen in the transition of ancMal-Ima to ancMal and ancIma1–4.
While most aspects of our data fit with the EAC model, some results are more difficult to reconcile with the EAC theory. Specifically, one of the pillars of the EAC model is that positive selection drives the specialization of both paralogs after duplication. While our data demonstrate that duplication of ancMAL-IMA has led to optimization of both subfunctions in different duplicate lineages (maltase-like activity in ancMAL and isomaltase-like activity in ancIMA1–4), our selection tests only reveal indications of positive selection in the ancIMA1–4 lineage but not in the ancMAL lineage. Moreover, as discussed above, positive selection is difficult to prove [44],[49], and we cannot exclude the possibility of both false positive and false negative artifacts.
Recently, some other likely examples of the EAC mechanism have been described [16],[17],[50]–[52]. These studies also presented plausible arguments for ancestral multifunctionality, adaptive conflict, and/or adaptive optimization of subfunctions in different paralogs, but as in the present case, none could provide strong experimental evidence for all three predictions made by the EAC model [48],[53]. Instead of classifying the evolutionary trajectory of particular gene duplicates into one of the many models for gene duplication, it may prove more useful to distill a more general picture of duplicate evolution across a gene family that includes aspects of dosage selection, and sub- and neofunctionalization, like the one depicted in Figure 7.
Our study is the first to investigate multiple duplication events in the same gene family in detail. Interestingly, we found that evolution has taken two different molecular routes to optimize isomaltase-like activity (the evolution of ancMAL-IMA to ancIMA1–4 and ancIMA5 to IMA5). In both cases, only a few key mutations in the active pocket are needed to cause shifts in substrate specificity. Some of these key mutations exhibit epistatic interactions. For example, the shift in substrate specificity occurring on the path from ancMAL-IMA to ancIMA1–4 depends in part on mutations at three co-evolving positions (218, 278, and 279), but only one mutational path (279-218-278) shows a continuous increase in isomaltase-like activity. Interestingly, there is also a different path in the opposite direction (218-279-278) that shows a continuous increase in the ancestral maltase-like activity. This implies that the complex co-evolution at these three positions may be reversible. Interestingly, a recent study of the evolutionary history of plant secondary metabolism enzymes also identified AA changes that appear to be reversible [51], in contrast to the situation for, for example, glucocorticoid receptor evolution, where evidence was found for an “epistatic ratchet” that prevents reversal to the ancestral function [54].
It is tempting to speculate that complex mechanisms like those driving the evolution of the MALS gene family may be a fairly common theme. Many proteins display some degree of multifunctionality or “promiscuity” [55]–[57], just like the ancestral ancMal enzyme. Moreover, directed “in vitro” protein evolution experiments have shown that novel protein functions often develop from pre-existing minor functions [58],[59]. Although the different functions within an enzyme often exhibit weak trade-offs, allowing optimization of the minor activity without affecting the original function of the enzyme [55],[59],[60], this may not always be the case. If there are stronger trade-offs between different subfunctions, duplication may enable the optimization of the conflicting functions in different paralogs.
While it is difficult to obtain accurate dating of the various duplication events, the duplication events studied here appear to postdate the divergence of Saccharomyces and Kluyveromyces clades, estimated to have occurred 150 mya [61], but predate the divergence of Saccharomyces and Lachancea and the yeast whole genome duplication, about 100 mya. MALS diversification may thus have happened around the appearance and spread of angiosperms (Early Cretaceous, between 140 and 100 mya [62]) and fleshy fruits (around 100 mya). Tentative dating results can be found in Table S6, but these should be approached with caution (see Text S1). The major shift in the earth's vegetation caused by the rise of the angiosperms almost certainly opened up new niches, and it is tempting to speculate that duplication and diversification of the MALS genes may have allowed fungi to colonize new niches containing sugars hydrolyzed by the novel Mal (Ima) alleles. In other words, the availability of novel carbon sources in angiosperms and fleshy fruits could have provided a selective pressure that promoted the retention of MALS duplicates and the ensuing resolution of adaptive conflicts among paralogs.
In total, the nucleotide and protein sequences of 169 extant maltases were collected for yeast species ranging from Saccharomyces cerevisiae to Pichia and Candida species. For Kluyveromyces thermotolerans, Saccharomyces kluyveri, and Kluyveromyces lactis, sequences were downloaded from Génolevures (www.genolevures.org). Sequences for many of the Saccharomyces cerevisiae and Saccharomyces paradoxus genes were obtained from the sequence assemblies provided by the Wellcome Trust Sanger Institute (http://www.sanger.ac.uk/research/projects/genomeinformatics/sgrp.html). All of the remaining extant maltase sequences were downloaded from NCBI (www.ncbi.nlm.nih.gov/). Sequences with greater than 92% pairwise protein sequence similarity to other sequences in the dataset were removed to reduce the phylogenetic complexity. All seven Saccharomyces cerevisiae S288c alleles were kept, however, yielding a final dataset of 50 sequences (see also Dataset S1).
We used ProtTest 2.4 [63] to score different models of protein evolution for constructing an AA-based phylogenetic tree. All possible models with all improvements implemented in the program were taken into account. An initial tree was obtained by Neighbor-Joining (BioNJ), and the branch lengths and topology were subsequently optimized for each evolutionary model independently. The LG+I+G model came out as best with a substantial lead over other protein models using −lnL, AIC, and AICc selection criteria (AICc = 43,061.26 with AICw = 1.00, while the second best model was WAG(+I+G) with AICc = 43,158.00 and AICw = 0.00). Consequently, an AA-based phylogeny for the 50 sequences was determined using MrBayes 3.1.2 [36] with a LG invariant+gamma rates model (four rate categories). Since the LG model is not implemented by default in MrBayes, we used a GTR model and fixed the substitution rate and state frequency parameters to those specified by the LG model. The BMCMC was run for 106 generations, sampling every 100 generations, with two parallel runs of four chains each. A burn-in of 2,500 samples was used, and the remaining 7,501 samples were used to construct a 50% majority-rule consensus phylogeny (Figure S1). The AWTY program [64] was used to check proper MCMC convergence under the given burn-in conditions. MrBayes AA tree constructions were also performed under other evolutionary models (WAG, JTT). Additional tests were performed to exclude Long Branch Attraction (LBA) artifacts (see Text S1). We also inferred a maximum likelihood (ML) tree using PhyML under the LG+I+G model with four rate categories [37]. The initial tree was again obtained by BioNJ; tree topology, branch lengths, and rate parameters were optimized in a bootstrap analysis with 1,000 replicates.
We also used MrBayes to construct a codon-based phylogeny, using a GTR codon model of evolution. The original dataset of 50 sequences contained 18 sequences for species that employ the alternative yeast nuclear genetic code (all of them outgroup species). These sequences were removed from the dataset, resulting in a reduced dataset of 32 sequences. The codon alignment was obtained by translating the AA alignment obtained earlier. BMCMC analysis and consensus phylogeny construction were performed as described above for the AA trees. We contrasted models that did and did not allow for ω rate variation (i.e., the “Equal” versus “M3” codon model in MrBayes). AWTY analysis indicated that the latter was not able to converge properly, so we used the results of the Equal model.
The PAML package [65] was used to infer the posterior AA probability per site in the ancestors of interest under several commonly used models of protein evolution (LG, WAG, JTT), using the corresponding Bayesian consensus phylogenies. Both marginal and joint probability reconstructions were performed. The marginal reconstructions are presented in Table S1. Protein sequences resulting from marginal reconstructions under the JTT model were used to synthetize ancestral enzymes.
We performed tests for positive selection on the codon-based phylogeny obtained as described above. Various branch methods and branch-site methods included in the PAML [65] and HyPhy [66] packages were used.
Co-evolving residues in the MALS gene family were detected using the framework described by [71]. The NCBI Blast server was used to collect Saccharomyces cerevisiae S288c MAL12 maltase homologs, with an E-value <10e-70, resulting in a set of 1,211 sequences. Proteins were removed that were shorter than 400 AAs, longer than 800 AAs, and more than 95% similar to another protein in the dataset. This resulted in a dataset of 640 maltase homologs with sequence similarity >40% compared to Saccharomyces cerevisiae S288c MAL12. These sequences were aligned with MAFFT and only the most reproducible residue–residue couplings (present in at least 90% of the splits) were retained.
A two-way ANOVA using log-transformed kcat/Km (to obtain values that are normally distributed) as the variable and the different enzymes and sugars as factors was performed using the aovSufficient function from the HH package in R. This analysis was followed by pairwise comparisons using the Games-Howell post hoc test (since samples had unequal variances, as demonstrated by Levene's test). Results can be found in Table S3.
Ancestral maltase genes were synthesized and cloned into vectors for overexpression in E. coli host cells by GENEART (www.geneart.com). Sequences can be found in Table S1 and Dataset S2. The inferred protein sequences were reverse translated in order to optimize their codon usage for E. coli. These gene sequences were synthesized including an N-terminal 6xHis tag (ATGGGCAGCAGCCATCATCATCATCATCACAGCAGCGGCCTGGTGCCGCGCGGCAGCCAT) and 5′UTR (TCTAGAAATAATTTTGTTTAACTTTAAGAAGGAGA TATACC), cloned into in-house vectors at GENEART, and then sequenced. Subsequently, the inserts were subcloned into pET-28(a) vectors (Merck) via XbaI/XhoI sites. All of the overexpression plasmids were transformed into E. coli strain BL21*. All E. coli strains were grown under selection in standard LB media+kanamycin (Sigma Aldrich). Details on protein expression and purification can be found in Text S1.
The activities of the purified ancestral and present-day enzymes were determined by measuring glucose release from α-glucosides (maltose, sucrose, turanose, maltotriose, maltulose, isomaltose, palatinose, and methyl-α-glucoside) using a standard glucose oxidase/peroxidase coupled reaction. All sugars were purchased in their highest available purity. More information on the purchased sugars as well as a detailed protocol can be found in Text S1.
For each protein and substrate, the reaction velocity (amount of glucose produced per time unit) was determined. Subsequently, reaction velocities normalized by enzyme concentration as a function of substrate concentration were plotted and fitted using a nonlinear least squares fitting routine (Levenberg-Marquardt algorithm) both to Michaelis-Menten-style kinetics and Hill-style kinetics:The data fits were compared using an F statistic (i.e., Michaelis-Menten is a specific case of Hill kinetics with n = 1), and the Michaelis-Menten model was rejected with α = 5%. From these fits, errors (standard deviations) were computed by jack-knifing over the individual substrate concentrations (12 data points in total). For numerical optimization, code was written in Python using NumPy. Model parameters of interest, along with their associated errors, were extracted (i.e., kcat and Km; see Table S2). Processing (http://processing.org) was used to draw Figure 2 and Figure 5F by writing code. Enzyme efficiencies were plotted (as vertical lines) at different points on the tree, and values between were interpolated.
Relative Malthusian fitness was determined by competing unlabelled WT (KV1042), mal12 (KV1151), and mal32 (KV1153) strains against a reference strain (KV3261), expressing GFP from the TDH3p. Details can be found in the Supporting Information section.
All molecular modeling was performed using the MOE 2010.10 package (The Molecular Operating Environment, The Chemical Computing Group, Montréal, Canada). The recently released crystal structure of the Ima1 protein (pdb entry: 3A4A), with glucose in the binding pocket, was used as a template to construct the different MALS homology models, with implementation of the Amber99 force field. Since the AAs contacting this glucose molecule are conserved within the different MALS subgroups, this glucose was used to model the different sugar substrates within the active sites, using the MOE 2010.10 ligX implementation.
Full methods and any associated references can be found in the Supporting Information section.
|
10.1371/journal.ppat.1004522 | Structure and Specificity of the Bacterial Cysteine Methyltransferase Effector NleE Suggests a Novel Substrate in Human DNA Repair Pathway | Enteropathogenic E. coli (EPEC) and related enterobacteria rely on a type III secretion system (T3SS) effector NleE to block host NF-κB signaling. NleE is a first in class, novel S-adenosyl-L-methionine (SAM)-dependent methyltransferase that methylates a zinc-coordinating cysteine in the Npl4-like Zinc Finger (NZF) domains in TAB2/3 adaptors in the NF-κB pathway, but its mechanism of action and other human substrates are unknown. Here we solve crystal structure of NleE-SAM complex, which reveals a methyltransferase fold different from those of known ones. The SAM, cradled snugly at the bottom of a deep and narrow cavity, adopts a unique conformation ready for nucleophilic attack by the methyl acceptor. The substrate NZF domain can be well docked into the cavity, and molecular dynamic simulation indicates that Cys673 in TAB2-NZF is spatially and energetically favorable for attacking the SAM. We further identify a new NleE substrate, ZRANB3, that functions in PCNA binding and remodeling of stalled replication forks at the DNA damage sites. Specific inactivation of the NZF domain in ZRANB3 by NleE offers a unique opportunity to suggest that ZRANB3-NZF domain functions in DNA repair processes other than ZRANB3 recruitment to DNA damage sites. Our analyses suggest a novel and unexpected link between EPEC infection, virulence proteins and genome integrity.
| Pathogens often manipulate host functions by posttranslational modifications such as ubiquitination and methylation. The NF-κB pathway is most critical for immune defense against infection, thereby frequently targeted by bacterial virulence factors. NleE, a virulence effector from EPEC, is a SAM-dependent methyltransferase that modifies a zinc-finger cysteine in TAB2/3 in the NF-κB pathway. NleE is not homologous to any known methyltransferases. We present the crystal structure of SAM-bound NleE that shows a novel methyltransferase fold with a unique SAM-binding mode. Computational docking and molecular dynamics simulation illustrate a structural and chemical mechanism underlying NleE recognition of the NZF and catalyzing site-specific cysteine methylation. Subsequent substrate specificity analyses identify an N-terminal region in TAB3 required for efficient NleE recognition as well as another NZF protein ZRANB3 being a new substrate of NleE. NleE-catalyzed cysteine methylation also disrupts the ubiquitin chain-binding of ZRANB3-NZF domain, providing new insights into ZRANB3-NZF functioning in DNA damage repair. These results reinforce the idea of harnessing bacterial effectors as a tool for dissecting eukaryotic functions.
| NF-κB signaling plays a central role in defending against bacterial infection [1], [2]. The NF-κB signaling initiates innate immune responses and inflammation via a myriad of pathogen-recognition or cytokine receptors. These receptors generate ubiquitin-chain signals that are directly recognized by the TAB2/3 adaptors, thereby activating the TAK1 and IKK kinase cascade, leading to transcription of genes involved in immune defense. EPEC and the related enterohaemorrhagic E. coli (EHEC) block NF-κB signaling using virulence effector proteins injected into host cells by the type III secretion system (T3SS). The NleE effector, conserved in Shigella and Salmonella, plays a major role in EPEC suppression of the NF-κB signaling in cell culture infection [3], [4], [5]. We recently discovered that NleE is a SAM-dependent methyltransferase that modifies a cysteine in the NZF domains of TAB2/3, thereby disrupting ubiquitin-chain sensing of TAB2/3 and abolishing NF-κB-mediated proinflammatory responses [6].
Protein methylation is of great importance in a plethora of cellular processes including biosynthesis, signal transduction, protein repair, chromatin regulation and gene silencing [7]. SAM-dependent methyltransferases are diverse in their primary sequence, three dimensional structure and SAM-binding mode, and have been classified into five different families (Class I-V) [8]. The five families of methyltransferases generally catalyze lysine or arginine methylation. NleE-catalyzed cysteine methylation of TAB2/3 is the first example of enzyme-catalyzed protein cysteine methylation, representing a novel mechanism in regulating signal transduction in eukaryotes. NleE harbors no sequence homology to known methyltransferases. The structural basis for NleE methyltransferase activity and substrate specificity are unknown.
Here we determine the crystal structure NleE-SAM complex, which reveals a novel methyltransferase fold and a unique mode of SAM binding. Molecular dynamic simulation of the docked NleE-SAM-NZF complex indicates that Cys673 in TAB2-NZF is structurally and energetically favorable for attacking the SAM. Profiling of a large number of zinc fingers identifies ZRANB3 as a new NleE substrate. ZRANB3 is recruited to damaged DNA replication forks and functions in maintaining genome integrity [9], [10], [11]. NleE-methylated ZRANB3-NZF domain lost the ubiquitin chain-binding activity, suggesting an unexpected link between EPEC infection, virulence proteins and genome integrity. These structural and functional analyses suggest that NleE may target ZRANB3 or other zinc-finger proteins for cysteine methylation in promoting bacterial virulence.
Due to the lack of an antibody capable of recognizing methylated cysteine, we developed a back-methylation assay by examining the sensitivity of TAB2 purified from NleE-transfected mammalian cells to in vitro re-methylation by NleE (Figure 1A). Flag-TAB2 from 293T cells was efficiently re-methylated, whereas that from cells co-transfected with wild-type NleE resisted further in vitro methylation by NleE, suggesting full methylation of cellular TAB2 by transfected NleE. Furthermore, tandem mass spectrometric analysis of TAB2/3 from infected 293T cells confirmed the methylation modification as a Cys673-methylated peptide from Flag-TAB2 was detected upon infection with wild-type EPEC but not the ΔnleE strain (Figure 1B). Complementation of ΔnleE strain with NleE expressed from a high-copy plasmid resulted in complete methylation of Cys673 (Figure 1B). This provides direct evidences that NleE carries out cysteine methylation of TAB2/3 during EPEC infection. In addition to TAB2/3, components of the linear ubiquitin chain assembly complex (LUBAC), HOIP, HOIL-1L and Sharpin, also contain NZF domains and play important roles in NF-κB signaling [12], [13], [14]. Consistent with our previous in vitro data, the HOIL-1L and Sharpin-derived tryptic peptides bearing the cysteine corresponding to Cys673 in TAB2 were not methylated even when the infection was performed with the NleE-proficient EPEC strain (Figure 1C and 1D). These data support that NleE inhibition of NF-κB signaling results from its specific targeting TAB2/3-NZF domains for cysteine methylation.
To understand the mechanism of NleE function, we attempted to determine its crystal structure. Wild-type NleE yielded poor crystals, but NleE K181A protein, out of 15 Lys-to-Ala mutants designed to improve the crystallization [15], produced sufficient-quality crystals in the C2 space group. A model obtained from 2.6-Å diffraction data collected on the selenomethionine (Se-Met) protein was further refined and a final resolution of 2.3 Å was achieved (Table S1). The structure shows that the mutated K181A is exposed and located at the interface of crystal contacts (Figure S1). Despite the presence of four NleE in an asymmetric unit (chain A–D) (Figure S2A), NleE was exclusively a monomer in solution as judged by gel filtration chromatography analysis. The size of the buried surface area formed between different chains ranged from 590 Å2 to 1722 Å2 (Figure S2B). PISA (http://www.ebi.ac.uk/pdbe/pisa/) analysis of protein interface present in the crystal also suggested that the NleE tetramer is unlikely to be stable in solution (Figure S2C), indicating that the tetrameric assembly of NleE in the asymmetric unit results from crystal packing likely with no physiological relevance. No meaningful structural difference was found among the four molecules and therefore only chain A was analyzed hereinafter.
The structure of NleE (residues 21–220: residues 1–20 and 220–224 lacked electron density) adopts α/β doubly wound topology with a central three-stranded anti-paralleled β-sheet (β1–β3) sandwiched by α-helices (α1–α10) (Figure 2A). β1–β3 are arranged in the left-right-middle order, which, together with the flanking α-helices (α8–α10), generates a deep and narrow cavity on the left side of NleE. The SAM molecule fits snugly into the cavity (Figure 2B), which buries 2374 Å2 solvent accessible surface area, corresponding to 76% of the total surface area of SAM and leaving the rest of 24% exposing to the solvent.
We first analyzed the structural details of SAM binding in NleE. The interior of the SAM-binding cavity is filled with hydrophobic side chains, but polar interactions appear to play a key role in riveting the SAM into the cavity (Figure 2C). Specifically, the α-carboxylic group of the amino acid moiety of the SAM is coordinated by the side chain of Arg107 situated on a loop connecting β1 and α6. The hydroxyl group of the ribose of the SAM is hydrogen-bonded with the side-chain carboxylic group of Glu191. The adenine ring of SAM is oriented by the aromatic ring of Tyr212 residing at α10 (residues 206–219) through a π-π stacking interaction. This explains the complete functional loss of the NleEΔ6 mutant [4], [6] as deletion of 209IDSYMK214 is expected to disrupt the α10. The interactions embed the SAM at the bottom of the cavity with a narrow opening slit, through which the buried ligand presents its methylthio in a direction favorable for the SN2 methyltransfer.
Mutation of Arg107, Glu191 or Tyr212 in NleE all abolished in vitro methylation of GST-TAB2-NZF, revealed by native gel mobility analysis of the NZF domain (Figure 3A). When exogenous SAM was added, NleE-E191A and Y212A showed activity equal to wild-type NleE, whereas NleE-R107A and NleEΔ6 remained completely inactive (Figure 3A). Consistently, NleE-Y212A and NleE-E191A were partially impaired in transferring 3H-methyl from radiolabeled SAM onto TAB2-NZF; NleE-R107A showed absolutely no activity in this assay (Figure 3B). Thus, Arg107 is most critical for SAM binding while NleE-Y212A and NleE-E191A are severely impaired. The in vivo activity of NleE-R107A, E191A or Y212A mutants in inhibiting NF-κB in transfected 293T cells was concordant with their in vitro methylation activity (Figure 3C). Further supporting the structural analysis, the NleE-R107A mutant, when complemented into EPEC ΔnleE strain, failed to restore methylation of TAB2 in bacteria infected cells (Figure 3D).
Among the five families of SAM-dependent methyltransferases [8], the most abundant Class I has a central seven-stranded β-sheet and a GxGxG SAM-binding motif; Class II has long β-strands and a shallow groove with a RxxxGY SAM-binding motif; Class III is a homodimer with each monomer adopting an αβα structure and the SAM moiety bound between the two monomers; Class IV (the SPOUT family of RNA methyltransferases) bears a C-terminal SAM-binding knot structure; Class V contains a SET-domain SAM binding motif composed of three small β-sheets. Remarkably, the overall architecture of NleE does not resemble any of the five families (Figure 4A), thus representing a completely novel class of methyltransferases. Moreover, the conformation of the SAM in NleE is much different from other methyltransferases as reflected in the adenosine and methionine conformations (Figure 4B and 4C).The adenine base in NleE-bound SAM is characterized by a C4′-C1′-N9-C4 dihedral angle of 60°, significantly smaller than that in other methyltransferase-bound SAM or S-adenosylhomocysteine (SAH) (Figure 4D). The O4′-C4′-C5′-Sδ dihedral angle in NleE-bound SAM is 160°, comparable to the ∼180° in that in Class I methyltransferases, whereas this dihedral angle is approximately −90° in Class II-IV and 80° in Class V methyltransferases (Figure 4D). According to the C4′-C5′-Sδ-Cγ dihedral angle, the SAM/SAH molecules in NleE, Class I and II methyltransferases adopt a relatively extended conformation while those in other three classes adopt a more compact structure (Figure 4C).
NleE specifically methylates Cys673 in TAB2 (Cys692 in TAB3) among the four Zn-coordinating cysteines in TAB2/3-NZF domains despite that they are predicted to be chemically inert due to protection by hydrogen bonds [16] (Figure S3A). In the TAB2-NZF structure (PDB ID code: 3A9J), Cys673 and Cys687 are largely exposed, whereas Cys670 and Cys684 are completely buried (Figure S3B). To understand the mechanism of site-specific methylation by NleE, a hierarchical protein-protein docking approach with enforced distance restraints between the methyl group of SAM and the sulfur of Cys673/Cys687 was employed and molecular dynamics (MD) simulation was performed. The Cys687-restricted simulation showed a dramatic motion with pronounced root-mean-square deviation (RMSD) values, high interaction energy and a large distance from Cys687 to the methyl donor (Figure 5A and Figure S3C). In contrast, the Cys673-restricted simulation showed a relatively limited motion and lower energy with a close distance from Cys673 to the methyl donor. Thus, a most energetically favorable and structurally stable NleE-SAM-NZF complex was in silico modeled (Figure 5B), which clearly showed that Cys673 is the most favorable substrate residue.
We previously observed that deletion of the NZF from TAB3 (TAB3ΔNZF) does not affect its binding to NleE [6] (Figure 6A, and Figure S4A and S4B). This suggested that specific recognition by NleE requires another region in TAB3. Progressive truncations from both the C and N termini of TAB3ΔNZF identified residues 52–194 as the minimal fragment sufficient for binding to NleE in the yeast two-hybrid interaction assay (Figure 6A and 6B, and Figure S4). Co-immunoprecipitation assay in transfected 293T confirmed that residues 52–194 of TAB3, in contrast to the NZF alone, were competent in efficient binding to NleE (Figure 6C). Thus, binding of the N-terminal region in TAB3 (possibly also TAB2) may serve as a docking mechanism for recognition and methylation of TAB2/3-NZF by NleE. However, it is worth noting here that this region, involved in docking TAB3 onto NleE, does not appear to be sufficient for NleE methylating of other NZF domain as an NleE-resistant ZRANB2-NZF (see below) remained unmodified by NleE even when positioned in place of TAB3-NZF in the TAB3 ΔNZF construct (Figure S5).
Given that Zn coordination is required for cysteine methylation by NleE, we investigated whether other Zn fingers could also be a substrate of NleE. Among a total of more than 50 Zn fingers including C2H2, RING, RBCC/TRIM, FOG, PHD, as well as all the 13 NZF C4 fingers (Table S2 and Figure S6), NleE efficiently methylated the NZF domain of ZRANB3 with similar efficiency to that of the NZF domains of TAB2/3 and yeast Vps36 (Figure 7A). NleE did not modify Npl4, Sharpin, HOIP, HOIL-1L, Trabid-NZF1/2/3 and ZRANB2-NZF among the NZF subfamily [6] (Figure 7A and Table S2). Tandem mass spectrometry analysis identified the second cysteine in ZRANB3 (Cys630) being the methylation site, which echoes the situation with TAB2/3-NZF domains (Figure S7). Full-length ZRANB3 purified from 293T cells was also a robust substrate in the in vitro methylation assay (Figure 7B). In the back-methylation assay, recombinant NleE failed to methylate ZRANB3 purified from NleE-expression 293T cells (Figure 7B), suggesting a full methylation of ZRANB3 in transfected mammalian cells. Agreeing with that reported in previous studies [9], [11], GST-ZRANB3-NZF could bind to polyubiquitin chains of Lys63, Lys48, as well as tetra ubiquitin with linear linkage (Figure 7C). However, methylation by NleE was found to abolish the binding of ZRANB3-NZF to all ubiquitin chains. NleE could also abolish the ubiquitin chain binding of full-length ZRANB3 in transfected 293T cells, whereas the methyltransferase-deficient NleEΔ6 mutant failed to do so (Figure 7D and 7E). In EPEC-infected cells, the majority of Flag-ZRANB3 appeared to be methylated in an NleE-dependent manner (Figure 7F). The diminished ZRANB3 methylation in ΔnleE EPEC-infected cells could be fully restored by wild-type NleE but not the SAM-binding deficient R107A mutant (Figure 7F). Consistently, NleE could completely disrupt the ubiquitin chain-binding ability of Flag-ZRANB3 during EPEC infection, which also required Arg107 in NleE (Figure 7G). Thus, ZRANB3, like TAB2/3, is a bona fide target of NleE methyltransferase activity under physiological conditions.
Overexpression of ZRANB3 in the presence or absence of NleE did not affect NF-κB activation (Figure S8). ZRANB3 has 1, 077 amino acids; its N-terminal half is a helicase domain and the C-terminus harbors multiple domains including the NZF domain. Recent studies suggest that ZRANB3 is localized in nucleus and functions in DNA replication stress response to maintain genome stability [9], [10], [11]. ZRANB3 is recruited to damaged replication forks to promote fork restart. We also observed that EGFP-ZRANB3 was recruited to laser-generated stripes where DNA damage occurred (Figure S9). Notably, co-expression of NleE, which was found distributed in both the cytoplasm and nucleus (Figure S10) and resulted in complete methylation of ZRANB3 (Figure 7B), did not affect ZRANB3 recruitment to DNA damage sites (Figure S9). It has been proposed that damage-induced recruitment of ZRANB3 is mediated by its binding to K63-linked polyubiquitin chains on PCNA, a protein playing a central role in promoting faithful DNA replication [9], [11]. Our results suggest that another structural region in ZRANB3 is more likely responsible for its recruitment to DNA damage sites and the NZF domain-mediated polyubiquitin-chain binding probably participates in other aspects of ZRANB3 function that remains to be defined. It is also worth noting here that the activity of NleE offered us a unique approach to achieve functional disruption of a single domain within a large multiple-domain protein.
NleE is a unique SAM-dependent methyltransferase in catalyzing cysteine methylation. The structure of NleE bears an overall Rossmann-like fold and more resembles that of Class I SAM-dependent methyltransferase, but its SAM-binding mode and conformation are completely different. This indicates an independent evolution of the two sub-lineages within the methyltransferase family and highlights the convergent evolution of bacterial virulence activity. The unique fold of NleE expands the repertoire of SAM-dependent methyltransferases and highlights the convergence on methylation chemistry from different three dimensional folds.
Cysteine methylation is rare; a recent example is methylation of Cys39 in Rps27a, a nonessential yeast ribosomal protein [17]. Rps27a is structurally similar to the N-terminal domain of Ada protein and Cys39 is also one of the four Zn-coordinating cysteines, suggesting a similar non-enzymatic methyl transfer. This supports that Zn coordination facilitates methyl transfer onto the cysteine thiol. The high abundance of zinc finger [18] also indicates that other zinc-finger motifs might be potential methylation targets of some methyltransferases. A recent study on a radical SAM methyltransferase RlmN shows methylation of a cysteine not bound to the zinc [19], [20], further highlighting a chemical diversity of cysteine methylation.
In addition TAB2/3, we now also identify as ZRANB3 as another efficient methylation substrate of NleE (EPEC 2348/69 strain). As NleE homologues are also present in other pathogenic E. coli strains as well as Salmonella and Shigella spp., it is possible that different NleE homologues, produced by different bacterial pathogens, may target different host substrates. NleE efficiently methylates ZRANB3-NZF and abolishes its ubiquitin-chain binding but does not affect ZRANB3 recruitment to DNA damage sites. Proper function of ZRANB3 depends on its interaction with PCNA [9], [10], [11] and three domains, the PCNA-interacting protein motif (PIP-box), the AlkB2 PCNA-interaction motif (APIM), and the NZF domain, are proposed to be involved. These studies are all based on arbitrary deletion of an internal fragment in ZRANB3, which might complicate data interpretation. NleE offers an unprecedented opportunity for specifically inactivating the NZF in ZRANB3 in situ without interfering with other domain functions, which reveals a dispensable role of the NZF domain in DNA damage recruitment. Thus, the NZF either plays little role in the recruitment or is functionally redundant to other domains. It is also plausible that NZF-mediated poly-ubiquitin chain binding may regulate the activity of ZRANB3 itself or fulfill functions as yet undefined.
The cDNA expression constructs for NleE, TAB2/3, TAB2/3-NZF, LUBAC-NZFs and NEMO-NZF were described previously [6]. NleE point mutations were generated by QuickChange Site-Directed Mutagenesis Kit (Stratagene). cDNA for mTrabid was kindly provided by Dr. Paul Evans (University of Sheffield, UK). IMAGE clones for ZRANB2 and ZRANB3 were purchased from Source BioScience LifeSciences Inc. NZFs of Trabid, ZRANB2 and ZRANB3 were PCR-amplified and inserted into the pGEX-6p-2 vector for recombinant expression in E. coli, and the full-length ZRANB3 was inserted into pCS2-Flag for transient expression in mammalian cells. pEF-Flag-TAB3ΔNZF-ZRANB2-NZF chimera were constructed as previously described [6], [21]. Luciferase plasmids were also described previously [22]. All the plasmids were verified by DNA sequencing. Antibodies for EGFP (sc8334), GAL4 AD (C-10, sc-1663) and ubiquitin (P4D1, sc8017) were purchased from Santa Cruz; Anti-Flag (M2) antibody, anti-tubulin antibody and EZview Red ANTI-FLAG M2 Affinity Gel were from Sigma. Antibody for c-Myc (9E10, MMS-150R-200) was purchased from Covance. Cell culture products were from Invitrogen, and all other reagents were Sigma-Aldrich products unless noted.
His-SUMO-NleE was expressed in E. coli BL21 (DE3) Gold strain. Se-Met labeled NleE was expressed in E. coli B834 (DE3) as previously described [23]. NleE was purified sequentially by nickel affinity chromatography, Ulp1 digestion and Mono Q+ Superdex 75 chromatography. Expression and purification of NleE mutants and GST-NZFs were essentially the same as previously described [6]. Purified NleE was concentrated to 20 mg/ml in a buffer containing 20 mM Tris-HCl (pH 8.0) and 100 mM NaCl. Crystals were grown using vapor-diffusion hanging-drop method at 19°C for one week against a reservoir buffer containing 22% PEG3350 and 0.2 M ammonium citrate dibasic. Se-Met crystals were obtained using Se-Met labeled NleE plus 1 mM SAM against 20% PEG3350, 0.125 M ammonium citrate dibasic and 0.1 M sodium malonate (pH 7.0). Crystals were cryo-protected in the well buffer supplemented with 25% glycerol and flash-freezed in liquid nitrogen. Diffraction data were collected at Shanghai Synchrotron Radiation Facility (SSRF) BL-17U at the wavelength of 0.9789 Å for Se-Met crystals and 0.9792 Å for native crystals. All data were processed in the HKL2000 [24].
The phase for NleE was determined from the Se-Met crystal data using the single wavelength anomalous dispersion method [25]. Phasing and initial model building were accomplished using the AutoSol function of PHENIX. Automatic model building was performed using the 2.3-Å native data in PHENIX.Autobuild [26]. The autobuild model was manually adjusted in Coot [27]. The final model was refined in PHENIX.Refine. All the structural figures were prepared using the PyMol program (http://www.pymol.org).
Mammalian cell culture, transfection, immunoprecipitation and luciferase assays were basically the same as those described previously [6]. Rhodamine-Phalloidin staining of F-actin also follows that described the previous literature [28]. EPEC strains and infection protocols were described previously [6].
Yeast whole cell extracts were prepared as previously described with some minor modifications [29]. 20 OD600 units of yeast cells were harvested and freezed in liquid nitrogen. 600 µl of yeast lysis buffer (1.85 M NaOH and 7.4% β-mercaptoethanol) were added and cells were kept on ice for 10 min. Trichloroacetic acid (TCA) was then added to a final concentration of 25% and cell lysates were incubated on ice for another 10 min. After centrifugation at 4°C for 30 min, the pellet was washed with cold acetone for four times. The air-dried pellet was solubilized in the SDS loading buffer and the supernatant was loaded onto an SDS-PAGE for further immunoblotting analysis.
Flag-TAB2/HOIL-1L-V195R/Sharpin-S346R immunopurified from infected 293T cells, Flag-TAB3/TAB3ΔNZF-ZRANB2-NZF co-expressed with NleE in 293T cells, and GST-Vps36/GST-ZRANB3-NZF treated with NleE are subjected to in-gel trypsin digest and subsequent tandem mass spectrometry analysis similarly as previously described [6]. The V195R and S346R mutations in HOIL-1L and Sharpin, respectively, were introduced to facilitate mass spectrometry identification of the target tryptic peptides.
For mass spectrometry analysis of ZRANB3 methylation by NleE, Flag-ZRANB3 was digested by Glu-C in solution. 293T cells transfected with Flag-ZRANB3-expressing plasmid and infected with EPEC were first harvested in buffer A (50 mM Tris-HCl, pH 7.5, 150 mM NaCl, 20 mM n-octyl-β-D-glucopyranoside (INALCO) and 5% glycerol) supplemented with an EDTA-free protease inhibitor mixture (Roche Molecular Biochemicals). Cells were lysed by ultrasonication. The supernatant was pre-cleared by protein G–Sepharose at 4°C for 1 h and subjected to anti-Flag immunoprecipitation. Following 4-h incubation, the beads were washed once with buffer A and then five times with TBS buffer (50 mM Tris-HCl, pH 7.5, and 150 mM NaCl). Bound proteins were eluted with 600 mg/ml Flag peptide (Sigma) in the TBS buffer. The eluted protein was diluted 8 times with 50 mM Tris-HCl (pH 8.5) and then concentrated to 20∼30 µl using the Vivaspin (30,000 MWCO, 500 µl, Sartorius). Tris-HCl (pH 8.5) and Urea were then added to the final concentrations of 0.1 M and 0.8 M, respectively. Ultrasonication was performed to facilitate solubilization of the denatured proteins. The proteins were reduced in 5 mM TCEP (Tris-(carboxyethyl) phosphine hydrochloride) at 55°C for 20 min and then alkylated in 10 mM iodoacetamide at room temperature in dark for 15 min. The alkylated proteins were digested with the sequencing grade Glu-C (Roche Molecular Biochemicals) at 25°C overnight. An aliquot of peptide solution was analyzed by tandem mass spectrometry as previously described [6].
The NZF domain of TAB2 (PDB code 3A9J) [30] was used to model the NleE-NZF complex. The cysteine residues coordinating the Zn (Cys670, Cys673, Cys684 and Cys687) were deprotonated and hydrogen atoms were added using the Protein Local Optimization Program [31], [32], [33]. Protein-protein docking was carried out using the RosettaDock program (Rosetta 3.1) [34], [35] with distance restraints enforced between the carbon atom of donor methyl group in SAM and the sulfur atoms of Cys673 or Cys687 in NZF (cutoff value of 10 Å). Distance restraints between the Zn and sulfur atoms of the four cysteines were also added to ensure the correct spatial geometry of the zinc finger motif. The docking poses were clustered using the NMRCLUST program [36] according to the RMSD values of Cα atoms of TAB2-NZF using the NleE structure as the reference. Representative models of the largest four clusters were selected for further MD simulation refinement.
All the MD simulations were set up by employing the Gromacs 4.07 package [37] with amber 99SB force filed [38] in the TIP3P explicit water model [39]. The tetrahedron-shaped zinc parameters were applied in the MD simulation [40], [41]. After minimization and equilibration, the production run was performed in NVT for 15 ns (300 K) without positional restraints. The short-range electrostatic and Lennard-Jones interactions in the simulations were calculated using a force-shifted cutoff value of 12 Å and 10 Å, respectively. The Long-range electrostatic interactions were computed by the Particle Mesh Ewald method [42]. The covalent bonds involving hydrogen atoms were constrained with the LINCS algorithm [43]. The non-bound interaction energy between TAB2-NZF and NleE was computed by accounting the sum of electrostatic (Ecoul-SR) and vander Waals (ELJ-SR) interaction terms in short range. The surface accessible area is calculated in DSSP program [44], and the related solvent accessibility is measured on the ASA-View Server [45]. The trajectories of last 10-ns MD simulations were saved every 100 ps and further analyzed. The interaction energy between TAB2-NZF and NleE, the RMSD, and the distance of polarized “CH3” group of SAM to the sulfur of Cys673 and Cys687 were measured and compared to obtain the near-native complex structure. The model derived from the cluster 1 of distance restraint sampling was the best model, on which another 15-ns MD simulation was carried out for further optimization.
8 µg of GST-TAB2-NZF was incubated with 6 µg of NleE or its mutants (without exogenous SAM) or 2 µg of NleE (with 0.8 mM exogenous SAM) for 30 min at 37°C in 30 µl of buffer containing 50 mM Tris-HCl (pH 7.5), 150 mM NaCl, 5 mM DTT and 0.1% NP-40. The reaction mixtures were separated on a 12% native-PAGE gel, followed by Coomassie blue staining. 3H-SAM labeling of different GST-NZFs were carried out as previously described [6]. To examine NleE methylation of TAB2/ZRANB3 in vivo, Flag-TAB2/ZRANB3 co-expressed with an empty vector or NleE was immunopurified from 293T cells and subjected to in vitro methylation using 0.6 µg of recombinant NleE and 0.55 µCi of 3H-SAM. To examine the effect of NleE modification on the ubiquitin-chain binding activity of ZRANB3 in vitro, 20 µg of GST-ZRANB3-NZF was incubated with 3 µg of NleE for 30 min at 37°C in a 40-µl reaction containing 0.8 mM SAM. The GST-tagged proteins were then immobilized onto Glutathione Sepharose 4B beads (GE Healthcare) for GST pulldown of Lys48, or Lys63-linked ubiquitin chains or linear tetra-ubiquitin similarly as that described previously [6]. To assay NleE modification and inactivation of cellular ZRANB3, 293T cells, co-transfected with Flag-ZRANB3 and EGFP-NleE as indicated, were harvested and re-suspended in 50 mM Tris-HCl (pH 7.5), 150 mM NaCl, 0.1% Triton X-100 and 5% glycerol. Cells were lysed by ultrasonication. The supernatant was pre-cleared using Protein G Sepharose (GE Healthcare) and then subjected to overnight pulldown by Lys63-linked ubiquitin chains or SBP (streptavidin binding peptide)-tagged linear tetra-ubiquitin chains [6].
U2OS cells cultured in 35-mm glass bottom culture dish (MatTek) were co-transfected with EGFP-ZRANB3 and RFP-NleE expression plasmids as indicated by using the Vigofect reagent (Vigorous). Cells were sensitized by addition of 10 µM BrdU for 16 h and then transferred to the environmental chamber (5% CO2, 37°C) in the spinning disk confocal imaging system (PerkinElmer UltraVIEW VOX). Following visualization under Nikon Eclipse Ti inverted microscope, cells with both EGFP and RFP fluorescence were subjected to laser microirradiation using the FRAP (Fluorescence recovery after photobleaching) module and live images were then taken at indicated times points after the microirradiation.
The coordinates of the NleE structure together with the structure factors have been deposited in the Protein Data Bank with the accession code 4R29.
|
10.1371/journal.pcbi.1000107 | Quantitative Predictions of Peptide Binding to Any HLA-DR Molecule of Known Sequence: NetMHCIIpan | CD4 positive T helper cells control many aspects of specific immunity. These cells are specific for peptides derived from protein antigens and presented by molecules of the extremely polymorphic major histocompatibility complex (MHC) class II system. The identification of peptides that bind to MHC class II molecules is therefore of pivotal importance for rational discovery of immune epitopes. HLA-DR is a prominent example of a human MHC class II. Here, we present a method, NetMHCIIpan, that allows for pan-specific predictions of peptide binding to any HLA-DR molecule of known sequence. The method is derived from a large compilation of quantitative HLA-DR binding events covering 14 of the more than 500 known HLA-DR alleles. Taking both peptide and HLA sequence information into account, the method can generalize and predict peptide binding also for HLA-DR molecules where experimental data is absent. Validation of the method includes identification of endogenously derived HLA class II ligands, cross-validation, leave-one-molecule-out, and binding motif identification for hitherto uncharacterized HLA-DR molecules. The validation shows that the method can successfully predict binding for HLA-DR molecules—even in the absence of specific data for the particular molecule in question. Moreover, when compared to TEPITOPE, currently the only other publicly available prediction method aiming at providing broad HLA-DR allelic coverage, NetMHCIIpan performs equivalently for alleles included in the training of TEPITOPE while outperforming TEPITOPE on novel alleles. We propose that the method can be used to identify those hitherto uncharacterized alleles, which should be addressed experimentally in future updates of the method to cover the polymorphism of HLA-DR most efficiently. We thus conclude that the presented method meets the challenge of keeping up with the MHC polymorphism discovery rate and that it can be used to sample the MHC “space,” enabling a highly efficient iterative process for improving MHC class II binding predictions.
| CD4 positive T helper cells provide essential help for stimulation of both cellular and humoral immune reactions. T helper cells recognize peptides presented by molecules of the major histocompatibility complex (MHC) class II system. HLA-DR is a prominent example of a human MHC class II locus. The HLA molecules are extremely polymorphic, and more than 500 different HLA-DR protein sequences are known today. Each HLA-DR molecule potentially binds a unique set of antigenic peptides, and experimental characterization of the binding specificity for each molecule would be an immense and highly costly task. Only a very limited set of MHC molecules has been characterized experimentally. We have demonstrated earlier that it is possible to derive accurate predictions for MHC class I proteins by interpolating information from neighboring molecules. It is not straightforward to take a similar approach to derive pan-specific HLA-DR class II predictions because the HLA class II molecules can bind peptides of very different lengths. Here, we nonetheless show that this is indeed possible. We develop an HLA-DR pan-specific method that allows for prediction of binding to any HLA-DR molecule of known sequence—even in the absence of specific data for the particular molecule in question.
| Major histocompatibility complex (MHC) molecules play an essential role in the host-pathogen interactions determining the onset and outcome of many host immune responses. While peptides derived from foreign, intracellular proteins and presented in complex with MHC class I molecules can trigger a response from cytotoxic T lymphocytes (CTL), MHC class II molecules present peptides derived from proteins taken up from the extra-cellular environment. They stimulate cellular and humoral immunity against pathogenic microorganisms through the actions of helper T lymphocytes. Only a small fraction of the possible peptides that can be generated from proteins of pathogenic organisms actually generate an immune response. In order for a peptide to stimulate a helper T lymphocyte response, it must bind MHC II in the endocytic organelles [1].
MHC molecules are extremely polymorphic. The number of identified human MHC (HLA) molecules has surpassed 1500 for class I and many thousands for class II [2]. This high degree of polymorphism constitutes a challenge for T cell epitope discovery, since each of these molecules potentially has a unique binding specificity, and hence a unique preference for which peptides to present to the immune system. Even though many of the alleles could be functionally very similar (i.e. have binding pockets that are similar to other alleles) it is often very difficult a priori to identify such similarities since subtle differences in binding pocket amino acids can lead to dramatic changes in binding specificity [3].
During the last decades, prediction of T cell epitopes has reached a level of accuracy which makes prediction algorithms a natural and integral part of most major large scale rational epitope discovery projects [4]–[6]. The single most selective event defining T cell epitopes is the binding of peptide fragments to the MHC complexes [7],[8]. However, most efforts in developing accurate prediction algorithms for MHC/peptide binding has focused on MHC class I (for review see [9]). Here, large-scale epitope discovery projects integrating high-throughput immunoassays [10] with bioinformatics has achieved highly accurate prediction algorithms covering large proportions of the human MHC class I allelic polymorphism [3],[11],[12]. The situation for MHC class II is quite different. Here, most prediction algorithms have been developed from small data sets covering a single or a few different MHC molecules [13]–[24]. Very limited work has been done on deriving HLA class II prediction algorithms with broad allelic coverage. To our knowledge, only three such publicly available method exists: Propred [25], ARB [17], and NetMHCII [26]. Propred is a publicly available version of the TEPITOPE method [27], which is an experimentally derived virtual matrix-based prediction method that covers 50 different HLA-DR alleles, and relies on the approximation that the peptide binding specificity can be determined solely from alignment of MHC pockets amino acids. NetMHCII and ARB are weight matrix data-driven methods derived from quantitative peptide/MHC binding data and covers 14 HLA-DR alleles (as well as some mouse MHC class II alleles). Most other HLA class II prediction methods have been trained and evaluated on very limited data sets covering only a single or a few different HLA class II alleles [13]–[23].
We have previously shown that a minimum number of 100–200 peptides with characterized binding affinity is needed to derive an accurate description of the binding motif for MHC class II alleles [26]. Characterizing the binding preference of each MHC molecule would therefore be an immense and very costly undertaking. In a recent paper, we have demonstrated that is a possible to derive accurate predictions for any HLA class I A and B loci protein of known sequence, by interpolating information from neighboring HLA class I molecules which have been experimentally addressed [3]. It would therefore seem natural to attempt as similar approach to derive a pan-specific HLA class II prediction algorithm. For two major reasons, however, the situation for HLA class II is very different from HLA class I. Firstly, quantitative binding data is only available for a few HLA class II alleles (only 14 HLA-DR alleles are characterized by more than 100 quantitative binding data points, the IEDB database November 2007, [28]). Secondly, the HLA class II binding groove is open at both ends allowing binding of peptides extended beyond the nonamer-binding core [29],[30]. A prerequisite for deriving a pan-specific binding prediction algorithm is therefore a precise alignment of the peptide-binding core to the HLA binding cleft. This alignment is essential since the algorithm underlying the pan-specific binding predictions relies on the ability to capture general features of the relationship between peptides and HLA sequences and interpret these in terms of a binding affinity. Such relationships can only by captured if the peptide is correctly aligned relative to the residues in the HLA binding cleft. We have recently published a method [26] for prediction of peptide-MHC class II binding that covers the 14 HLA-DR alleles which are populated with large amounts of quantitative peptide data in the IEDB database. This method provides a predicted binding affinity value for each peptide, together with an identification of the peptide-binding core, and it is based upon these predictions, we have developed this HLA-DR pan-specific method following the strategy described in [3].
In this work, we demonstrate how a pan-specific HLA-DR prediction method exploiting both peptide and primary HLA sequence can be used to accurately predict quantitative binding predictions for all HLA-DR molecules of known protein sequence. In particular, the method is capable of predicting the specificity of HLA-DR molecules with previously uncharacterized binding specificities thus demonstrating the true pan-specific nature of the method. The method and the benchmark data sets are available at http://www.cbs.dtu.dk/services/NetMHCIIpan.
We trained the pan-specific HLA-DR prediction method as schematically illustrated in Figure 1. Both peptide sequences and HLA primary sequence information were used as input to the method. The peptide core and peptide flanking residues (PFR) were identified using the stabilized matrix alignment method [26]. Multiple register peptides were presented to the method in terms of the normalized measured binding affinity as illustrated in Figure 1B. By including both the peptide and HLA primary sequence, the pan-specific method is able to predict binding of peptides to all HLA-DR molecules even in the absence of data characterizing its binding specificity.
To validate the pan-specific method, we conducted a leave-one-molecule out (LOO) experiment covering all 14 HLA-DR alleles included in the IEDB data set. For each allele, an artificial neural network (ANN) pan-specific predictor was trained as described in Materials and Method using all peptide data from the IEDB data set except the data for the HLA-DR molecule in question. Next, peptide binding affinity values for the HLA-DR molecule in question were obtained as the ANN prediction score for the optimal nonamer peptide core. The experiment thus simulates prediction of binding to hitherto un-characterized HLA-DR molecules. The predictive performance for each HLA allele was measured in terms of the AUC value [31] and Pearson's correlation [32]. Values for the Spearman's rank correlation [32] are given in Table S1. For each allele, we compared the LOO performance to that of the TEPITOPE method [27] for the alleles covered by this method, and a conventional single allele predictor (SMM-align [26]) trained on data from the most closely related HLA molecule as identified by similarity between the HLA sequences (Neighbor).
The results shown in Table 1 clearly demonstrate the predictive power of the pan-specific LOO method. The LOO approach achieves the highest predictive performance for all 11 alleles covered by TEPITOPE, and only for two alleles (DRB1*1302, and DRB4*0101) is the performance of the single allele neighbor method (SMM-align) better than that of the pan-specific LOO method. These differences are statistically significant (p<0.001 and p = 0.001, respectively, Binomial test).
The predictive performance of the pan-specific method relies on the ability to interpolate information from “neighboring” alleles in HLA specificity space and interpret this information in terms of binding affinities. It is thus expected that the pan-specific method should perform best in cases where closely related HLA molecules are included in the training of the method. The data in Table 1 and Figure 2 illustrates that this is indeed the case. Except for the two outliers DRB1*1302, and DRB1*0701 the plot shows the clear relation that alleles with close nearest neighbors tend to be predicted with a higher accuracy compared to alleles with large distances to their nearest neighbor.
Next, the final NetMHCIIpan method was trained on the complete datasets in a fivefold cross-validated manner abandoning the leave-one-out approach (see Materials and Methods). We compare the performance of the NetMHCIIpan method to that of a conventional single allele prediction method (SMM-align) and the TEPITOPE method in terms of both the AUC values and the Pearson's correlation coefficient (the latter is only included for the NetMHCIIpan and SMM-align methods, since the TEPITOPE method does not provide output values that are linearly related to the peptide binding affinity). The summary of this benchmark calculation is shown in Figure 3 (for details see Table S2).
The results show how the pan-specific method is capable of integrating information from neighboring HLA-DR molecules, and thus boosting the predictive performance beyond that of the conventional single allele methods like SMM-align and TEPITOPE. For all 14 alleles included in the benchmark, the pan-specific method outperforms the two other methods (p<0.001, Binominal test).
The ultimate validation of a pan-specific method for HLA-DR peptide binding predictions would be to identify which peptides that will bind to a hitherto un-characterized HLA-DR molecule. We therefore conducted such an experiment where a set of 256 15mer peptides were tested in an in vitro binding assay for binding to the HLA-DRB1*0813 molecule (described in Materials and Methods). Of the 20 top scoring peptides, 75% were shown to bind with a KD values below 1000 nM, and 50% were shown to bind stronger than 50 nM. A performance summary of this experiment is shown in Table 2. This experiment demonstrates how the pan-specific prediction approach can identify peptide-binding motifs even in the absence of any data for the specific query HLA-DR molecule.
The NetMHCIIpan method was further validated using a large set of data from the SYFPEITHI database [29], which were not included in the training data of the NetMHCIIpan method. This set consists of 584 HLA ligands restricted to 28 different HLA-DR alleles. For every peptide, the source protein was found in the SwissProt database [33]. If more than one source protein was possible, the longest protein was chosen. The source protein was split into overlapping peptide sequences of the length of the HLA ligand. All peptides except the annotated HLA ligand were taken as negative peptides. We are aware that this is a strong assumption, since suboptimal peptides that could be presented on the HLA molecule are counted as negatives. For each protein-HLA ligand pair the predictive performance was estimated as the AUC value. The summary of this benchmark calculation is shown in Figure 4 (for details see Table S3).
The NetMHCIIpan and TEPITOPE methods have similar predictive performance on the subset of 17 alleles covered by both methods. The TEPITOPE method has the highest performance for 10 alleles and the NetMHCIIpan the highest performance for 7 alleles (this difference is not significant p>0.3, Binomial test). For the 11 alleles not covered by the TEPITOPE method, NetMHCIIpan achieves the highest performance for 9 alleles, and the TEPITOPE method the highest performance for 2 alleles. For these alleles, NetMHCIIpan thus performs significantly better than the TEPITOPE method (p<0.01, Binominal test). Finally, for the 14 alleles not covered by the SMM-align method, and thus not included in the training of the pan-specific method, NetMHCIIpan achieves a higher performance than the TEPITOPE method. However, this difference is not significant. Also, in this experiment the NetMHCIIpan method performs particularly poorly compared to the TEPITOPE method on the DRB1*13 alleles. Using a network ensemble trained by leaving out the binding data for the DRB1*1302 allele, the average predictive performance for the DRB1*1302 allele is improved from 0.567 to 0.747 (data not shown). This result confirms our earlier observation that the DRB1*1302 allelic data included in the training of the NetMHCIIpan method forms an outlier group with unusual binding specificity characteristics.
To validate the ability of the NetMHCIIpan method to correctly identify the binding core of peptides bound to MHC class II molecules, we compiled from the PDB database [34] a set of 15 peptides which have been crystallized in complex with an HLA-DR allele. For these peptides, we can identify the exact peptide binding by manual extracting which peptide residue is bound in the P1 pocket and subsequently test if this core can be identify by the prediction method. As demonstrated in Table 3, both the TEPITOPE and NetMHCIIpan methods are capable of identifying the binding core of the 15 peptides. TEPITOPE correctly identifies all 15 binding cores, whereas the NetMHCIIpan misaligns one peptide by a single amino acid residue.
It has previously been shown that HLA-A and HLA-B class I molecules can be clustered into a limited number of groups also known as supertypes sharing common binding specificity characteristics. A similar clustering of HLA-DR alleles has also been proposed [35]. In order to validate and extend this clustering, the NetMHCIIpan method was used to cluster HLA-DR molecules according to predicted peptide binding specificity. Pruned HLA distance trees were calculated as described in Materials and Methods. Figure 5 depicts a tree including 76 representatives of the currently known HLA-DR molecules.
The overall structure of the HLA-DR specificity tree is in accordance with the previously proposed clustering [35] containing 12 main supertypes. It is, however, striking to observe the high degree of serotype mixing between the different supertype clusters. Almost all of the proposed supertypes contain HLA-DR molecules from more than one serotype. This has earlier been observed when defining HLA-DR specific clusters based on the TEPITOPE binding matrices [35], but not to the degree suggested by the analysis presented here.
The MHC molecules are extremely polymorphic giving rise to many different peptide-binding specificities being expressed in the human population. More than 500 different HLA-DR molecules and more than 2000 different HLA-DQ and HLA-DP molecules have been described [2]. The only partially pan-specific HLA-DR prediction algorithm publicly available is the TEPITOPE method [27]. This method describes binding of peptides to 50 HLA-DR molecules. However, as shown in this work, the TEPITOPE method leaves large portions of the HLA-DR allelic polymorphism undescribed.
In the present work, we develop a HLA-DR pan-specific method, NetMHCIIpan, capable of providing quantitative predictions of peptide binding to all HLA-DR molecules with known protein sequence. The method is based on artificial neural networks and is trained on quantitative peptide HLA-DR binding data including the peptide-binding core, peptide flanking residues, and the HLA-DR residues estimated to be within interaction distance of the bound peptide. The natural strength of the method is the ability to predict binding of peptides to any HLA-DR molecule, thus being truly HLA-DR pan-specific. Further, since the method is artificial neural network based, it can capture non-linear relationships defining the binding specificity both within the peptide and between the peptide and the HLA molecule. This is fundamentally different from the methodology underlying the TEPITOPE method, that relies on the approximation that peptide binding specificities can be determined as a summation over independent HLA pockets preferences. The method is validated in terms of prediction of peptide binding to hitherto un-characterized HLA-DR molecules, large-scale leave-one-out experiments, cross-validation and identification of endogenously presented peptides and experimentally validated binding cores. In all validation experiments, the NetMHCIIpan method was shown to perform better than or comparable to TEPITOPE, the only other partially HLA-DR pan-specific binding prediction method publicly available.
A powerful application of the HLA-DR pan-specific prediction algorithm would be to search for highly promiscuous peptide sequences that will bind to most HLA-DR alleles. Such peptides could be of high value in the development of synthetic and recombinant vaccines, since they would bind universally in most humans independently of MHC class II genetic background and thus potentially provide universal helper T cell activation. By way of example, we applied the pan-specific method to identify peptides, predicted to bind a set of prevalent HLA-DR alleles. Prevalent alleles were selected as HLA-DR alleles with a maximal allelic frequency above 1% in an ethnic population as reported by Middleton et al. [36]. In doing so, we could identify peptides predicted to bind promiscuously to all prevalent HLA-DR molecules. Earlier efforts have been made to identify such highly promiscuous peptides. The PADRE sequence [37] is one of the most prominent examples of such peptides. Using the pan-specific method, the PADRE sequence is predicted to bind to less than 40% of the prevalent HLA-DR molecules. The analysis shown here demonstrates that exhaustive searches for truly pan-promiscuous HLA-DR are indeed feasible using the proposed pan-specific method.
The pan-specific approach relies on the ability of the neural networks to capture general features of the relationship between peptides and HLA sequences and interpret these in terms of a binding affinity. For this approach to provide reliable predictions, it is essential that polymorphism of the HLA molecules described by the pan-specific method is to some degree covered by the data included in the training of the method. For the NetMHCIIpan prediction method, we have included binding data covering only 14 of the more than 500 known HLA-DR molecules [2], thus very likely leaving large regions of the HLA specificity space uncovered. On the basis of the specificity clustering shown in Figure 5, we can identify HLA-DR alleles with un-characterized binding specificities as these alleles are found far from the alleles included in the training of the pan-specific method. Such novel HLA-DR molecules include the DRB1*14 molecules, i.e., DRB1*1407 (12.5%) and some of the DRB1*11, like DRB1*1103 (5%), as well as DRB1*12 alleles like DRB1*1202 (35%) placed close to center of the tree. The number in parenthesis after each allele is the maximal allelic frequency in an ethnic population as reported by Middleton et al. 2003 [36].
We have previously shown how integrative approaches combining bioinformatics and immunoassays to identify and experimental assay peptide with uncharacterized binding affinity can improve the prediction accuracy of peptide/MHC class I prediction algorithms [38]. Using the pan-specific approach to identify HLA class II molecules with uncharacterized binding specificities, we suggest extending this search strategy into the dimension of MHC polymorphism. A schematic illustration of this search strategy integrating bioinformatics and high throughput immunoassays is shown in Figure 6.
Here, we illustrate an iterative cycle that identifies novel MHC molecules with predicted binding specificities that are dissimilar to the specificities included in the training of the pan-specific method. Next, immunoassays should be developed describing the binding specificity of these molecules by identifying peptides with un-characterized binding affinity, and experimentally assay these peptides. Such an approach should allow for rapid and efficient sampling of both the MHC polymorphism and the diversity of peptide binding.
The current version of NetMHCIIpan and the benchmark data used in this work is available at http://www.cbs.dtu.dk/services/NetMHCIIpan. The service covers all HLA-DR alleles with known protein sequence. The method will be updated as more data becomes available. In the future, it is our hope to extend the method to also cover HLA-DQ and HLA-DP molecules.
Quantitative HLA-DR restricted peptide-binding data was obtained from the IEDB database [28] and from an in-house collection of unpublished data [Bjorn Peters, private communication]. For external evaluation of the pan-specific method, we included a set of HLA-DR class II ligands from the SYFPEITHI database [29]. Only ligands not included in the quantitative HLA-DR restricted peptide binding data set were used. The SYFPEITHI data set consists of 584 MHC ligands restricted to 28 HLA-DR alleles. The details on the data set is given is Tables S4 and S5 (the complete data sets are available at http://www.cbs.dtu.dk/suppl/immunology/NetMHCIIpan.php).
The pan-specific HLA-DR method was constructed as described in Figure 1. The peptide nonamer core and peptide-flanking residues (PFR) were identified for each of the peptides in the IEDB dataset using the SMM-align method [26]. The SMM-align method identifies of the maximal scoring nonamer peptide core for each peptide sequence. This approach will thus leave out information on the suboptimal nonamer sequences that are predicted not to bind or to bind with a weaker affinity. To include information on the binding affinity for these suboptimal nonamer peptides, we assign a normalized binding score, Snorm, to suboptimal nonamer peptides given as the ratio of the SMM-align score for the peptide to the SMM-align score of the optimal peptide multiplied with the log-transformed experimental IC50 binding value of the peptide. That is Snorm = (S/SM)M, where S is the SMM-align score for the (suboptimal) peptide, SM is the SMM-align score of the optimal peptide, and M is the binding value log-transformed as 1−log50k(aff), where aff is the experimental IC50 binding value of the full-length peptide, and log50k is the logarithm with base 50.000. In case the SMM-align method assigns the maximal scoring nonamer peptide a log-transform binding value of 0, the log-transformed experimental IC50 binding value is assigned randomly to one of the suboptimal peptides and all other nonamer peptides are given a binding value of 0. In doing this expansion using sub-optimal nonamer peptides, the size of the IEDB dataset was enlarged from 14,607 to more than 100,000 data points. This more than 5 fold increase of the data gave consistent improvements to the accuracy of the prediction method in all benchmark calculations (data not shown).
For each peptide core, the PFRs were identified as the amino acids flanking the peptide core up to a maximum of three at either end.
The HLA sequence was encoded in terms of a pseudo-sequence consisting of amino acid residues in contact with the peptide. The contact residues are defined as being within 4.0 Å of the peptide in any of a representative set of HLA class II structures. Only residues polymorphic in any known HLA-DR, DQ and DP protein sequence were included giving rise to a pseudo-sequence consisting of 21 amino acid residues. The HLA class II pseudo-sequence is described in detail in Table S6.
Artificial neural networks (ANN) were trained to quantitatively predict peptide-HLA binding as described in Nielsen et al. [3]. The input sequences were presented to the neural network in three distinct manners: (a) conventional sparse encoding (i.e., encoded by 19 zeros and a one), (b) Blosum encoding, where each amino acid was encoded by the BLOSUM50 matrix score vector [39], and (c) a mixture of the two, where the peptide was sparse encoded and the HLA pseudo sequence was Blosum encoded. PFRs were calculated as the average BLOSUM62 score over a maximum length of three amino acids [26]. The PFR length was encoded as LPFR/3, 1−LPFR/3, where LPFR is the length of the PFR (between 0 and 3), and the peptide length was encode as LPEP, 1−LPEP, where LPEP = 1/(1+exp((L−15)/2)) and L is the peptide length. For each data point, the input to the neural network thus consists of the peptide sequence (9×20 = 180 inputs), the PFRs (2×20 = 40 inputs), the HLA pseudo sequence (21×20 = 420 inputs), the peptide length (2 inputs), and the length of the C and N terminal PFR's (2×2 = 4 inputs) resulting in a total of 646 input values.
To estimate the predictive performance of the method, the leave-one-out (LOO) experiment was conducted as described by Nielsen et al. [3]. For each HLA-DR molecule, a neural network ensemble was trained using all available data, excluding all data specific for the HLA-DR allele in question. Network architectures with hidden neurons of 22, 44, 56, and 66 were used. The network training was performed in a fivefold cross-validated manner using the three encoding schemes described above resulting in an ensemble of 60 neural networks (3 encoding schemes, 4 architectures, and 5 folds). The predicted affinity for a peptide was then determined as prediction value for the maximal scoring nonamer peptide core (including PFRs), where each nonamer peptide core is scored as the average of the 60 predictions in the neural network ensemble.
For the final NetMHCIIpan method, a conventional five-fold cross-validated training was performed. The pool of unique peptides was randomly split into five groups with all HLA binding data for a given peptide placed in the same group (in this way, no peptide can belong to more than one group).
The nearest neighbor distance between two HLA alleles is estimated from the alignment score of the HLA pseudo sequences using the relation d = 1−s(A,B)/(s(A,A)s(B,B))1/2, where s(A,B) is the BLOSUM50 alignment score [39] between the pseudo sequences A and B, respectively.
HLA distance trees were derived from correlations between predicted binding affinities as described by Nielsen et al. [3]. In order to visualize the HLA distance tree, only a subset of the leaves in the tree was displayed. The subset was selected in a Hobohm 1-like manner, where the alleles were clustered at a 0.95 distance level and only a single allele from each cluster selected for display [40].
The extracellular parts of HLA DRA1*0101 and HLA DRB1*0813 were fused to the Fos Jun leucine zipper dimerization motifs as previously described [41]. Both chains were separately expressed as inclusion bodies in E. coli (BL21) using standard IPTG induction. The two chains were extracted from inclusion bodies and purified by anion exchange and gel filtration chromatography under denaturing conditions. Equimolar concentrations of alpha and beta chain were diluted into a refolding buffer containing a titration of peptide (0–15 µM). After 48 h of incubation at 18°C the concentration of formed complex was determined by a quantitative ELISA using the HLA-DR specific monoclonal antibody L243. The data was fitted to a saturation curve using non-linear regression and the Kd value determined. |
10.1371/journal.ppat.1004721 | Human T-Cell Leukemia Virus Type 1 (HTLV-1) Tax Requires CADM1/TSLC1 for Inactivation of the NF-κB Inhibitor A20 and Constitutive NF-κB Signaling | Persistent activation of NF-κB by the Human T-cell leukemia virus type 1 (HTLV-1) oncoprotein, Tax, is vital for the development and pathogenesis of adult T-cell leukemia (ATL) and HTLV-1-associated myelopathy/tropical spastic paraparesis (HAM/TSP). K63-linked polyubiquitinated Tax activates the IKK complex in the plasma membrane-associated lipid raft microdomain. Tax also interacts with TAX1BP1 to inactivate the NF-κB negative regulatory ubiquitin-editing A20 enzyme complex. However, the molecular mechanisms of Tax-mediated IKK activation and A20 protein complex inactivation are poorly understood. Here, we demonstrated that membrane associated CADM1 (Cell adhesion molecule1) recruits Ubc13 to Tax, causing K63-linked polyubiquitination of Tax, and IKK complex activation in the membrane lipid raft. The c-terminal cytoplasmic tail containing PDZ binding motif of CADM1 is critical for Tax to maintain persistent NF-κB activation. Finally, Tax failed to inactivate the NF-κB negative regulator ubiquitin-editing enzyme A20 complex, and activate the IKK complex in the lipid raft in absence of CADM1. Our results thus indicate that CADM1 functions as a critical scaffold molecule for Tax and Ubc13 to form a cellular complex with NEMO, TAX1BP1 and NRP, to activate the IKK complex in the plasma membrane-associated lipid rafts, to inactivate NF-κB negative regulators, and maintain persistent NF-κB activation in HTLV-1 infected cells.
| HTLV-1 infection leads to the development of Adult T-cell Leukemia (ATL) or HTLV-1 associated myelopathy/ tropical spastic paraparesis (HAM/TSP). One of the major causes responsible for the development of HTLV-1 associated diseases is chronic inflammation directed by NF-kappaB (NF-κB). NF-κB activation in response to a wide variety of signals is transient and tightly controlled by ubiquitin-editing enzyme A20. One of the mechanisms of persistent NF-κB activation in HTLV-1 infected cells is inactivation of NF-κB negative regulators; however, the precise mechanism is unknown. Here, we focused on host tumor suppressor Cell adhesion molecule 1 (CADM1) that is robustly upregulated in HTLV-1 infected cells. The expression of CADM1 is frequently silenced in several cancers; however, it is critical for HTLV-1 associated ATL tumor cell survival. We characterized the role of CADM1 in persistent NF-κB activation in HTLV-1 infected cells. We found that CADM1 is required for the HTLV-1 oncoprotein, Tax, to form a cellular complex with Ubc13, TAX1BP1, NRP and NEMO in the membrane lipid rafts micorodomain. We further demonstrated that Tax requires CADM1 to inactivate NF-κB negative regulator and maintain persistent NF-κB activation. Our study reveals a novel mechanism of chronic NF-κB activation by CADM1 in HTLV-1 infected cells.
| Infection with human T-cell leukemia virus type 1 (HTLV-1), an oncogenic retrovirus, is associated with the development of adult T-cell leukemia (ATL), an aggressive and lethal malignancy of CD4+ T lymphocytes and a chronic neuroinflammatory disease termed HTLV-1-associated myelopathy/tropical spastic paraparesis (HAM/TSP). HTLV-1 encodes a 40-kDa oncoprotein Tax that regulates viral gene expression and plays vital roles in ATL leukemogenesis [1–3]. Tax regulates the expression of viral and cellular genes involved in cell transformation, immortalization, and tumor initiation through NF-κB, cyclic AMP response element-binding protein (CREB), and serum responsive factor (SRF) signaling pathways [4,5]. Tax also promotes cellular transformation by inducing post-translational modifications of multiple cellular factors, inactivating tumor suppressors, and dysregulating cellular signaling pathways and cell cycle machinery [6–12]. The carboxyl-terminal PDZ-binding domain motif (PBM) of Tax recruits PDZ domain-containing cellular factors, which play critical roles in the dysregulation of signaling pathways, proliferation, and immortalization of primary T-cells [13]. One of the key functions of Tax is the persistent activation of the nuclear factor kappa-B (NF-κB) transcription factor signaling pathways that are important for transformation, proliferation, and survival of HTLV-1 infected T-cells [14–16]. Tax also maintains persistent NF-κB activation by inactivating NF-κB negative regulators, such as A20 and cylindromatosis (CYLD) [17–19]. However, the underlying mechanisms of Tax-mediated inactivation of NF-κB negative regulators and persistent NF-κB activation remain poorly understood.
NF-κB plays critical roles in inflammation and the development of innate and adaptive immunity [20]. The NF-κB family is composed of five members, NF-κB1 (p50/p105), NF-κB2 (p52/p100), p65 (RelA), RelB, and c-Rel, and each of these proteins can form homo- and heterodimers [21]. Upon stimulation of TNF receptor 1 (TNFR1) with TNF or the T-cell receptor (TCR) with antigen, NF-κB activation is triggered in the membrane microdomains, termed lipid rafts [22,23]. NF-κB is sequestered in the cytoplasm as an inactive form by the family of IκB proteins and can be rapidly activated in response to stimulation [24]. NF-κB activating signals converge at the IκB kinase (IKK) complex containing the catalytic kinase subunits IKKα, IKKβ, and the regulatory subunit IKKγ (also known as NEMO) [25]. Based on the involvement of specific receptors and extracellular stimuli, NF-κB pathways are classified into either canonical (classical) or noncanonical (alternative) pathways. The classical NF-κB pathway is dependent on IKKβ and NEMO, whereas the alternative NF-κB pathway is dependent on IKKα.
In the classical NF-κB pathway, IKKβ phosphorylates IκBα on two serine residues in a NEMO-dependent manner which triggers its polyubiquitination followed by proteasome degradation of IB, thus liberating NF-κB and allowing its rapid mobilization into the nucleus where it regulates the expression of a plethora of genes involved in cell growth, inflammation and survival [26]. In the alternative NF-κB pathway, IKKα is activated by the NF-κB inducing kinase (NIK) in response to specific ligands of the TNF superfamily, including BAFF, lymphotoxin- and CD40L [27]. Activated IKKα phosphorylates p100 that results in the ubiquitination and partial degradation of p100 by the proteasome to generate p52. Tax constitutively activates both the classical and alternative NF-κB pathways by interacting with NEMO in the absence of extracellular stimuli [28]. It has been demonstrated that Ubc13-dependent Tax ubiquitination is essential for interaction with NEMO [15]. Interestingly, Tax-mediated NF-κB activation is initiated in the Golgi-associated lipid rafts [29]; however, the mechanism is poorly understood. It has also been recently demonstrated that Tax requires Tax1 binding protein1 (TAX1BP1) to activate NF-κB [19,30]. TAX1BP1 is an adaptor molecule for the NF-κB negative regulatory ubiquitin-editing enzyme A20 complex in receptor mediated NF-κB signaling pathways [31,32]. Phosphorylation of TAX1BP1 by IKKα in a stimulus-dependent manner assembles the ubiquitin-editing enzyme A20 complex (A20, TAX1BP1, Itch and RNF11) to terminate NF-B signaling and thus maintain transient NF-κB activation. Interestingly, TAX1BP1-mediated assembly of the A20 ubiquitin-editing enzyme complex was impaired due to the disruption of TAX1BP1 and IKKα interactions by Tax [19] in TNF- or IL-1-stimulated cells. However, the mechanism of TAX1BP1 and IKKα dissociation in Tax expressing cells is poorly understood.
Cell adhesion molecule 1 (CADM1; also known as TSLC1, IGSF4 or NECL2) is encoded by chromosomal region 11q23 and was first identified as a tumor suppressor gene in non-small cell lung cancer (NSCLC) [33]. CADM1 is a 442 amino acid protein that is highly upregulated in ATL cells [34]. Although CADM1 is localized in the cytoplasm and cell membrane, it is mostly localized in the basolateral membrane of the cells [35–39]. The N-terminal three extracellular Ig-like loops and the C-terminal cytoplasmic tail are the main functional regions of the CADM1 protein. The extracellular loops interact with the EGF receptor/ErbB family members, cell surface receptor class-I-restricted T-cell-associated molecule (CRTAM), and integrin α6β4 [40–42] and the cytoplasmic tail contains a protein 4.1-binding motif (protein 4.1-BM) and a type II PDZ-binding motif (PDZ-BM) which are critical for CADM1 function [43]. CADM1 plays a critical role in the adhesion of spermatogenic cells to Sertoli cells [44]. The interaction between CRTAM expressed on activated cytotoxic T-cells and CADM1 expressed on antigen presenting cells drives the production of IFN-γ and IL-22 by the activated CD8+ T-cells [45]. Studies with CADM1-deficient mice have revealed that the interaction between CADM1 and the TCR ζ-chain is critical for T-cell functions [46]. It was also demonstrated that CADM1 expression is reduced in lung cancer cell lines and the reintroduction of CADM1 into A549, a NSCLC cell line, significantly inhibited tumorigenicity in nude mice [47]. Intriguingly, although CADM1 functions as a tumor suppressor in NSCLC, it may function as an oncoprotein in ATL cells [47]. The fundamental mechanistic role of CADM1 as an oncoprotein in ATL cell is not known.
Here we demonstrated that the expression of CADM1 was regulated by Tax. CADM1 was required for K63-linked polyubiqutination of Tax in the membrane lipid rafts. Interestingly, CADM1 was also required for Tax to interact with Ubc13, NEMO, TAX1BP1, and NRP, and for activation of the IKK complex in the membrane-associated lipid raft in HTLV-1 infected T-cells. Finally, Tax failed to inhibit TAX1BP1 phosphorylation, which is critical for the assembly of the NF-κB negative regulator, ubiquitin-editing enzyme A20 complex, in Cadm1-deficient MEFs stimulated with TNF or IL-1, indicating that CADM1 is not only required for Tax to activate NF-κB but also to target NF-κB negative regulators in HTLV-1 mediated tumor cells.
The expression of CADM1 has been shown to be upregulated in HTLV-1 transformed and primary ATL cells derived from acute type ATL patients; however, the mechanism of CADM1 regulation in HTLV-1 infected cells is currently unknown [34,48]. The Tax point mutant M47 activates NF-κB (but not CREB/ATF), and Tax M22 activates ATF/CREB (but not NF-κB) transcription factors [10]. Recent findings have showed massive inflammation and spontaneous tumor development at different sites in Tax transgenic mice [49,50]. Therefore, we first examined CADM1 expression in the spontaneous tumors that developed in Tax transgenic mice, and found that all tumors from these mice displayed elevated levels of CADM1 expression compared to normal mice (Tax-negative control mice) (Fig. 1A). Lentiviral-mediated expression of wildtype Tax, and Tax mutants M22 and M47 showed increased CADM1 mRNA (S1A-S1B Fig.) and protein expression in primary murine embryonic fibroblasts (MEFs) and Jurkat T-cells (Fig. 1B-C). Interestingly, lentiviral expressing Tax double mutant M22 and M47, that could activate neither CREB nor NF-κB, failed to induce CADM1 protein expression in MEFs cells (S7 Fig.). Recent reports suggest that the expression of suppressor of cytokine signaling 1 (SOCS1) is regulated by Tax in HTLV-1 infected cells through NF-κB but not through the CREB pathway [51,52], and we found that wildtype Tax and the Tax mutant M47, but not the Tax mutant M22, induce the expression of SOCS1 mRNA (S1A-S1B Fig.) and protein (Fig. 1B-C). We also observed elevated CADM1 protein expression in Tax expressing HTLV-1 infected C8166, MT-2 and MT-4 cells as compared to Jurkat T-cells (S1C Fig.). Collectively, these results suggest that Tax regulates the expression of CADM1 through the mechanisms dependent of NF-κB and CREB pathways.
To gain more insight into the regulation of CADM1 expression by Tax, we performed co-immunoprecipitation (Co-IP) assays to determine if there was a physiologically relevant association of Tax with CADM1 in Jurkat T-cells. Tax was expressed by lentiviral transduction and was subsequently immunoprecipitated with anti-Tax using JurkaT-cell lysate. Surprisingly, Tax interacted with endogenous CADM1 (Fig. 2A); however, no binding was observed when immunoprecipitations were performed with a control mouse immunoglobulin antibody (Fig. 2A). We also detected the interaction between Tax and CADM1 in HTLV-1 transformed cell lines C8166, MT-2 and MT-4 proving that Tax indeed physically associated with CADM1 in HTLV-1 infected T-cells (Fig. 2B). These results suggest that Tax is tightly and stably associated with CADM1. We generated several CADM1 deletion mutants to examine the Tax interacting region (Fig. 2C). Cadm1 −/− MEFs were reconstituted with these Flag-tagged CADM1 deletion mutants to determine the domain(s) important for Tax interaction. Deletion of the extracellular region (Δ-EC) had no effect on the Tax and CADM1 interaction (Fig. 2D). However, deletion of the cytoplasmic tail (Δ-CP), specifically PDZ-binding motif (Δ-PDZ-BM) abrogated CADM1 and Tax interaction (Fig. 2D). These studies revealed that the PDZ binding motif of the cytoplasmic tail of CADM1, but not the extracellular region, is critical for Tax interaction.
Since previous studies showed that Tax interaction with its adaptor molecules, TAX1BP1, NEMO, NEMO-Related Protein (NRP), and the ubiquitin-conjugating enzyme Ubc13, is essential for the activation of NF-κB [15,19,53], we next examined whether CADM1 was required for Tax to interact with these adaptor molecules. To address this possibility, HTLV-1 transformed MT-2 cells were stably transduced with four distinct CADM1 shRNAs and CADM1 knockdown was confirmed by immunoblotting, which revealed that shRNA number 3 was the most efficient in the knockdown of endogenous CADM1 expression in MT-2 cells (S2A Fig.). We next examined by co-IPs the endogenous interactions of Tax with TAX1BP1, NEMO, NRP and Ubc13 in the HTLV-1 transformed C8166, MT-2 and MT-4 cell lines expressing control or CADM1 shRNA. Tax interaction with endogenous TAX1BP1, NEMO, NRP and Ubc13 occurred in the presence of control shRNA, but not CADM1 shRNA (S2B Fig.). Similarly, Tax interacted with its adaptor molecules in Cadm1+/+ but not in Cadm1 −/− MEFs (S2C Fig.). Collectively, these results suggest that CADM1 is required for Tax to interact with its adaptor molecules TAX1BP1, NEMO, NRP and Ubc13.
The mechanistic and functional role of the cytoplasmic tail of CADM1 has not been well understood. Several reports have documented that K63-linked polyubiquitination of Tax is indispensable for its activity and interaction with NEMO and NRP, and subsequent activation of the NEMO/IKK complex [15,53]. Therefore, we next analyzed Tax K63-linked polyubiquitination in Jurkat T-cells transduced with lentiviruses expressing Tax and either control or CADM1 shRNA number 3 (hereafter referred to as CADM1 shRNA). Lysates were immunoprecipitated with Tax antibody, eluted with 1% SDS, diluted in lysis buffer and re-immunoprecipitated with Tax antibody to ensure that we were examining Tax ubiquitination, and not that of an associated protein. Tax polyubiquitination was detected by immunoblotting with antibodies to K63-Ubi or Tax. Knockdown of CADM1 greatly reduced the K63-linked polyubiqutination of Tax compared to cells expressing control shRNA (Fig. 3A). CADM1 knockdown had no effect on the expression of Tax as confirmed by immunoblotting (Fig. 3A). We next examined the K63-linked polyubiquitination of Tax in Cadm1 −/− MEFs. The Tax K63-linked polyubiquitination assay was carried out as above using lysates from Cadm1+/+ and Cadm1 −/− MEFs transduced with Tax-expressing lentiviruses. The K63-linked polyubiquitination of Tax was markedly impaired in Cadm1 −/− MEFs compared to Cadm1+/+ control cells (Fig. 3B). Furthermore, knockdown of endogenous CADM1 with lentiviral shRNA in C8166, MT-2 and MT-4 cells also showed loss of K63-linked polyubiquitination of endogenous Tax (Fig. 3C). These collective results strongly suggest that CADM1 is absolutely essential for K63-linked polyubiquitination of Tax. Since CADM1 does not possess ubiquitin-conjugating or ubiquitin-ligase enzyme activity, this suggests that it may be recruiting E2 ubiquitin-conjugating enzyme Ubc13 on Tax.
Previous studies have demonstrated that K63-linked polyubiquitination of Tax is critical for activation of NF-κB pathways, and Tax also activates noncanonical NF-κB pathways [15,53]. We therefore examined the functional role of CADM1 in both the canonical (NF-κB1) and noncanonical (NF-κB2) NF-κB activation pathways by Tax. First, to examine canonical NF-κB activation, endogenous CADM1 expression was stably suppressed with lentiviral shRNA in Jurkat T-cells, which were then transfected with Tax and NF-κB luciferase plasmids. NF-κB activation was examined by both luciferase assays and immunoblotting for phosphorylated IκBα. As expected, Tax-mediated activation of NF-κB was impaired in CADM1 knockdown Jurkat T-cells as determined by luciferase assays (Fig. 4A) and phosphorylation of IκBα (Fig. 5A). Similarly, induction of IκBα phosphorylation, and NF-κB DNA binding were impaired in Cadm1 −/− MEFs compared to wildtype MEFs (Figs. 4B, 5B). A control Oct-1 EMSA demonstrated similar DNA binding in all of the nuclear extracts (Fig. 5C). We next examined the motifs of CADM1 that were required for Tax to activate canonical NF-κB. Using primary Cadm1 −/− MEFs transfected with Tax, in the absence or presence of wildtype or deletion mutants of CADM1, we found that the overexpression of CADM1 alone induced NF-κB more robustly than Tax and the cytoplasmic tail and PDZ-BM of CADM1 were required for Tax to activate canonical NF-κB (Fig. 4C). Phosphorylation of IκBα and NF-κB DNA binding were also abrogated in HTLV-1 transformed C8166, MT-2 and MT-4 cells (Fig. 5D, E) after knockdown of CADM1 by CADM1 shRNA, suggesting that constitutive NF-κB activation in these cells was CADM1-dependent. Finally, the induction of NF-κB target genes by Tax was examined in Cadm1+/+ and Cadm1 −/− MEFs by RT-PCR. Induction of NF-κB target genes, A20, IL-6, and Bfl-1 by Tax were defective in Cadm1 −/− MEFs (Fig. 5F). These results suggest that CADM1 is required for Tax to activate the canonical NF-κB pathway.
Next, we examined if CADM1 was required for Tax to activate the noncanonical NF-κB pathway. When Tax was overexpressed in Jurkat T-cells that were stably knocked down for CADM1, we found that Tax-mediated processing of p100 to p52 was completely impaired in CADM1 knockdown cells compared to cells expressing control shRNA (S3A Fig.). We also confirmed that Tax-mediated processing of p100 to p52 was impaired in Cadm1-deficient MEFs (S3B Fig.). Thus, CADM1 is essential for Tax to activate both the canonical and noncanonical NF-κB pathways.
It has been reported that NF-κB activation in T-cells downstream of TCR engagement strictly occurs in plasma membrane lipid rafts [23], and that Tax also mediates persistent NF-κB activation in the membrane lipid rafts [29,54]. To further characterize the CADM1 and Tax sub cellular localization, we used confocal fluorescence imaging. CADM1 staining was substantially localized in the plasma membrane and cis-Golgi (as determined by Golgi marker GM-130) and Tax was localized in the plasma membrane, cytoplasm and nucleus (consistent with previously published reports [29,55,56] in MT-2 (Fig. 6A), MT-4 (S5A Fig.), and in C8166 cells (S9A Fig.). Next, we co-stained Tax and CADM1 with cholera toxin B labeled with red fluorescence dye, which binds specifically to the sphingolipid-enriched microdomains and found that significant portions of Tax and CADM1 proteins overlapped with GM1 in MT-2 (Fig. 6B), MT-4 (S5B Fig.) and C8166 cells (S9B Fig.). Thus, significant amounts of Tax co-localizes with CADM1 in the lipid rafts.
Since K63-linked polyubiquitination of Tax is critical for IKK complex activation [57], which is initiated in the membrane lipid rafts [29], further experiments were conducted to determine whether Tax undergoes K63-linked polyubiquitination in the membrane lipid raft or in cytoplasmic portion in HTLV-1 transformed cells. We made lysates from MT-2 cells stably expressing lentiviral control shRNA or CADM1 shRNA and subjected these lysates to density gradient ultracentrifugation. Lysates obtained from density gradient ultracentrifugation were split equally into two parts, one was used for immunoprecipitation with anti-Tax and immunoblot with anti-K63-Ubi and anti-Tax, the other was used for immunoprecipitation with anti-CADM1 and immunoblot with known Tax-interacting molecules TAX1BP1, NEMO, Ubc13 and NRP. As shown in MT-2 (Fig. 6C) and in MT-4 cells (S5C Fig.), Tax underwent K63-linked polyubiquitination only in the lipid raft fractions of control shRNA expressing cells; however, Tax was unable to undergo K63-linked polyubiquitination in CADM1 knockdown cells. Similarly, Tax interacted with TAX1BP1, NEMO, Ubc13 and NRP in the lipid raft fractions, corresponding to the fractions with the lipid raft markers GM1 and LAT (fractions 4, 5 and 6) from MT-2 and MT-4 cells expressing control shRNA; however, Tax was unable to interact with these molecules upon knockdown of CADM1. Examination of the total and phosphorylated IKKα/β and Tax-associated proteins (TAX1BP1, NEMO, Ubc13 and NRP) in the lysate fractions from MT-2 and MT-4 cells expressing control shRNA or CADM1 shRNA showed that IKKα/β was robustly phosphorylated only in the lipid raft fractions (fractions 4, 5 and 6) from MT-2 and MT-4 cells expressing control shRNA; however, IKK activation was impaired in the lipid raft fractions (fractions 4, 5 and 6) from MT-2 and MT-4 cells expressing CADM1 shRNA (Figs. 6C and S5C). We also observed that when CADM1 is present in the cytoplasm Tax does not undergo K63-linked polyubiquitination in the cytoplasm, nor does it interact with its associated proteins in the cytoplasm.
We also sought to determine whether Tax and its associated proteins, TAX1BP1, NEMO, Ubc13 and NRP were localized in the plasma membrane by co-staining with cholera toxin B labeled with red fluorescence dye and confocal fluorescence imaging technique in MT-2 cells. As expected, Tax and its associated proteins were mostly localized in the cytoplasm in MT2 cells and significant portions of Tax and TAX1BP1, NEMO, Ubc13 and NRP overlapped with GM1, shown in the merged pictures (S8 Fig.). These results strongly suggest that Tax and its associated proteins localized in the membrane lipid rafts of HTLV-1 transformed MT-2 cells.
To determine if K63-linked polyubiquitination of Tax and Tax interaction with its associated molecules occurred in the HTLV-1 transformed Tax expressing MT-2 cells treated with cholesterol-chelating agent, methyl-β-cyclodextrin (MβCD) − a selective cholesterol inhibitor that impairs formation of lipid rafts. Lysates obtained from density gradient ultracentrifugation were split equally in two parts, and immunoprecipitated as described above. As expected K63-linked polyubiquitination of Tax, IKK activation, and Tax interaction with CADM1, TAX1BP1, NEMO, Ubc13 and NRP were intact in the membrane lipid rafts of control treated (media) MT-2 cells. However, K63-linked polyubiquitination of Tax, IKK activation, and Tax interaction with CADM1, TAX1BP1, NEMO, Ubc13 and NRP were impaired in the MβCD treated MT-2 cells (S6 Fig.). These results strongly suggest that intact lipid raft is critical for K63-linked polyubiquitination of Tax and Tax interaction with its associated molecules to activate IKK complex in HTLV-1 infected cells.
It has been previously demonstrated that K63-linked polyubiquitination of Tax induced by cytosolic factors is sufficient to activate the IKK complex and the first round of IκBα phosphorylation in vitro cell-free system. Therefore, we utilized a cell-free system, which lacks plasma membrane associated CADM1, to ascertain whether CADM1 plays any role in Tax-mediated activation of the IKK complex, and the induction of the first round of IκBα phosphorylation and degradation. Cytosolic extracts prepared from Jurkat T-cells, NEMO deficient JM4.5.2 cells, and Cadm1 wildtype and deficient MEFs were incubated with recombinant Tax (S10A Fig.), and immunoprecipitated with anti-Tax followed by immunoblotting with anti-NEMO, anti-phospho-IKKα/β, anti-IKKα, anti-IKKβ, anti-CADM1, and anti-Tax. Lysates from these reaction mixtures were also examined for Tax-mediated phosphorylation and degradation of IκBα, expression levels of IKKα, IKKβ, NEMO, CADM1, and β-actin. As expected, IKK complex activation, phosphorylation and degradation of IκBα, and Tax-IKK interactions were impaired in NEMO-deficienT-cells. Interestingly, IKK complex activation, phosphorylation and degradation of IκBα, and Tax-NEMO interactions occurred in CADM1-deficient cytosolic extracts similar to Jurkat T-cells and Cadm1 wildtype extracts (S10B Fig.). We next sought to determine if endogenous CADM1 was required for Tax-NEMO interaction, IKK complex activation, IκBα phosphorylation and degradation in intacT-cells. Lysates from stably expressed lentiviral Tax in Jurkat T-cell, NEMO-deficient JM4.5.2 cells, and Cadm1 wildtype and deficient MEFs (S10C Fig.), or stably expressing control scrambled shRNA or CADM1 shRNA in HTLV-1 transformed C8166, MT-2, and MT-4 cells (S10D Fig.) were immunoprecipitated with anti-Tax followed by immunoblotting with anti-NEMO, anti-phospho-IKKα/β, anti-IKKα, anti-IKKβ, anti-CADM1, and anti-Tax. Lysates were also examined for Tax-mediated IκBα phosphorylation and degradation. As expected, we observed interactions between Tax and NEMO and IKK complex activation and the first round of IκBα phosphorylation and degradation in Jurkat and CADM1 wildtype MEFs cells; however, the interactions between Tax and NEMO and IKK complex activation and first round of IκBα phosphorylation and degradation were impaired in CADM1-deficienT-cells similar to NEMO-deficienT-cells (S10C-S10D Fig.). These results clearly indicate that the mechanisms of Tax-mediated activation of IKK and the induction of IκBα phosphorylation and degradation in an in vitro cell-free system and in intacT-cells are partly different.
Studies have demonstrated that the TNFR and TLR4/IL-1Rβ-mediated activation of the canonical NF-κB pathway is tightly regulated by the A20 enzyme complex [19,30]. TAX1BP1 is a critical adaptor molecule for the A20 enzyme complex and is essential for the negative regulation of the canonical NF-κB pathway [19,32,58], thus prompting us to examine whether CADM1 is required for Tax to target TAX1BP1 phosphorylation and inactivation of the A20 complex. Lentiviruses expressing empty vector or Tax were transduced in Cadm1 wildtype and deficient MEFs and stimulated with either TNF-α or IL-1β at various times to induce phosphorylation of TAX1BP1. As expected, TAX1BP1 was inducibly phosphorylated and interacted with A20 (Fig. 7). Transient phosphorylation and degradation of IκBα occurred after 15 minutes in empty vector expressing wildtype Cadm1 MEFs stimulated with TNF-α or IL-1β; however, Tax completely blocked the phosphorylation of TAX1BP1 and A20 interaction with TAX1BP1 and triggered the persistent phosphorylation and degradation of IκBα in wildtype Cadm1 MEFs treated with TNF-α (Fig. 7) or IL-1β (S4 Fig.). Interestingly, TAX1BP1 phosphorylation and interaction with A20, and phosphorylation and degradation of IκBα, were normal and transient in empty vector or Tax expressing Cadm1-deficient MEFs stimulated with TNF-α (Fig. 7) or IL-1β (S4 Fig.). These results clearly suggest that Tax requires CADM1 to inhibit TAX1BP1 phosphorylation and inactivation of A20, and to maintain persistent NF-κB activation in HTLV-1 infected T-cells.
The functional activities of CADM1 that have been described to-date include suppression of tumor growth (such as NSCLC), activation of NK or CD8+ T cells by serving as a tumor antigen, regulation of cell-cell interactions, and regulation of proper T-cell functions [59–61]. CADM1 expression is frequently down-regulated and its promoter hypermethylated in many human cancers, including lung, prostate, pancreatic, gastric, breast, esophageal and uterine cervix cancer [62,63]. In addition to the hypermethylation, Takai et al. have recently found that miR-214, miR-199, and hypoxia also down-regulate Necl-2 protein expression [64,65]. Restoration of CADM1 expression in NSCLC cells induced apoptosis and inhibited cell proliferation [66]. However, more than a 30-fold upregulation of CADM1 in HTLV-1-associated ATL tumor cells was critical for tumor cell progression and invasion [34]. Morishita et al. have reported that mice receiving CADM1-expressing T-lymphoma cell lines died due to massive tumor metastasis, suggesting that hyper-expression of CADM1 in T-lymphoma cells aggressively promotes leukemia/lymphoma [48]. However, the mechanistic role of CADM1-mediated leukemia/lymphoma cell proliferation and survival is not known. Although ATL cell survival and proliferation is partly dependent on persistent NF-κB activation [67,68] the mechanism of NF-κB activation in ATL cells is poorly understood.
In the current study, we have demonstrated that the HTLV-1 Tax protein induces the aberrant expression of CADM1 through NF-κB and CREB-dependent activation pathways in MEFs, consistent with recent studies by Kim et al. that demonstrated CADM1 mRNA and protein expression in Jurkat T-cells [46]. The findings from our study demonstrate that Tax preferably interacts with the cytoplasmic tail of CADM1 in HTLV-1 transformed T-cell lines. Moreover, Tax requires CADM1 for its K63-linked polyubiquitination, NF-κB activation, and inactivation of the NF-κB negative regulatory A20 complex.
Previous studies have demonstrated that Tax interactions with Ubc13, NEMO, TAX1BP1 and NRP are critical for activation of the IKK complex [15,19,53,69]. Our study suggests that CADM1 most likely recruits Ubc13 on Tax, which causes K63-linked polyubiquitination of Tax, and association of ubiquitin binding domain/motif containing NEMO, TAX1BP1 and NRP proteins to Tax. Stimulus-dependent activation of the IKK complex by TNFR and TCR engagement is generally initiated in the membrane-associated lipid rafts [70]. The upstream signaling molecules for these pathways (RIP1 for TNFR1, and ZAP70 and phosphatidylinositol 3-kinase for TCR) are the key proteins involved in IKK activation in the lipid rafts [71]. A previous study has demonstrated that Tax-mediated chronic NF-κB activation is initiated in lipid raft microdomains in intacT-cells [29]. Tax activates the IKK complex by interacting with its adaptor molecules in the membrane lipid rafts. Interestingly, Tax-mediated IKK complex activation in the membrane lipid rafts occurs in the absence of receptor engagement, suggesting that some lipid raft associated molecules are critical for Tax-mediated IKK complex activation. Another study has shown that cytosolic factors are sufficient to activate initial NF-κB activation in an in vitro cell free system, where lipid raft microdomains are absent, and claimed that the first round of NF-κB activation by Tax is critical for the induction of cytokines which are involved in NF-κB activation [57]. Although we also observed Tax-mediated IKK complex activation and the first round of NF-κB activation in an in vitro cell free system lacking lipid rafts, Tax failed to activate the IKK complex and the first round of NF-κB in CADM1-deficient intacT-cells. Tax-mediated activation of the IKK complex and the first round of NF-κB in the absence of lipid rafts in an in vitro cell free system is possibly due to easy and direct access to cytosolic factors that are normally assembled in response to upstream signals that activate the IKK complex in intacT-cells. We also observed in our study that the loss of Tax-NEMO interactions and Tax-mediated IKK activation in the absence of CADM1 is possibly due to lack of post-translational modifications on Tax. It is also possible that in the absence of CADM1 scaffolding function, Tax-associated molecules are not properly assembled in intact cells. More recent studies from Kim et al. demonstrated that CADM1 interacts with the ζ-chain of TCR to regulate TCR activation and T-cell interactions with APCs [46]. In agreement with this, our results indicate that membrane associated CADM1 is essential for Tax to interact with Ubc13, NEMO, TAX1BP1 and NRP and to activate the IKK complex in the membrane lipid rafts in the absence of cell stimulation (Figs. 6 and 8).
Previous studies have demonstrated that TAX1BP1 is a critical adaptor molecule for the NF-κB negative regulatory A20 complex in TNF-α or IL-1β stimulated cells [72]. IKKα-mediated phosphorylation of TAX1BP1 facilitates A20 complex assembly and subsequent inhibition of NF-κB activation in the TNF-α or IL-1β signaling pathways. Tax inhibits the phosphorylation of TAX1BP1 by disrupting IKKα and TAX1BP1 interactions as a mechanism of persistent NF-κB activation in TNF-α or IL-1 stimulated cells. [19]. However the mechanism of IKKα and TAX1BP1 interactions disruption has remained elusive. Here, we found that the stimulus-dependent phosphorylation of TAX1BP1 by IKKα is impaired in the absence of CADM1 in Tax expressing cells, suggesting that CADM1 is a crucial molecule for Tax to target the A20 complex (Figs. 7–8). It is possible that CADM1 serves as a critical lipid raft scaffold molecule for Tax, IKKα, TAX1BP1, and the other adaptor proteins. In future studies we will determine the mechanism of how CADM1 assist Tax to inhibit IKKα-mediated TAX1BP1 phosphorylation and A20 interaction.
Our results suggest that in HTLV-1 infected T-cell CADM1 does not have tumor-suppressor activity, but rather has gained tumor-promoting activity. It is highly likely that this switch in CADM1 function is triggered by multiple post-translational modifications including phosphorylation, SUMOylation, and ubiquitination. In future studies we will determine the post-translational modification in CADM1 that is responsible for this functional switch.
Generation of GZB-Tax transgenic mice was described previously [49], the use of tissues obtained from murine models in this study was carried in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Mice were housed under pathogen-free conditions according to the guidelines of the Division of Comparative Medicine and experiments were approved by the Animal Studies Committee, Washington University School of Medicine under ASC protocol #20100026. Tissues were removed from euthanized animals and placed on ice in PBS during the completion of the necropsy and then frozen on dry ice. Bone marrow was aspirated from long bones, and centrifuged at 2,500 rpm for 5 minutes. The supernatant was aspirated and the cell pellet was frozen on dry ice. The control was an age-matched, sex-matched Tax-negative littermate.
Generation of Cadm1 −/− mice was described previously [44], animals were housed under specific pathogen free conditions and experiments were carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was reviewed and approved by the University of Miami Institutional Animal Care and Use Committee (IACUC) (Protocol number: 12–104 RENEWAL 03).
The human T-cell lymphocytic cell line Jurkat was obtained from ATCC (Manassas, VA) and Tax-expressing HTLV-1-transformed T-cell lines MT-2, MT-4 and C8166 were obtained from the NIH AIDS Reagent Program. The NEMO-deficient JM4.5.2 cell line [69] was a gift from Dr. Sun Sc (The University of Texas MD Anderson, Houston, Texas). Jurkat, JM4.5.2, MT-2, MT-4 and C8166 cells were cultured in RPMI medium (Mediatech, Inc., Herndon, VA) supplemented with 10% fetal bovine serum, 100 U/ml penicillin, and 100 μg/ml streptomycin (Invitrogen, Carlsbad, CA). Cadm1+/+ and Cadm1 −/− MEFs were generated using a standard procedure [73]. Briefly, Cadm1 heterozygous mice described previously [44] were mated and E12.5 embryos were dissected free of surrounding tissues, washed in PBS (phosphate-buffered saline), and the heads and livers removed. The tissue was placed in 3 ml of 0.25% trypsin/EDTA and disrupted by forcing through a 6 cc syringe followed by vigorous pipetting, and the contents were transferred into a T25 tissue culture flask before placing in a tissue culture incubator at 370 C and 5% CO2 for 5 minutes. Cadm1+/+ and Cadm1 −/− MEFs were cultured in complete DMEM medium (Mediatech; Manassas, VA) containing 20% fetal bovine serum, heat inactivated, sterile-filtered (Sigma-Aldrich), L-glutamine, 1x penicillin-streptomycin (Invitrogen/Life Technologies). The plasmids pCAGI-Puro-FLAG-CADM1, pCAGI-Puro-FLAG-CADM1-ΔCP (deleting the cytoplasmic tail, aa 404–445), pCAGI-Puro-FLAG-CADM1-ΔEC (deleting the extracellular region, aa 1–362), pCAGI-Puro-FLAG-CADM1-ΔFERM (deleting the FERM domain-binding motif, aa 401–413), and pCAGI-Puro-FLAG- CADM1-ΔPDZ-BM (deleting the PDZ domain-binding motif, aa 442–445) were described previously [41]. The pCMV4-Tax, Tax M22, Tax M47 and NF-κB-TATA luciferase constructs have been described previously [32]. Tax M22 and M47 were constructed by replacing G137A and L138S, and L319R and L320S amino acid substitutions using a QuikChange site-directed mutagenesis kit (Stratagene, La Jolla, CA). All mutations were confirmed by DNA sequencing. The following antibodies were used in this study: anti-β-actin (Abcam), anti-TAX1BP1 (Abcam), anti-A20 (BD Biosciences Pharmingen and EMD Millipore), anti-phospho-TAX1BP1 described previously [19] was a gift from Dr. Edward Harhaj (Johns Hopkins School of Medicine), anti-ERK1/2, anti-phospho-TAK1, anti-phospho-IκBα, anti-phospho-IKKα/β (Cell Signaling), anti-CADM1 (MBL International Corporation), anti-CADM1, anti-TAK-1, anti-NEMO, anti-IKKα, anti-IKKβ, anti-IκBα (Santa Cruz Biotechnology), anti-Flag (Sigma), anti-NRP (Cayman Chemical), anti-LAT (Upstate Biotechnology), anti-Ubc13 (clone 4E11; Invitrogen), and antibody specific for ubiquitin Lys63 (HWA4C4; Millipore). Anti-Tax [31] was prepared from a Tax hybridoma (168B17-46-34) from the AIDS Research and Reference Program of the National Institute of Allergy and Infectious Diseases (US National Institutes of Health). Recombinant TNF and IL-1 were purchased from R&D Systems. The Optiprep was from Axis-Shield (Oslo, Norway).
To investigate the role of CADM1 in Tax-mediated NF-κB activation, a CADM1-specific shRNA construct was used to knockdown CADM1 expression. HEK 293-T-cells in 6-well culture plates were transfected with 1 μg of control scrambled shRNA or CADM1 shRNA with 2 μg of packaging plasmids (OriGene Technologies) containing puromycin selection marker using FuGENE 6 (Roche). Seventy-two hours post-transfection, the supernatants were collected and concentrated by ultracentrifugation and the pellets were resuspended in ice-cold PBS. Viral stocks were used to infect Jurkat T-cells, MEFs and HTLV-1 transformed (C8166, MT-2, and MT-4) cells and selected with puromycin. To overexpress Tax in MEFs and Jurkat T-cells, pCMV-Tax was used as a template for PCR-mediated cloning into the pDUET-GFP-hygromycin and pCDH-Cuo-MCS-EFI-GFP-T2A-puro lentiviral vector. Lentiviruses expressing Tax or control (GFP) empty vector were generated as described above. MEFs or Jurkat T-cells were infected with lentiviruses and selected with puromycin or hygromycin after 48 hours.
Transient transfections in MEFs and Jurkat T-cells were performed using FuGENE HD (Roche) according to the manufacturer's instructions. For luciferase assays, cells were harvested after 36–48 hours post-transfection and cell lysates were prepared in 1× Passive Lysis Buffer (Promega). Luciferase activity was assayed using the Dual Luciferase Assay system according to the manufacturer's instructions (Promega). All luciferase transfections included the Renilla luciferase reporter pRL-tk to normalize for transfection efficiency. Error bars indicate the standard error of the mean (s.e.m.) of triplicate samples from a representative experiment.
RT-PCR was done as described previously [15]. Total RNA was obtained from cells by using an RNeasy kit (QIAGEN, Valencia, CA) and converted to cDNA using a first-strand cDNA synthesis kit (Roche). The following sets of primers were used to amplify gene products for PCRs: glyceraldehyde-3-phosphate dehydrogenase (GAPDH) (263 bp) forward-5′-CCA CAG TCC ATG CCA TCA C-3’ and reverse-5′-GCT TCA CCA CCT TCT TGA TG-3’; Tax (429 bp) forward-5′-CGG ATA CCC AGT CTA CGT C-3’ and reverse-5′-GAG GTA CAT GCA GAC AAC GG-3’; IL-6 (460 bp) forward-5′-GAC TTC ACA GAG GAT ACC ACT C-3’ and reverse-5′-GTC CTT AGC CAC TCC TTC TG-3’; A20 (560 bp) forward-5′-GAC AGA AGT GTC CAG GCT TC-3’ and reverse-5′-GTG CTG GCT GTC ATA GCC TAG-3’; and CADM1 (477 bp) forward-5’-GAT GAT CGA TAT CCA GAA AGA CAC-3’ and reverse-5’-GTT TTG TTT AGG TTA TTG ATG AAC AG. Bfl-1 (394) forward-5’-TAC AGG TAC CCG CCT TTG AG-3’ and reverse-5’-TCT TCC CAA CCT CCA TTC TG-3’. Human SOCS1 (400) forward-5’- GACGCCTGCGGATTCTACTG-3’ and reverse-5’-GGAAGGAGCTCAGGTAGTCG-3’, mouse SOCS1 (453) forward-5’-GACACTCACTTCCGCACCTTCC-3’ and reverse-5’-GTCACGGAGTACCGGGTTAAGAG-3’.
Lipid raft fraction analysis was carried out as described previously [29]. Briefly, HTLV-1 transformed MT-2 cells were lysed in 2 ml of extraction buffer (20 mM Tris-Cl, pH 7.4, 150 mM NaCl, 1 mM EDTA, 1% Triton X-100 plus protease inhibitor cocktail). Lysates were combined with a 60% Optiprep solution to yield 40% and placed at the bottom of the ultracentrifuge tube followed by overlaying with an equal volume (4 ml) of discontinuous 30% and 5% OptiPrep Density Gradient medium. Samples were centrifuged at 100,000 × g for 4 hours at 4°C in an SW41 rotor. 1 ml of each fraction from the top to bottom was collected and equal volumes of each fraction were loaded onto SDS-PAGE gels. For the depletion of plasma and intracellular membrane cholesterol by MβCD in MT-2 cells cultured in RPMI medium supplemented with fetal bovine serum (10%) and penicillin-streptomycin (1%), were treated with or without 10 mM MβCD and incubated at 37°C for 45 min. Followed by this step lysates were subjected to density gradient ultracentrifugation for lipid raft fractionation analysis.
Tax was expressed from the pTaxH6 expression plasmid and purified as previously described [74,75].
Cytoplasmic extract from Jurkat, NEMO-deficient JM4.5.2 (Harhaj et al. 2000 Oncogene 19:1448–56), Cadm1+/+ and Cadm1 −/− MEFs were prepared as described earlier [57]. Briefly, cytoplasmic extracts were prepared by lysing the cells in a hypotonic buffer (10 mM Tris HCl (pH 7.5), 1.5 mM MgCl2, 10 mM KCl, 0.5 mM dithiothreitol (DTT) and a protease inhibitor cocktail (Roche)) and homogenized using a Dounce homogenizer. Lysates were placed on ice for another 10 minutes. After centrifugation at 100,000g for 1 hour at 4°C the supernatant (S100) was collected. Recombinant Tax was incubated in cytosolic extract (10 mg/ml) containing ATP buffer (50 mM Tris HCl (pH 7.5), 5 mM MgCl2, 2 mM ATP, 5 mM NaF, 20 mM β-glycerophosphate, 1 mM Na3VO4 and a protease inhibitor cocktail). After incubation at 30° C for 1 hour, the reaction mixtures were subjected to western blotting.
Ubiquitination assays were performed essentially as described previously [19,72]. Briefly, MEFs, Jurkat or HTLV-1 infected (C8166, MT-2 and MT-4) cells were lysed in RIPA buffer and immunoprecipitated with Tax antibody, and eluted with 1% SDS, diluted in lysis buffer, re-immunoprecipitated with Tax antibody, and detected by immunoblotting with antibodies to K63-Ubi or Tax. Similarly, the fractions obtained from density gradient ultracentrifugation or cells lysed in RIPA buffer were immunoprecipitated with specific antibodies. Immunoprecipitates were washed three times with respective buffers. Immunoblotting was performed with the indicated antibodies for co-IPs.
Immunoblotting was performed as described earlier [32,76]. Whole-cell lysates were resolved by SDS–PAGE, transferred to nitrocellulose membranes, blocked in 5% milk, incubated with the indicated primary and secondary antibodies and then detected with Western Lightning Enhanced Chemiluminescence reagent (Perkin Elmer).
The cells were seeded onto 12-mm poly-L-lysine-coated coverslips (BD Biosciences, Bedford, MA) and were briefly centrifuged prior to fixation. The cells were washed three times with PBS and fixed in 4% paraformaldehyde for 15 minutes at room temperature. The fixed cells were permeabilized with PBS containing 0.2% Triton X-100, and nonspecific binding was prevented by a 1 hour incubation in SuperBlock buffer (Thermo Scientific) followed by staining with primary antibodies: mouse anti-Tax, rabbit anti-CADM1 (Santa-Cruuz biotechnology), anti-Ubc13, anti-NEMO, anti-NRP (Fisher Scientific), anti-TAX1BP1, anti-GM130 (Abcam), and chicken anti-CADM1 (MBL International Corporation), diluted in PBS containing 1% BSA and incubated for 2 hours followed by five washes with PBS containing 1% BSA. Secondary antibodies: Alexa Fluor 555- donkey anti-mouse or Alexa Fluor 647- donkey anti-mouse IgG (for Tax), Alexa Fluor 488-donkey anti-rabbit IgG (for CADM1, TAX1BP1, Ubc13, NEMO, NRP), Alexa Fluor 647- donkey anti-rabbit IgG (for Golgi-130), Alexa Fluor 555-conjugated cholera toxin subunit B (Invitrogen), and Cy2 donkey anti-Chicken IgG, (for CADM1) Jackson ImmunoResearch) were incubated for 45 min followed by four washes with PBS. The cells were then incubated with DAPI 500ng/ml (Sigma). After washing three times with PBS, the coverslips were mounted onto the glass slides with ProLong Gold anti-fade reagent (Invitrogen) and then observed under SP5 confocal microscope (Leica).
Nuclear extracts were prepared from HTLV-1 transformed (C8166, MT-2, and MT-2), and MEFs cells. The NF-κB electrophoretic mobility shift assay (EMSA) was done as described previously [15,32]. The Oct-1 EMSA probe was generated by annealing the following oligonucleotides: forward 5′-TGTCGAATGCAAATCACTAGAA-3’ and reverse 5′-TTCTAGTGATTTGCATTCGACA-3’. The annealed oligonucleotides were labeled with (32P) dTTP in a fill-in reaction with Klenow fragment (Promega). Nuclear extract (4 μg) was incubated with buffer containing 1 mM dithiothreitol, 1 μg poly(dI-dC), dialysis buffer (25 mM HEPES, pH 7.9, 10% glycerol, 100 mM KCl, and 0.1 mM EDTA), and 32P-labeled probe for 15 minutes. The reaction was terminated by the addition of 5× loading dye, and the reaction mixture was run on 5% polyacrylamide gels in 0.25× Tris-borate-EDTA buffer, dried under vacuum, and subjected to autoradiography.
|
10.1371/journal.ppat.0030050 | Transcriptional Regulation of Chemical Diversity in Aspergillus fumigatus by LaeA | Secondary metabolites, including toxins and melanins, have been implicated as virulence attributes in invasive aspergillosis. Although not definitively proved, this supposition is supported by the decreased virulence of an Aspergillus fumigatus strain, ΔlaeA, that is crippled in the production of numerous secondary metabolites. However, loss of a single LaeA-regulated toxin, gliotoxin, did not recapitulate the hypovirulent ΔlaeA pathotype, thus implicating other toxins whose production is governed by LaeA. Toward this end, a whole-genome comparison of the transcriptional profile of wild-type, ΔlaeA, and complemented control strains showed that genes in 13 of 22 secondary metabolite gene clusters, including several A. fumigatus–specific mycotoxin clusters, were expressed at significantly lower levels in the ΔlaeA mutant. LaeA influences the expression of at least 9.5% of the genome (943 of 9,626 genes in A. fumigatus) but positively controls expression of 20% to 40% of major classes of secondary metabolite biosynthesis genes such as nonribosomal peptide synthetases (NRPSs), polyketide synthases, and P450 monooxygenases. Tight regulation of NRPS-encoding genes was highlighted by quantitative real-time reverse-transcription PCR analysis. In addition, expression of a putative siderophore biosynthesis NRPS (NRPS2/sidE) was greatly reduced in the ΔlaeA mutant in comparison to controls under inducing iron-deficient conditions. Comparative genomic analysis showed that A. fumigatus secondary metabolite gene clusters constitute evolutionarily diverse regions that may be important for niche adaptation and virulence attributes. Our findings suggest that LaeA is a novel target for comprehensive modification of chemical diversity and pathogenicity.
| Patients with suppressed immune systems due to cancer treatments, HIV/AIDS, or organ transplantation are at high risk of infection from microbes. Some of the most deadly infections for such patients arise from a fungal pathogen, Aspergillus fumigatus. This species, like several of its close relatives, can produce an array of small chemical compounds that influences both the infection process and its environmental niche outside of the host. The genes dedicated to production of each compound are clustered adjacent to each other in the genome. One protein named LaeA is a master regulator of such clustered small molecule genes, and removal of the gene encoding LaeA cripples the organism's ability to infect. We conducted a genome-wide microarray experiment to identify small molecule gene clusters controlled by the presence of LaeA in A. fumigatus. In doing so, we identified actively expressed gene clusters critical for small molecule production and potentially involved in disease progression. These results also provide insight into evolutionary events shaping the organism's collection of chemical compounds.
| Aspergillus fumigatus is a saprophytic filamentous fungus with no known sexual stage. Prolific production of asexual spores (conidia) and nearly ubiquitous distribution in the environment ensures constant host exposure to its spores, at a density of 1 to 100 conidia/m−3 [1]. The innate immune system enables spores to be eliminated from lung epithelial tissue with ease in immunocompetent vertebrates. However, immunocompromised individuals are at risk for pulmonary disease as a consequence of A. fumigatus infection. Of particular concern is invasive aspergillosis, which occurs when hyphal growth proliferates throughout pulmonary or other tissues. Invasive aspergillosis has an associated mortality rate ranging from 50% to 90% depending on the patient population [2]. As the number of immunocompromised patients has increased in recent decades due to immunosuppressive chemotherapy treatments, HIV/AIDS, and solid organ and bone marrow transplantation, the incidence of invasive aspergillosis has increased more than 4-fold in developed nations [2].
Several A. fumigatus secondary metabolites or natural products (e.g., conidial melanins and mycotoxins) have been implicated as affecting virulence [3–7]. However, the exact mechanisms by which many of these compounds might affect disease outcome are unknown, nor is it clear in most cases whether these factors play direct or indirect roles in pathogenicity. In contrast to most genes involved in primary metabolism, genes encoding secondary metabolite biosynthetic enzymes exist in contiguous clusters within the genome [8,9]. LaeA was originally identified as a transcriptional regulator of secondary metabolite gene clusters in Aspergillus nidulans and A. fumigatus [10,11], including gliotoxin in the latter. Gliotoxin has long been suggested to be a major virulence attribute in invasive aspergillosis [12–14]. However, whereas a ΔlaeA mutant shows reduced virulence in a mouse model of invasive aspergillosis [11], inactivation of gliotoxin biosynthesis alone does not [15–17]. Therefore, we reasoned that because LaeA is a transcriptional regulator, perhaps acting at a chromatin remodeling level [9,18], a microarray experiment comparing the transcriptomes of ΔlaeA, wild-type, and complemented ΔlaeA control strains would yield further insight into LaeA-mediated A. fumigatus virulence attributes. We uncovered an unprecedented view of LaeA global regulation of mycotoxin islands, nearly all found in nonsyntenic regions of the Aspergillus genome. Because secondary metabolite gene cluster regions are evolutionarily diverse and may affect virulence attributes, LaeA is a novel target for comprehensive modification of chemical diversity.
Transcriptional profiles of the wild-type, ΔlaeA, and ΔlaeA complemented strain were determined by comparisons of relative transcript levels between (1) ΔlaeA versus wild-type and (2) wild-type versus complemented control strain. All strains were grown under identical conditions (25 °C, liquid shaking culture, glucose minimal media, 60 h) for three biological replicates. The condition and time point were chosen on the basis of optimal production of secondary metabolites [10,11]. The comparison of ΔlaeA versus wild-type was used to determine gene expression patterns specific to the ΔlaeA mutant, while the wild-type versus complemented strain comparison was conducted as a control, because the difference between these two strains is the presence of an ectopic copy of a selectable marker for hygromycin resistance. The processed signal intensity ratios for the three ΔlaeA versus wild-type replicates were analyzed using the significance analysis of microarrays (SAM) method [19], as described in Materials and Methods. In total, 943 genes were significantly differentially expressed. Figure S1 shows a heat map of a subset of these loci, depicting normalized expression ratios for the three ΔlaeA versus wild-type experiments and the three wild-type versus complemented control experiments. The high quality of the data is indicated by the consistency of color between the replications and the relative lack of color in the control lanes.
Of the 943 genes showing significant differences in expression between ΔlaeA and wild-type by SAM analysis, 415 showed increased expression in ΔlaeA and 528 showed decreased expression. Table 1 and Figure S2 indicate functional categories for these genes (defined as described by the Gene Ontology Consortium, http://www.geneontology.org). The most remarkable discovery was the near-global suppression of secondary metabolite gene expression in the ΔlaeA mutant. Nearly all (97%) of the secondary metabolite gene cluster loci showed decreased expression in ΔlaeA, with a mere three genes in this category showing increased expression in ΔlaeA. This was in contrast to all other functional categories, which showed substantial proportions of both increased and decreased expression in the mutant, possibly reflecting indirect effects due to loss of production of multiple metabolites. In addition to genes with unknown function (39%) and genes involved in secondary metabolism (11%), other major categories included genes encoding proteins involved in transmembrane transport (8%) and those involved in information processing (4%), and cell wall biogenesis (4%). Statistical analysis of the overrepresentation of different Gene Ontology categories and Pfam protein domains within the set of 943 differentially regulated genes is shown in Tables S1 and S2, respectively.
Interestingly, LaeA appeared to influence expression of a subset of species- and lineage-specific genes not strongly conserved with other fungal species. Only 18% and 44% of all genes significantly differentially expressed in the mutant have putative orthologs in Saccharomyces cerevisiae and Neurospora crassa, respectively, compared to an average of 33% and 58% of all A. fumigatus genes. Many, but not all, of these genes were classified as secondary metabolism genes. Moreover, there are about 120 differentially expressed genes; again, most, but not all, are present in secondary metabolism clusters (Table 2), which have no detectable orthologs in Aspergillus oryzae and A. nidulans. Considering this overwhelming tight and directed transcriptional control of secondary metabolite loci by LaeA, below we focus on such genes as possible members of the LaeA-regulated A. fumigatus pathogenicity arsenal.
Although initial genome analysis suggested the presence of 26 secondary metabolite gene clusters [20], subsequent analysis (G. Turner, N. D. Fedorova, V. Joardar, J. R. Wortman, and W. C. Nierman, unpublished data) has provided support for only 22 clusters. Of the 13 secondary metabolite gene clusters whose expression was influenced by LaeA in the condition used for microarray analysis, ten are particularly strongly affected, with a majority of genes within these clusters being significantly down-regulated in ΔlaeA as indicated by SAM. Three additional clusters have at least one gene encoding a critical enzyme such as a nonribosomal peptide synthetase (NRPS) or a polyketide synthase showing decreased expression in ΔlaeA. Additionally, 38% (23 of 71) of all P450 monooxygenases show differential expression in ΔlaeA, also associated with secondary metabolite biosynthesis and/or detoxification. Fifteen of these genes encoding P450 monooxygenases are found in secondary metabolite gene clusters. Table S3 gives normalized expression ratio values for all 22 gene clusters in A. fumigatus. Table 2 summarizes the current state of knowledge regarding function of LaeA-regulated secondary metabolite gene clusters. These include clusters dedicated to production of conidial melanins, fumitremorigens, gliotoxin, and ergot alkaloids such as festuclavine, elymoclavine, and fumigaclavines A, B, and C (Table 2) [4,17,21–27]. Figure 1 depicts the chromosomal landscape of those regions most strongly regulated by LaeA.
To confirm these microarray results, quantitative real-time reverse-transcription (RT)-PCR (QRT-PCR) was performed on one major class of secondary metabolite genes, those encoding NRPSs [28]. As indicated in Table 3, relative expression levels for NRPSs that showed differential expression between mutant and wild-type in the microarray study also were dramatically reduced upon QRT-PCR analysis. In all cases, complementation of the laeA defect restored NRPS gene expression to wild-type levels (Table 3).
Notably, because the microarray analysis determines only relative expression and not absolute levels of transcript, we could not conclude whether secondary metabolite clusters not showing differential expression in mutant versus wild-type were not affected by LaeA or were simply not induced under the growth condition used. To further examine these possibilities, we assessed the expression of a subset of NRPSs thought to encode siderophore-biosynthesizing enzymes. Although siderophores do not fit neatly into a definition of secondary metabolites, which are dispensable in laboratory growth conditions [29], these molecules are produced from clustered genes and are critical for pathogen growth in blood serum [30]. Because iron was included in the media used for the microarray study, we investigated whether the Δlae mutant was deficient in expression of siderophore gene cluster NRPSs under iron-limiting conditions.
As previously reported [29], low iron conditions induced transcriptional upregulation of several NRPS genes known or predicted to be involved in siderophore biosynthesis (Table 4). Normalized expression levels of the siderophore NRPSs in the low iron condition relative to high iron conditions were NRPS2/sidC, 2.245 ± 0.449; NRPS3/sidE, 68.595 ± 13.725; and NRPS4/nps6/sidD, 28.509 ± 4.704. NRPS7 transcripts were not detectable in these experiments. In contrast to Reiber et al. [29], NRPS3/sidE showed the highest induction to the low iron conditions in our experiments. This discrepancy might be explained by differential sensitivity of the semiquantitative RT-PCR method used by Reiber et al. compared to our QRT-PCR methodology or subtle differences in culture conditions. We also noted that the complemented control strain with an ectopic copy of laeA showed increased expression of NRPS3/sidE in both low and high iron conditions.
Comparison of the ΔlaeA mutant and controls by QRT-PCR analysis indicated differential expression of NRPS3/sidE in low (inducing) iron conditions. In high iron conditions, NRPS4/nps6/sidD, NRPS3/sidE, and possibly NRPS2/sidC showed decreased expression in the ΔlaeA mutant (Table 4). Interestingly, NRPS7 was not detectable in these experiments. In the low iron condition, expression of actin did decrease in the ΔlaeA mutant. However, the dramatic decrease in expression of NRPS3/sidE (1,000-fold less) seen in the ΔlaeA background strongly suggests that LaeA regulates the expression of at least this NRPS. Little is known about the function of SidE, although it has been speculated to be involved in siderophore biosynthesis on the basis of homology to SidC [29]. It remains to be determined whether NRPS3/sidE is involved in siderophore production, a process known to be critical to virulence [31,32], or whether it synthesizes an iron-responsive compound with a distinct function. Regardless of the function of SidE, these experiments show that LaeA is also involved in controlling expression of other secondary metabolite clusters not induced by the environmental conditions used in the microarray experiments.
Cluster 18 (Figure 1) on Chromosome 6, strongly differentially expressed in ΔlaeA, encodes the genes required for gliotoxin biosynthesis. Gliotoxin is arguably the most well-studied mycotoxin produced by A. fumigatus. First identified in 1936, this compound has immunosuppressive properties in vitro [12] and in vivo [13,14], although its direct contribution to pathogenicity is only beginning to be understood [15–17]. Like all other compounds in the epipolythiodioxopiperazine class, gliotoxin is a cyclic dipeptide with an internal disulfide bridge that can undergo redox cycling (for a recent review, see [33]). Immunosuppressive activity of gliotoxin is due at least in part to negative regulation of the transcription factor nuclear factor–κB, which occurs by inhibition of proteasome-mediated degradation of the nuclear factor–κB inhibitor IκBα [34,35]. Gliotoxin is also known to be cytotoxic and can evoke both apoptotic [36–39] and necrotic [40,41] cell death. Recently, gliotoxin was shown to trigger the release of apoptogenic factors by the host mitochondrial protein Bak [42]. The secondary metabolism gene cluster responsible for gliotoxin production was recently identified by bioinformatic analysis [43] and has been experimentally confirmed [15,17]. Despite the known immunosuppressive activities of the molecule and its detection in blood serum of patients with invasive aspergillosis [44], three recent studies using genetic mutants of the gliotoxin gene cluster demonstrated that gliotoxin is not a virulence factor in murine models of invasive aspergillosis [15–17]. However, these same studies presented evidence that gliotoxin could adversely affect T cells, neutrophils, and mast cells and, we offer, likely acts synergistically with other LaeA-regulated toxins. The ΔlaeA mutant is impaired in gliotoxin production during growth in culture as well as growth in vivo in murine models of invasive aspergillosis [10,15], and the microarray results presented here confirm that LaeA strongly influences expression of genes in this cluster under the condition investigated.
Secondary metabolite cluster 1 on Chromosome 1, which is differentially expressed in ΔlaeA, contains an atypical NRPS called Afpes1 that is required for virulence in an insect model of invasive aspergillosis [4]. Afpes1 shows greatest homology to NRPSs that produce siderophores or destruxins, including one paralog required for virulence of the plant pathogen Alternaria brassicae [45]. However, the Afpes1 cluster is thought to be unlikely to produce either of these compounds, because destruxin toxin has not been detected in A. fumigatus [4] and expression of Afpes1 was not responsive to iron levels [4,21]. Deletion of Afpes1 alters conidial morphology and hydrophobicity as well as melanin synthesis and results in increased susceptibility to reactive oxygen species, implying altered conidial melanin and/or rodlet composition [4]. Most of these characteristics are common to the ΔlaeA phenotype [11], possibly implicating a role of the Afpes1 metabolite in the attenuated virulence of ΔlaeA.
A. fumigatus synthesizes several clavine ergot alkaloids, compounds that can be partial agonists or antagonists of serotonin, dopamine, and α-adrenalin receptors, thus affecting nervous, circulatory, reproductive, and immune system function [46]. The role of these compounds in invasive aspergillosis has not been determined. In addition to having the receptor-modulating activities mentioned, the festuclavine ergot alkaloid produced by A. fumigatus is cytostatic and is directly mutagenic in the Ames assay [47,48]. Recently, Coyle and Panaccione [25] showed that deletion of an A. fumigatus dimethylalleletryptophan synthase (DMAT synthase) homologous to dmaW of the ergot-producing species Claviceps purpurea eliminated all known ergot alkaloids, confirming its predicted function in the first committed step of ergot alkaloid production (i.e., addition of dimethylallyl diphosphate to l-tryptophan to result in 4-methylallyl-tryptophan). The biochemical activity of the A. fumgiatus DmaW enzyme was also confirmed by Unsöld and Li [22], who subsequently characterized a reverse prenyltransferase in the same gene cluster that converts fumigaclavine A to fumigaclavine C [23]. These genes are located in secondary metabolite gene cluster 4 on Chromosome 2, which is strongly differentially expressed in ΔlaeA.
Melanins found in conidia are one of the few described virulence factors in A. fumigatus [5,6,24]. Lack of melanins leads to increased susceptibility to reactive oxygen species produced by the host innate immune response during infection as well as altered (smooth) conidial morphology [5,7]. However, the scarcity of nonpigmented A. fumigatus spores in nature has drawn into question the clinical relevance of melanins as virulence factors [1]. Conidia of ΔlaeA are pigmented, but altered expression of alb1 in the mutant has been reported previously and at least one unidentified spore metabolite is missing in ΔlaeA [11]. There is significant differential expression of the 1,8-dihydroxynapthalene–melanin gene cluster in ΔlaeA under the condition investigated in this study. Expression of this gene cluster is also regulated by cAMP/protein kinase A signaling [49] as is LaeA itself [10], perhaps a suggestion that in this case LaeA control of this cluster may be both directly and indirectly mediated by protein kinase A.
Additionally, a LaeA-regulated supercluster on Chromosome 8 is likely to produce multiple compounds. Recently, two genes in this cluster have been reported to encode biosynthetic enzymes for the tremorgenic mycotoxin fumitremorgin B and related compounds [26,27]. The cyclo-l-Trp-l-Pro derivative fumitremorgin B is cytotoxic, inhibiting cell cycle progression at G2/M, and thus has been of interest as a potential anticancer agent. The pathway involves generation of the cyclic dipeptide brevianamide F by the NRPS brevianamide synthetase [27], prenylation of brevianamide F by the prenyltransferase FtmPT1 to tryptrostatin B [26], and subsequent conversion in several steps to fumitremorgen B. Thus, LaeA-mediated influence on expression of ftmPT1 and ftmPT2 would govern the production of this entire class of diketopiperozine compounds. Once again, however, the specific effects of these compounds on pathogenicity during invasive aspergillosis are unknown.
The fact that LaeA promotes expression not only of these secondary metabolite gene clusters but an additional eight others confirms its role as a master controller of secondary metabolism. The importance of several of these compounds in toxicity studies also underscores relevance of LaeA during infection [11]. We suggest the possibility that virulence attributes are not influenced as much by individual metabolites as by the blend of LaeA-regulated toxins, which, in combination, may confer an advantage to the pathogen.
Comparative genomic analysis between A. fumigatus and related species indicates overlap between A. fumigatus–specific genes and genes differentially expressed in ΔlaeA (N. D. Fedorova and W. C. Nierman, unpublished data). In total, 68% of A. fumigatus secondary metabolite genes do not have orthologs in the closely related species A. clavatus (N. D. Fedorova and W. C. Nierman, unpublished data). Additional secondary metabolite genes do not have orthologs in more distantly related Aspergilli such as A. oryzae and A. nidulans [20]. The variability of secondary metabolite clusters may be explained by the fact that many of them are located in highly divergent telomere-proximal regions characterized by frequent chromosomal rearrangements [20,50]. For example, 54% of the clusters showing differential expression in ΔlaeA in the conditions described here were found within 300 kb of telomeres. It should be noted that, in addition to the secondary metabolite clusters, other genes with significantly lower expression in ΔlaeA also show some positional specificity within the genome but to a much lesser extent (unpublished data). Further analysis also showed that A. fumigatus telomere-proximal clusters tend to have larger numbers of genes than clusters located closer to the centromeres, suggesting that the former may accumulate additional genes more easily (N. D. Fedorova, J. R. Wortman, and W. C. Nierman, unpublished data).
Initial comparative genome analyses indicate that the telomere-proximal regions (and to a lesser extent, synteny breakpoints and intrasyntenic regions) appear to be a hotbed of diversity, not only between Aspergillus species but even between different strains of the same species [51,52]. The genomes of two A. fumigatus strains have been sequenced: the clinical isolate Af293 (by The Institute for Genomic Research, Rockville, Maryland, United States) and isolate CEA10 (under contract from Elitra Pharmaceutical by Celera Genomics and made available by Merck; B. Jiang and W. C. Nierman, personal communication). These strains show an overall divergence of 2%, and the majority of this variation is in telomere-proximal and synteny breakpoint regions. Similarly, microarray experiments also supported high divergence in these regions when Af293 was compared to the unsequenced A. fumigatus strains Af294 and Af71 [20].
Many secondary metabolite clusters appear to be associated with transposons (known to be active in A. fumigatus [51]) and transposase-like sequences (Table S3). The finding that these transposable elements often flank or are embedded in many of the clusters may represent one mechanism for generating the diversity of secondary metabolites in aspergilli. Whether or not there is a connection between LaeA function and transposon activity has yet to be established. In total, these analyses suggest that secondary metabolite clusters are located in the regions that undergo extensive rearrangements, which may result in subsequent alterations in secondary metabolite production and, therefore, have major impacts on niche adaptation between different species of fungi or between strains of the same species. Other examples include a non–aflatoxin-producing clade of A. flavus, better known as the food-fermenting A. oryzae used in the production of traditional Asian products such as miso and soy sauce, which may have arisen as a result of telomere-proximal rearrangements [53]. Similarly, genotypic variability between strains of Fusarium compactum also proved to be a major determinant of metabolite production and geographic distribution [54]. In Fusarium graminearum, the major cause of wheat and barley head blight, intraspecific polymorphic variations in a trichothecene mycotoxin gene cluster were correlated with chemotype differences, host range, and fitness [55].
In light of such examples, it is interesting to speculate about the role of LaeA in chemotype evolution and niche adaptation. It is possible that variation at any particular secondary metabolite gene cluster could result in less efficient control by LaeA. This potential has been demonstrated in A. nidulans [18]. Conversely, LaeA itself is a major target for comprehensive changes in the entire complement of secondary metabolites. The clustering of secondary metabolite biosynthetic genes has been suggested to reflect their evolutionary history [8,9,20,51,56,57]. Several models have been proposed to explain the establishment and maintenance of secondary metabolic gene clusters in filamentous fungi. The “selfish cluster” hypothesis proposes that selection occurs at the level of the cluster and promotes maintenance of the cluster as a unit, possibly through horizontal transfer events [56]. However, there is only limited evidence for widespread horizontal transfer of secondary metabolism gene clusters, with penicillin being a notable exception [58]. Alternative models suggest that clusters are maintained due to coregulation mechanisms, likely at the level of chromatin regulation [8,9]. LaeA may provide a mechanistic means of secondary metabolism gene cluster coregulation and maintenance. Certainly LaeA demonstrates a positional bias for local gene regulation, as transfer of genes into or out of a secondary metabolite cluster leads to respective gain or loss of transcriptional regulation by LaeA [18]. This has been speculated to occur through regulation of nucleosome positioning and heterochromatin formation [9].
Our results confirm that LaeA plays a central role in regulation of chemical diversity in A. fumigatus. Furthermore, genomic regions that are transcriptionally controlled by LaeA are species and even strain specific, suggesting that they may serve as niche adaptation factors. The loss of laeA results in a great decrease in repertoire of secondary metabolites, which appears to impact the infection process. Therefore, LaeA constitutes a novel target for the production of an array of factors critical to success during pathogenesis. Furthermore, LaeA is a tool to identify metabolite gene clusters that may impact virulence, allowing the correlation of specific secondary metabolite clusters with virulence even in absence of knowledge about the mycotoxin itself.
Three prototrophic A. fumigatus fungal strains were used in this study. Af293 (the wild-type clinical isolate used in the A. fumigatus genome sequencing project [20]), TJW54.2 (ΔlaeA) [11], and a complemented control strain TJW68.6 (ΔlaeA + laeA) [11] were grown in triplicate at 25 °C in liquid minimal media [59] with shaking (280 rpm) for 60 h. Profiles of secondary metabolites extracted from the media with chloroform were compared by thin-layer chromatography, and the results confirmed that the ΔlaeA strain showed reduced levels of multiple secondary metabolites under this condition ([10] and unpublished data). Total RNA was isolated from fungal mats, labeled, and hybridized with a DNA whole-genome amplicon microarray [20,60] in three independent biological replicates.
To analyze siderophore NRPS gene expression under low- or high-iron conditions, 50-ml liquid cultures were grown as described [29], with low-iron media containing 25 g/L glucose, 3.5 g/L (NH4)2SO4, 2.0 g/L KH2PO4, 0.5 g/L MgSO4 (heptahydrate), and 8 mg/L ZnSO4 (heptahydrate) (pH 6.3). High-iron media was identical except for the addition of Fe(III)Cl3 to a final concentration of 300 μM. Cultures were grown at 37 °C, 280 rpm, and samples were collected at 24 h postinoculation. All glassware was subjected to sequential treatment with 1 mM and 5% HCl as described [29].
Total RNA was extracted from Aspergillus strains by use of TriZOL reagent (Invitrogen, http://www.invitrogen.com) according to the manufacturer's instructions. RNA was further purified by two extractions with phenol:chloroform:isoamyl alcohol (25:24:1) and then labeled with Cy-3 or Cy-5 dye and hybridized as previously described [20]. The generation of the whole genome array has been described [20].
QRT-PCR was used to (1) confirm the expression level trends observed in the microarray experiment and (2) investigate NRPS gene expression under iron-limiting conditions. Expression of select NRPSs putatively regulated by LaeA was examined. Total RNA from two or three biological replicates was pooled in equal amounts (2 μg per sample) for each Aspergillus strain, wild-type AF293, TW54.2, and TW68.6, and treated with Ambion Turbo DNA-free DNase I (Ambion, http://www.ambion.com) to remove contaminating genomic DNA. A total of 500 ng of DNase I–treated total RNA from each sample was reverse transcribed with Superscript III reverse transcriptase (Invitrogen). Real-time RT-PCR was conducted with 20-μl reaction volumes with the iQ SYBR green supermix (Bio-Rad, http://www.bio-rad.com), 2 μl of a 1:6 dilution of first-strand cDNA, and 0.4 μl of each 10 μM primer stock. Primer sequences were previously reported [28]. No reverse transcriptase controls (NRT) were used to confirm elimination of contaminating genomic DNA. Real-time RT-PCR was performed using an iQ Cycler Real-Time PCR detection system (Bio-Rad). PCRs for each NRPS were done in triplicate and melt curve analysis was performed immediately following the PCR to confirm the absence of nonspecific amplification products and primer dimers. The relative expression levels of NRPS genes between A. fumigatus wild-type strain AF293, the ΔlaeA mutant, and the complemented control strain were calculated using 2−ΔΔCt method with iQ cycler system software. All values were normalized to expression of the A. fumigatus actin gene and relative to the wild-type strain for each condition analyzed.
Gene expression ratios were determined for triplicate comparisons of (1) wild-type and ΔlaeA and (2) ΔlaeA and the complemented control strain. Prior to statistical analysis, the LOWESS normalization method was used to remove any systematic bias from the raw expression ratios [61]. Loci showing significantly different expression were identified using the SAM method for one-class designs that has been previously described in detail [19], implemented in the TM4 suite's MultiExperiment Viewer (http://www.tm4.org) [62,63]. This allowed identification of genes whose mean expression across experiments is significantly different from a user-specified mean (log2 = 0, corresponding to identical mRNA levels in the mutant and wild-type strains). Genes with scores above the significance threshold and exceeding the cutoff value of zero for the false discovery rate (the most conservative setting) were designated as significantly differentially expressed between mutant and wild-type. The delta value cutoff in SAM was chosen to capture the maximum number of significant genes while maintaining the reported estimated false discovery rate at zero. Genes down-regulated in ΔlaeA were further analyzed by the Expression Analysis Systematic Explorer (EASE) [64] within TM4 to identify overrepresented Gene Ontology terms and Pfam domains. Fisher's exact test probabilities and step-down Bonferroni corrected probabilities are reported from the EASE analysis to indicate which terms are overrepresented in the down-regulated gene set. |
10.1371/journal.pntd.0001998 | Typhoid Fever and Its Association with Environmental Factors in the Dhaka Metropolitan Area of Bangladesh: A Spatial and Time-Series Approach | Typhoid fever is a major cause of death worldwide with a major part of the disease burden in developing regions such as the Indian sub-continent. Bangladesh is part of this highly endemic region, yet little is known about the spatial and temporal distribution of the disease at a regional scale. This research used a Geographic Information System to explore, spatially and temporally, the prevalence of typhoid in Dhaka Metropolitan Area (DMA) of Bangladesh over the period 2005–9. This paper provides the first study of the spatio-temporal epidemiology of typhoid for this region. The aims of the study were: (i) to analyse the epidemiology of cases from 2005 to 2009; (ii) to identify spatial patterns of infection based on two spatial hypotheses; and (iii) to determine the hydro-climatological factors associated with typhoid prevalence. Case occurrences data were collected from 11 major hospitals in DMA, geocoded to census tract level, and used in a spatio-temporal analysis with a range of demographic, environmental and meteorological variables. Analyses revealed distinct seasonality as well as age and gender differences, with males and very young children being disproportionately infected. The male-female ratio of typhoid cases was found to be 1.36, and the median age of the cases was 14 years. Typhoid incidence was higher in male population than female (χ2 = 5.88, p<0.05). The age-specific incidence rate was highest for the 0–4 years age group (277 cases), followed by the 60+ years age group (51 cases), then there were 45 cases for 15–17 years, 37 cases for 18–34 years, 34 cases for 35–39 years and 11 cases for 10–14 years per 100,000 people. Monsoon months had the highest disease occurrences (44.62%) followed by the pre-monsoon (30.54%) and post-monsoon (24.85%) season. The Student's t test revealed that there is no significant difference on the occurrence of typhoid between urban and rural environments (p>0.05). A statistically significant inverse association was found between typhoid incidence and distance to major waterbodies. Spatial pattern analysis showed that there was a significant clustering of typhoid distribution in the study area. Moran's I was highest (0.879; p<0.01) in 2008 and lowest (0.075; p<0.05) in 2009. Incidence rates were found to form three large, multi-centred, spatial clusters with no significant difference between urban and rural rates. Temporally, typhoid incidence was seen to increase with temperature, rainfall and river level at time lags ranging from three to five weeks. For example, for a 0.1 metre rise in river levels, the number of typhoid cases increased by 4.6% (95% CI: 2.4–2.8) above the threshold of 4.0 metres (95% CI: 2.4–4.3). On the other hand, with a 1°C rise in temperature, the number of typhoid cases could increase by 14.2% (95% CI: 4.4–25.0).
| This research studies the spatial and temporal distribution of typhoid infections in the Dhaka metropolitan area of Bangladesh in the period 2005 to 2009. Data from hospital admission records was analysed together with a range of demographic, environmental and climatic data, in what is believed to be the first study of this nature; clear periodicity was found in the timing of case occurrences, with most cases occurring in the monsoon season. Men and very young children appear to be at greatest risk of contracting the disease. Closeness to rivers was also found to be a contributor to increased typhoid risk. While a difference in rates between urban and rural locations suggested by other studies was not found, distinct clustering of the disease was uncovered. Two of these clusters are located in central Dhaka with a third in the north of the metropolitan area.
| Typhoid fever is one of the leading causes of morbidity and mortality across the world [1].Typhoid is caused by a bacterium of the genus Salmonella. Salmonella infection in humans can be categorised into two broad types, that caused by low virulence serotypes of Salmonella enterica which cause food poisoning, and that caused by the high virulence serotypes Salmonella enterica typhi (S. typhi), that causes typhoid,and a group of serovars, known as S Paratyphi A, B and C, which cause Paratyphoid [2]. Humans are the only host of this latter group of pathogens. S. Typhi is a highly adapted human-specific pathogen [3], and the illness caused by these bacteria is a serious public health concern, particularly in developing countries [4]. A recent estimate found that 22 million new typhoid cases occur each year in the world with some 200,000 of these resulting in death [5], indicating that the global burden of this disease has increased steadily from a previous estimate of 16 million [6] however, case-fatality rates have decreased markedly [5]. The highest number of cases (>100 per 100,000 persons/year) and consequent fatalities are believed to occur in South Central and Southeast Asia [1]. Generally, typhoid is endemic in impoverished areas of the world where the provision of safe drinking water and sanitation is inadequate and the quality of life is poor.
Although contaminated food [7]–[11] and water [9], [12]–[15] have been identified as the major risk factors for typhoid prevalence, a range of other factors have been reported in different endemic settings such as poor sanitation [16], close contact with typhoid cases or carriers [17], level of education, larger household size, closer location to water bodies [17], [18], flooding [19], personal hygiene [12], poor life style [20], and travelling to endemic areas [21]. In addition, climatic variables such as, rainfall, vapour pressure and temperature have an important effect on the transmission and distribution of typhoid infections in human populations [12], [22].
On the Indian subcontinent, Pakistan has the highest incidence (451.7 per 100,000 persons/year) of typhoid fever followed by India (214.2 per 100,000 persons/year) [23]. The mean age of those infected with typhoid is 15.5 years in India and 7.0 years in Pakistan. Bangladesh, located in South Asia, has a population that is mostly impoverished; thus, it is probable that typhoid incidence will be high. A population-based study reported that children and young adults had the highest age-specific rates of all enteric infection [24]. Typhoid disproportionately affects children, with the highest incidence rate being observed in children <5 years old [25].Their study also indicated distinct seasonal patterns in the occurrence of typhoid. A community-based study in an urban slum in Bangladesh, by Brook et al. [26] suggested that the overall incidence was 3.9/1000 persons/year and the rate was higher in preschool children aged between 0 and 4 years (18.7 per 1000 persons/years). A recent study revealed that typhoid fever was endemic in urban areas with a high-incidence of multi-drug resistant strains [4]. This study also found that the incidence rate is higher in children aged <5 years (10.5/1000 persons/year) with an overall incidence rate of 2.0/1000 persons/year. Age-wise data demonstrated that the infection was lower in an older group (0.9/1000 persons-year) than in children [4]. Both of these studies [4], [26]) reported that the incidence of typhoid fever among populations in the urban slums of Dhaka was higher than that of paratyphoid. In a group of low income children in a semi-urban setting, the prevalence of typhoid was also found to be high in school age children [27].
Although Bangladesh is located in a region in which typhoid is highly endemic [5], very little is known about the spatial and temporal distribution of typhoid on a regional scale. Population-based studies in Dhaka have highlighted the high disease burden but did not address the spatial distribution of the population at risk from this disease; therefore, they cannot be used to make generalizations for the entire metropolitan population [20].
Timely data are essential not only to introduce vaccination programs for typhoid [23] but also to identify populations at risk so that public health interventions can be carried out effectively. A better understanding of the spatial distribution of diseases such as typhoid, together with the associated environmental factors can support more targeted disease control efforts. In addition, a spatio-temporal analysis can assist in understanding the disease dynamics in regions where the infection is already a concern.
Geographic Information Systems (GIS) are an important tool in understanding the distribution of diseases over space, and such systems have contributed significantly to spatial epidemiological research [28], [29]. GIS enables the factors associated with a disease can be investigated spatially, thereby allowing researchers to test different hypotheses. This is achieved by incorporating geocoded data from a wide range of sources such as census, surveys and remote sensing satellites [30]. As a GIS is capable of analysing disease data using a range of spatial analytical tools and spatial statistics, the output can provide valuable information in the study of public health issues enabling health officials to plan for informed decision making [29], [31]. GIS and spatial statistics have previously been applied to identify spatial clustering, risk areas and causative factors in historical studies of S. typhi in the USA [32] and India [18].
We aimed to study typhoid infection in the Dhaka Metropolitan Area (hereinafter, DMA) of Bangladesh with the following specific goals: (i) to analyse the epidemiology of typhoid cases from 2005 to 2009; (ii) to identify spatial patterns of typhoid infection based on two spatial hypotheses; and (iii) to determine the hydro-climatological factors associated with the prevalence of typhoid.
The spatial hypotheses in ii) above are: a), the infection rate of typhoid is higher among people living close to water bodies, and b) there is no significant spatial clustering of the disease in the study area.
The study area of DMA is that area contained within the Dhaka Metropolitan Development Plan (DMDP). The DMA comprises three municipalities, Dhaka City Corporation (DCC), the municipalities of Savar and Tongi, and numerous other local government areas (unions). The area is located between 23.61° and 23.97°N and 90.22° and 90.59°E, and has an area of 878 km2 (Figure 1). Based on the most recent published census (2001), the total population of this area was more than 8 million with an average literacy rate of 65% [33]. Topographically, the area is flat with a surface elevation ranging from 1 to 16 meters. The study area is surrounded by five major river systems, namely the Buriganga, Turag, Tongi, Lakhya and the Balu rivers, which flow to the south, west, north, east and northeast, respectively. These rivers are primarily fed by local rainfall but some are also distributaries of the considerably larger Ganges, Brahmaputra and Meghna rivers. The city has a humid sub-tropical monsoon climate and receives approximately 2000 mm of rainfall annually, more than 80% of which falls during the monsoon season from July to October. Most of the inhabitants in the three municipal areas have access to piped-water but outside the municipalities, drinking water sources may include ponds, rivers, tube wells etc.
Data on typhoid case occurrence for the period 2005 to 2009 was collected from the record rooms of each of the 11 major hospitals located in the study area (Figure 1) during the period from April to December 2009. The 30-member data collection team included medical graduates, information technologists and geographers. The team members were trained for two days by various professionals including medical practitioners, epidemiologists, information technologists and environmentalists. Four of the team members were responsible for the overall supervision of data collection and four were responsible for data encoding. A standardized abstraction form was created after consultation with epidemiologists, medical practitioners and environmentalists. The form was designed to document the patient's residence, demographic and clinical data, date of admission, date of discharge and outcome (alive/dead). Only those admitted to hospital with typhoid fever were included in the database, and outpatients were not included. Outpatient records in the study area are not systematic enough to enable their inclusion in this study and may include patients with other less serious febrile and diarrheic conditions. The diagnosis of typhoid was made by physicians at the respective hospitals using either blood culture or Widal tests [3]. To avoid data duplication, we first matched data using all the demographic variables and then cross-checked the data against the corresponding day/year in the log books of the hospitals. If a case satisfies both of these records, then we included in the database. We excluded cases residing outside of DMA along with the duplicates (n = 1231). This resulted in a total of 4355 cases pertaining to the study area. To minimize error in case mapping, we also cross-referenced individual case's place of residence with the 2001 census tract names by Bangladesh Bureau of Statistics (BBS). When place of residence inconsistencies were found, we used the smallest suitable mapping unit since people in the study area are more familiar with local names than with administrative units. In practice these units were the mahalla the mauza. A mahalla is the smallest urban administrative geographic unit and has a Ward Council Institution. It is a settlement of homogenous group of people of urban municipal areas. A mauza is the smallest rural geographic unit having Jurisdiction List Number (JLN). It has a definite area demarcated by the revenue department and refers exclusively to rural areas.
A spatio-temporal database of the census tracts in the study area was created from the small area atlas of BBS, digital databases from Bangladesh Space Research and Remote Sensing Organization (SPARRSO) and the Centre for Environment and Geographic Information Systems (CEGIS) database. Whilst this database was being created, it was found that 25 new census tracts used in the 2001 census, were not identified in the existing digital spatial data. To identify these, the 1991 census tracts names were first matched with the 2001 census tracts names using the community series of BBS. A hard copy map from BBS, which highlighted the road networks that were used to split the original (1991) census tracts to create new census tracts for 2001 census, was used to digitise the tracts created between decennial censuses. Field visits using a high resolution mobile mapping GPS (Trimble Nomad 800GXE) were used to confirm and correct the road network locations. The final census tract layer consists of 1212 polygons which include 441 rural villages (mauza) and 771 urban communities (mahallas). All typhoid cases were then aggregated and matched with the census tract polygon features using ArcGIS 10 software since downscaling spatial units into census districts has been shown to provide valuable information on the spatial distribution of disease [33]. Data for basic geographic features such as major waterbodies were obtained from the CEGIS.
The population data were acquired from the BBS community series [34], encoded in a spreadsheet and then linked with the census tract boundary using a unique ID. The population data between 2005 and 2009 were interpolated assuming a uniform rate of population growth during the decade in each census tract. The following equation was used to estimate the population for the years 2005 to 2009:(1)Where, Pi is population for year of interest, P2001 is the population of 2001, x is the annual growth rate between 1991 and 2001, and n is the number of years between 2001 and the year of interest.
The interpolated population data were used as the denominators to calculate the typhoid incidence rate for each year. We have not attempted to interpolate other demographic variables such as age, sex and age-wise population, hence the relevant fractions from the 2001 census were used to calculate the age-specific incidence rate for each age group reporting typhoid fever.
Daily river level data for four rivers (the Buriganga, Tongi, Turag and Balu) across the DMA were acquired from Bangladesh Water Development Board (BWDB). The daily average of the maximum water levels at the four stations was used in the analysis. In addition, climatic data including daily rainfall, temperature and humidity were obtained from Bangladesh Meteorological Department. The mean daily temperature and total weekly rainfall were calculated from this daily climate database.
The typhoid cases and demographic data were used to examine typhoid epidemiology in DMA. The annual incidence rate for each census district for each year was calculated as:(2)To investigate the monthly and seasonal variations, we have categorised annual disease data into three distinct seasons, i.e. pre-monsoon (March to June), monsoon (July–October) and post-monsoon (November to February). To determine the most vulnerable age group to typhoid, we estimated age-specific incidence rate per 100,000 persons. Geographic variation of typhoid occurrences in terms of urban and rural areas was determined by classifying our census tracts into two geographic entities (e.g. urban and rural) defined by BBS [35] (2001)1. The difference between the means for urban and rural cases was tested for statistical significance using the t-test. Temporal patterns of typhoid cases were also investigated and an epidemic curve was prepared based on the annual incidence of typhoid divided by total population for each year multiplied by 100,000 persons.
A nonparametric statistical analysis was first performed to verify the spatial association between geographic data and typhoid incidence rate. We assumed that people living closer to water bodies (e.g. rivers) had a higher rate of typhoid infection than people living further away. A surface representing distance from waterbodies was generated and the distance from the centroid of each census district from the water bodies was then measured. This resulted in a table of, typhoid incidence rate and distance from water bodies for each census district. The incidence rate was categorised into five classes: 0–5 cases; 5–10 cases; 10–15 cases; 15–20 cases and 20 + cases per 100,000 persons. Likewise, the distances were classified as: 0–1 km; 1–2 km; 2–3 km; 3–4 km and 4 km +. Finally, a contingency table was created and a nonparametric chi-square (χ2) test was performed.
In order to ascertain the results of noparametric chi-square (χ2), spatial regression in terms of Geographically Weighted Regression (GWR) was performed. In carrying out the GWR modelling, the data was subset to include only study units (mauza and mahalla) that registered at least one typhoid case over the study period. This was done to ensure that all areas participating in the analysis were coming from within the catchment areas of the 11 major hospitals from which the disease data was collected. It was assumed that areas that recorded zero incidences of typhoid infection over the five year periods never sought medical attention from these hospitals.
First of all, an Ordinary Least Square (OLS) regression was performed using incidence (log-transformed) data as dependent variable against distance to major waterbodies as independent variable. This resulted in the following statistics (R2 = 0.052, coefficient = −0.00031, p<.05), AICc = 2520.925). Following the OLS, we computed spatial weights using the Queen's case contiguity rule (first order) and used for deriving Moran's I statistics to account for spatial autocorrelation with the following outcome (I = 0.151, p<0.05), suggesting the presence of spatial dependency in the data, which explains the poor correlation result found by OLS. Next, GWR was carried out to investigate the dependency and highlight local variability. Since the spatial configuration of features being analysed were inhomogeneous [36], we used adaptive kernel to solve regression analysis. In order to understand the model fit and compare the results of global model (e.g. OLS) with local model (e.g. GWR) [37], the Akaike Information Criterion (AICc) was used to modulate the bandwidth. Local collinearity, independency and normality of residuals of GWR were further evaluated by condition number and achieving a largest condition number of 17, much smaller than a typical value of 30, revealed that our model was free from statistical concerns. The regression diagnostics from GWR is as follows (R2 = 0.475, AICc = 2386.95). The spatial regression was carried out using ArcGIS (v. 10) and the R software (SPGWR Package).
The second hypothesis of our study was that there is no association in typhoid occurrence among neighbouring spatial units. The alternative hypothesis was that neighbouring locations have similar typhoid rates: in other words, spatial clustering exists.
Since the census tracts considered in this study are highly variable in terms of shape, size and population distribution, use of raw incidence rate may not fully represent the relative magnitude of underlying risk since variation in population size between census districts means that standardised mortality/morbidity ratios (SMR) imprecisely estimated, from only a few cases, may dominate the spatial pattern. To overcome this problem, we have applied the Empirical Bayesian (EB) smoothing technique [38] to our typhoid data. This method adjusts raw rates by incorporating data from other neighbouring spatial units [39]. Essentially, the raw rates get ‘shrunk’ towards an overall mean which is an inverse function of the variance. Application of EB smoothed incidence rate not only provides better visualization compared to unsmoothed rate but also serves to find true outliers [40]. The next step was to conceptualise the spatial relationship that defines neighbourhood structure around each census district. We defined the neighbourhood structure of each census tract as being first order Queen's Case. Global Moran's I was used to initially evaluate spatial clustering. If spatial autocorrelation was found, we then evaluated the location of typhoid clusters using local indictors of spatial autocorrelation (LISA) [41]. Inference for significance of both global Moran's I and LISA was based on 999 Monte Carlo randomizations using the GeoDa software [42]. An alpha level of 0.01 was set to test the statistical significance. Based on these permutations and threshold, we have plotted values on a map to display the location of typhoid clusters in DMA.
We examined the relationship between the weekly number of typhoid cases and river levels and weather variables (temperature and rainfall) using generalised linear Poisson regression models allowing for overdispersion [43]. To account for the seasonality of typhoid counts not directly related to the river levels and weather, Fourier terms up to the 8th harmonic were introduced into the model. Fourier terms can capture repeated periodic (e.g. seasonal) patterns comprising a combination of pairs of sine and cosine terms (harmonics) of varying wavelengths [44]. This number of harmonics was chosen as that which minimised the Akaike Information Criteria. Additional indicator variables for the years of the study were incorporated into the model to allow for long-term trends and other variations between years. In this case an indicator variable for public holidays was incorporated into the model to control bias in the event that holidays affected access to hospital. To allow for the temporal autocorrelations an autoregressive term of order 1 was also incorporated into the model [45].
To account for delays in the effect of river levels and weather variables on the number of typhoid cases, temporally lagged river level and weather variables were incorporated into the model. To identify the optimum lag period, linear and spline terms for average river levels and weather variables up to each lag were incorporated into a model comprising indicator variables of years and public holidays and Fourier terms. The smallest lag with a statistically significant association (p<0.05 by Wald test) was chosen as the optimal lag. The optimal lag for river level effect was from 0 to 5 weeks (the average of the river levels in a given week and in the 5 previous weeks), and the optimal lags for rainfall and temperature effect were found to be 0 to 3 and 0 to 4 weeks, respectively.
The initial analysis was designed to identify the broad shape of any association. We fitted natural cubic splines (3 df) [46] to the average river level over lags 0–5 weeks, to the average rainfall over lags 0–3 weeks and to the average temperature over lags 0–4 weeks, as separate splines simultaneously included in the model. We included all the variables in the models to control potential mutual confounding. The model took the following form:(3)where, E(Y) is the expected weekly case count, “river level”, “temp” and “rain” indicate the average weekly river level, temperature and rainfall in each lag. NS indicates a natural cubic spline function. Fourier represents the Fourier (trigonometric) terms. i.year represents indicator variables for the year. i.holiday represents an indicator variable for weeks that include public holidays.
Because the initial analysis suggested a log-linear association through the whole range of river levels and temperature, we fitted simple linear models for river levels and temperature. A log-linear association above and below a threshold was suggested for river levels and rainfall, respectively, and so we fitted a linear threshold model for river levels and rainfall. The thresholds were chosen based on the maximum likelihood estimation for the river levels and rainfall over a grid of all possible one decimal and integer values within the range indicated on the river levels and rainfall-typhoid graphs, respectively. Likelihood profile confidence intervals (CIs) for the threshold were calculated as the thresholds for which deviance of the model was 3.84 more than the minimum. An increase in the number of cases that were associated with a 0.1 metre, 10 mm and 1°C increases in a given measure of river levels, rainfall and temperature, respectively, estimated as coefficients from the regression model, were reported as percentage change.
To investigate whether or not the main results were sensitive to the levels of control for seasonal patterns, the analyses were repeated using Fourier terms of weeks up to the 10th harmonic per year adding one harmonic at a time (0 to 10 pairs of harmonics). Indicator variables of the month were also examined instead of Fourier terms.
The annual typhoid incidence is shown in Figure 2, with the average number of typhoid cases in DMA being 871/year. Figure 3 shows the number of typhoid cases per week, and river level, temperature and rainfall data in the DMA. July–October had the highest typhoid occurrences followed by April–June (Fig. 3), and most typhoid cases were reported in monsoon season (44.62%) followed by the pre-monsoon (30.54%) and post-monsoon (24.85%) season (Figure 4).
The male-female ratio of typhoid cases was found to be 1.36, suggesting that in this population males are either more susceptible to typhoid, or more likely to present for hospital treatment, than females. The age-wise incidence rate showed a very interesting pattern because surprisingly, no cases were reported for the 5–9 years age group during our study period. The age-specific incidence rate was highest for the 0–4 years age group (277 cases), followed by the 60+ years age group (51 cases), then there were 45 cases for 15–17 years, 37 cases for 18–34 years, 34 cases for 35–39 years and 11 cases for 10–14 years per 100,000 person (Table 1). The median age of the typhoid cases was 14 years. Out of the 4,355 typhoid cases reported during the study period, 35 people died. The majority of these deaths were in patient's 35–59 years (37%) followed by 18–34 years (26%) and 60+ (14%). The highest fatality rate occurred in males (58%).
Table 2 shows the result of the χ2 test of the relationship between distance to water bodies and the number of typhoid cases per 100,000 persons. The association was found to be significant (p<0.05). Hence, the null hypothesis that there was no association between distance to water bodies and typhoid incidence rate was rejected. Since χ2 does not indicate the strength and direction of the association, Kendall's τ-B and Goodman-Kruskal γ tests were used (Table 3). Both statistics revealed a negative association between typhoid incidences and distance to water bodies, demonstrating that people living closer to water bodies had a higher infection rate than people living farther away.
The comparison of global and local regression models indicated that GWR outperformed OLS model in terms of AICc and coefficient of determination (R2). The AICc value from the OLS model for the independent variable was 2520.925. In contrast, AICc value by GWR was 2386.95. The multiple R2 (the coefficient of determination), also showed tremendous improvement. For example, OLS derived R2 was 0.052 which increased to 0.475 by GWR, demonstrating a substantial improvement in the model fit to the data. Since AICc is an effective way of comparing two models [47], a difference of three points implied an important improvement to the model fit [37].
Spatial regression analysis corroborates that spatial location was a factor in the distribution of typhoid incidence across the study area, meaning that typhoid incidence rates were associated with distance to major waterbodies (Figure 5 a&b). However, the direction of this association was two-fold with negative association being the dominant one (115 out of 197 significant observations (mahalla & mauza) showed negative relationship - 40% more than the positively associated ones, Fig. 5b). Figure 4b only shows parameter estimates (coefficients) that are significant at 90% confidence. The areas shown on the map (Fig. 5b) with significant coefficients represent a total population of 1.8 million and all together accounts for 24% of all typhoid cases (1027 out of 4355 cases), signifying the burden of disease.
The spatial pattern of the disease can be recognized easily from the smoothed and unsmoothed incidence maps (Figure 6). Geographically, the highest incidence rate was observed in the central and southern parts of DMA where the population density was relatively high. The global spatial autocorrelation analysis using the Moran's I is presented in Table 4. The result revealed that there was a significant clustering of typhoid distribution in DMA for each year. The Moran's I was highest (0.879) in 2008 (p = 0.001) and lowest (0.075) in 2009 (p = 0.0004).
Analysis of the distribution of typhoid cases during 2005–2009 in DMA showed that typhoid was not randomly distributed. A local indictor of spatial autocorrelation (LISA) map provides the most useful information in the form of significant locations of spatial autocorrelation. While high-high and low-low locations are typically known as hot and cold spots of an event, high-low and low-high are referred to as spatial outliers [39]. The LISA map of 2005–2009 (Figure 6) revealed that there were three large multi-centred typhoid clusters and five single-centred clusters. The first multi-centred cluster was located in the old town, and was centred on 13 locations, namely Babu Bazar, A.C. Roy Road, D.C. Roy Road, Dhaka Steamerghat, Islampur, Nawabbari, Jindabazar 3rd Lane, Kazi Ziauddin Road, Kumartuly, Islampur Road, Shakhari Bazar, Nalgola and Zinjira (see inset in Fig. 6). These areas are shown in the inset map in Figure 6. EB smoothed rates for these locations ranged from 96 to 2754 per 100,000 people (mean (SD): 572 (752)). The second multi-centred cluster is also located in the older part of Dhaka and comprised 7 census districts, namely Jail Road, Chak Circular Road, Purba Begum Bazar, Paschim Begum Bazar, Nazimuddin Road, Uttar Makim Katara and Ruihatta (also shown inset in Figure 6). EB-smoothed typhoid rates for these locations varied from 96 to 781 cases per 100,000 persons ((mean (SD): 246 (245)). The third multi-centred cluster is located in the northern part of DMA, also comprising 7 census tracts, including Diabari and Ashutia of Harirampur Union, Bhadam, Chak Badam, Jamaldia, Kakil Sataish and Gusulia of Tongi Municipality (see inset in Fig. 6). EB smoothed rates for these locations ranged from 283 to 1658 cases per 100,000 persons (mean (SD): 744 (559)). The single-centred locations were distributed throughout the DMA. They are Kaliganj, Gaider Tek, Chairman Bari, Bharalia and Malancha. EB-smoothed rates for these locations were 2898, 381, 904, 914 and 252 cases per 100,000 persons, respectively.
The relationships between the number of typhoid cases and river levels, rainfall and temperature, adjusted for season, inter-annual variations and holidays, are shown in Figure 7. An increase in typhoid cases was seen with increase in temperature, rainfall and river levels at lags of 0–4, 0–3 and 0–5 weeks, respectively. For a 0.1 metre increase in river levels, the number of typhoid cases increased by 4.6% (95% CI: 2.4–6.8) above the threshold of 4.0 meters (95% CI: 2.4–4.3). Maximum likelihood estimates of the threshold for rainfall calculated using a linear-threshold model coincided at 77 millimetres (95% CI: 54–90) for average rainfall over lags of 0–3 weeks. For a 1°C increase in temperature, the number of typhoid cases increased by 14.2% (95% CI: 4.4–25.0).
The chosen time lags of 3, 4 and 5 weeks for rainfall, temperature and river levels, respectively, are similar to those of previous studies [48], [49] although causal pathogens and study areas were different. The lag time from exposure to weather factors and access to hospitals may differ between causal pathogens. It may also differ between cities, countries and areas for topographical and population behaviour reasons.
In the sensitivity analyses, when the degree of seasonal control in the model was varied from 3 to 10 harmonics, the estimates of the effect of high river levels hardly changed, while the estimates decreased when no seasonal control was introduced or when the degree of seasonal control was 1 and 2 harmonics was introduced in the models (Figure S4).
To our knowledge, this is the first typhoid incidence study for this region that plots typhoid data over a five-year period together with a GIS to characterize its epidemiology and spatial patterns. Even though spatial tools have been available for quite some time, in Bangladesh, these tools have been applied mostly to diarrhoea, cholera, dengue risk, malaria and avian influenza mapping [50]–[56]. Although these studies have shown great potential in spatially identifying risk areas and associated causal factors for a range of diseases, no study has been found in the literature that applied spatial analytical tools to the geographical distribution of typhoid infections. Hence, this spatial epidemiological study of typhoid fever will be a valuable contribution to the efforts aimed at typhoid control and vaccination and prevention systems.
In common with previous findings [4], [25], we found that the age group of 0–4 had the highest incidence of typhoid. This confirms that that the rate of infection is highest in young children suggest that the current vaccination policy (which targets older age group children [25]) may need to be reassessed. Interestingly, we noticed that typhoid was prevalent in all age groups except 5–9 children group, and disproportionately affects the male population. The number of male cases surpassed female cases in the age groups of 0–4, 10–14, 15–17 and 18–34 with the highest incidence observed between the age groups of 0–4, 15–17 and 18–34. In contrast, female cases were higher in the age groups of 35–39 and 60+. Similar findings are also reported in Bangladesh [57] and in other endemic settings [16], [17], [58]. This gender preponderance might be the reflection of health-care seeking behaviour in Bangladesh which is largely controlled by cultural beliefs such as religion and patriarchy [57], [59]–[61].
Further, a conjectural explanation for apparently higher incidence rate in males in the age groups 0–17, may be that young sons are more highly valued that young daughters and are therefore more likely to be taken for hospital treatment. The reversing of the situation in later years of life may more truly reflect the exposure of women to infection during acts of cleaning and caring.
The reason for there not being a single reported case in the age group of 5–9 both for male and female is currently unknown, but may be related to vaccination being provided to school-age children. In Bangladesh, the male population is more exposed for working and other purposes than females, which may explain the higher infection rates obtained for the males in the population [58]. Other factors such as greater mobility, social behavioral attitudes as mentioned above, and lack of immunity because of lack of previous exposure may also attribute to the disproportionate number of cases in the male population [16]. Further study is therefore needed to determine the underlying risk factors, to prevent certain gender and age groups being infected by typhoid fever.
The mean and median age of the cases was found to be 17 and 14 respectively, which contrasts earlier findings in Dhaka [8] however it is similar to the results of studies from Pakistan, Indonesia and India [18], [19], [23]. The present study further confirms that there is not significant variation in the occurrences of typhoid between urban and rural environment (urban, m: 5.72 SD: 6.34; rural, mean: 5.94, SD: 12.02, p>0.05). Since previous population-based studies have mainly been conducted in urban locations in South Asia, some bias may have occurred, implying that the disease is largely confined to urban areas [16]. Urban areas in South Asia are rapidly growing compared to other parts of the world, and often characterized by inadequate provision of safe water and sanitation, hence the burden of this disease seems to be higher in urban places than its rural counterpart. This may also be introduced due to the fact that urban populations can, and do, seek medical help more often than rural populations, which could affect the number of cases that are recorded in these two locations.
A distinct seasonal variation was found with almost half (45%) of the reported cases found to have occurred in the monsoon. This is contrasting to the finding of a prospective community-based study [4] but supports other results [18], [25]. Monthly distribution revealed that August to September had the highest cases while December to February showed relatively low cases. Environmental factors such as rainfall may have substantial influence to the occurrence of typhoid [12], [22] with increasing transmission of water borne pathogens during wet periods [62]. Because of heavy rainfall during the monsoon in South Asia, a peak of disease occurrence during July to October is not surprising as chances of surface water contamination is also high [18], particularly in densely populated areas like DMA. Although the case-fatality rate was relatively low during the study period, improvements to the water and sanitation infrastructures could reduce the risk of infection and fatality, hence reducing the disease burden.
The spatial association between water bodies and the incidences of typhoid showed significant relationships. This finding suggests that people living closer to water bodies may have elevated risk of infection. This relationship has not been reported earlier, however, case-control studies in India [18] and Vietnam [17] revealed that residents close to water bodies, and who use surface water for drinking tend to have more typhoid risk. A similar observation was also reported for diarrhoea incidence [56]. The areas supporting our hypothesis of inverse relationship between typhoid occurrence and distance to waterbodies might explained by the fact that there is a higher faecal contamination load in rivers [63]. As surface and groundwater water quality get severely degraded due to increasing anthropogenic activities in DMA, this may have significant impact on the transmission and distribution of typhoid. In addition, low income inhabitants in the study area frequently use surface water for cooking, bathing etc. As a result, contamination of these water bodies may have substantial impact on the disease dynamics in the communities. As S.Typhi bacteria can survive in water for days [64], contaminated surface water such as sewage, freshwater and groundwater would act as etiological agents of typhoid [65]. Inspection of the t-value and parameter estimate maps of typhoid infection and distance to water bodies further corroborates the spatial association of these two variables (Figure 4a&b). We found that mostly communities living close to the rivers Buriganga, Turag, and Balu had an elevated risk of typhoid infection compared with communities in other locations. These three rivers have been found to have extreme pollution loads throughout the year, measured in terms of coliform counts and other physio-chemical parameters [66]–[71], hence the assumption of an increase in the disease burden is warranted. Also, risk factor investigations for typhoid have shown that all source of drinking water, including pipe water, tube wells and surface water are perpetually highly contaminated in the study area [8], [25], and therefore increases the chance of water borne infection among people living in that area. The transmission dynamics of typhoid in relation to water quality, therefore remains a very promising area for further investigation. It is important to note that we have used major water bodies to regress against dependent variable which is in coarse resolution. Using a finer resolution water bodies map may provide further detail as people in the study area depend on small waterbodies such as ponds for their domestic and bathing purposes.
The global autocorrelation analysis using the Moran's I demonstrated that the spatial distribution of typhoid was clustered for all years (2005–2009) (Table 4), signifying that the disease is not uniformly or randomly distributed over DMA. This information can guide public health professionals in their search for possible interventions. An interesting distribution pattern was observed in the typhoid incidence map (Figure 5), namely, that typhoid infections reported in the mahalla's were often located close to water bodies such as river network, lakes and ponds. One may conclude from this distribution that people closer to water bodies are more likely to be affected by typhoid fever because of huge pollution loads of surface water bodies, and the spatial regression analysis carried out in this study also supports this finding.
The LISA map (Figure 6) indicated that significant spatial clustering of census tracts with regard to typhoid endemicity in DMA. Our result suggests that empirical Bayesian-smoothed typhoid rates were spatially dependent for the years 2005–2009. This study identified 3 multi-centred and five single-centred clusters. These spatial cluster maps can be used as an initial step in the development of disease risk prediction map since neighbouring spatial units tend to share similar environments and are often connected by the spread of communicable disease [72]. Typhoid incidences in the study area have been reported to be correlated with socio-economic, environmental and sanitation factors [8], [25]. Therefore, an integrated study considering socio-economic, environmental and other relevant factors would greatly benefit public health community in deeper understanding of the dynamics and transmission of typhoid risk in DMA or elsewhere. Since rapid urbanization and food habits tend to alter the prevalence of typhoid [2], this study underscore the necessity of the implementation of sustained safe water and sanitation associated with rapid urban expansion in DMA.
The temporal analysis of the relationship between typhoid cases and hydro-meteorological factors revealed that the number of reported cases was amplified by increases in temperature, rainfall and river levels (Figure 8). While the seasonal distribution that we found in this study was similar to the distributions reported in earlier studies, one study by Lin et. al. [73] reported a contradictory finding for the association between river levels and typhoid incidences in Vietnam. Vapour pressure, temperature and precipitation have elsewhere been found to have significant associations with enteric diseases [12], [22], [74], which substantiates the result of this study. Our statistical model further stipulates that increase in rainfall and temperature lead to the higher typhoid cases in the study area. Since flooding is pervasive during the monsoon in DMA, increases in rainfall during the rainy season pollute the surface water which may have caused higher incidences of typhoid [75]. In addition, tube wells that are also flooded during the monsoon may be another source of infection due to contamination with faecal organisms [76], [77]. This study suggests that safe water supply remains a key issue in developing strategies for controlling typhoid infection in DMA.
Our study is not without limitations. First of all, the disease data that were acquired from hospitals may have underestimated or overestimated the typhoid records. Because the data were historical records and documented from the record room of each hospital, we had no valid method to ascertain repeated hospitalizations of an individual patient. In addition, hospital-based surveillance may underestimate actual infected population because only people in a severely weakened state tend to get admitted for treatment. Secondly, we only consider 11 major health service providers, the majority of which were public hospitals. The study could be improved by including data from private clinics where most of the affluent members of the population seek health services. Thirdly, we also could not separate cases into typhoid and paratyphoid groups. Isolation of these two types would allow us to estimate the disease dynamics and identify the most prevalent disease in DMA. Fourthly, the use of two or more methods to identify clustering is suggested as different analytical methods may recognize different underlying spatial patterns in the same dataset [78]. In this study, only one clustering method was used. Therefore, a future study should employ other spatial analytical technique to validate the result.
Despite the limitations above, the major strength of this study is the derivation of the first fine-scale regional map of the spatial distribution of typhoid and its epidemiology in Bangladesh.
Using multi-temporal typhoid data and spatial analytical methods, this study explored the epidemiology and spatial patterns of typhoid infection in DMA of Bangladesh. Epidemiological characteristics showed that the disease disproportionately affects the male population and certain age groups. We did not notice any significance on the occurrence of typhoid between urban and rural areas. Seasonal analysis showed that the risk of typhoid infection is high during monsoon. Temporal distribution suggested that the disease is increasing with time which underscores the importance of prevention. Cluster maps that have developed in this study would help planners to assess spatial risk for typhoid incidences in DMA or elsewhere, and to derive appropriate health policy.
The findings of this study could contribute to the understanding of spatial variability of the burden of disease at the community level and may be useful in making decisions about vaccination. Local public health officials can use the information to identify the areas having higher disease occurrences and prepare for targeted interventions. For example, children can be targeted for immunization as other measures such as improvement of water supply and sanitation require what would be a huge investment for a resource-poor country. In this study, spatial and environmental factors were used to identify possible causal factor for typhoid incidences. In addition to these factors, other variables such as population density can be used to examine the factors that are most responsible at the local level. To prevent the spread of typhoid, awareness program should be initiated for the people who rely on nearby water bodies for drinking and domestic purposes. Because of recurrent flooding in the study area in the monsoon season, infected debris could have been another source of disease transmission that would increase the risk of acquiring the disease. Therefore, typhoid prevention can be addressed through both short- and long-term measures. As a short-term measure, people should be informed through a targeted campaign program of the dangers of using unboiled surface water during the monsoon. Medium-term measures could include the improvement of drainage facilities to minimize runoff of human waste into water bodies and long-term measures may be the development of a strong surveillance system to identify both cases and carriers. Finally, an efficient vaccination program can be undertaken for age-specific population at risk, though vaccines are not an alternative to safe water and good hygiene practices [79].
|
10.1371/journal.pcbi.1005004 | Properties of Neurons in External Globus Pallidus Can Support Optimal Action Selection | The external globus pallidus (GPe) is a key nucleus within basal ganglia circuits that are thought to be involved in action selection. A class of computational models assumes that, during action selection, the basal ganglia compute for all actions available in a given context the probabilities that they should be selected. These models suggest that a network of GPe and subthalamic nucleus (STN) neurons computes the normalization term in Bayes’ equation. In order to perform such computation, the GPe needs to send feedback to the STN equal to a particular function of the activity of STN neurons. However, the complex form of this function makes it unlikely that individual GPe neurons, or even a single GPe cell type, could compute it. Here, we demonstrate how this function could be computed within a network containing two types of GABAergic GPe projection neuron, so-called ‘prototypic’ and ‘arkypallidal’ neurons, that have different response properties in vivo and distinct connections. We compare our model predictions with the experimentally-reported connectivity and input-output functions (f-I curves) of the two populations of GPe neurons. We show that, together, these dichotomous cell types fulfil the requirements necessary to compute the function needed for optimal action selection. We conclude that, by virtue of their distinct response properties and connectivities, a network of arkypallidal and prototypic GPe neurons comprises a neural substrate capable of supporting the computation of the posterior probabilities of actions.
| Choosing an appropriate action as quickly and accurately as possible in a given situation is critical for the survival of animals and humans. One of the brain regions involved in action selection is a set of subcortical nuclei known as the basal ganglia. The importance of understanding information processing in the basal ganglia is further emphasised by the fact that their disturbed interactions in Parkinson’s disease results in profound difficulties in movement. Computational models have suggested how the basal ganglia could select actions in the fastest possible way for the required accuracy level. These models further predict that a part of basal ganglia, called the external globus pallidus (GPe), needs to calculate a particular function of its inputs. This paper proposes how this function could be computed in a mathematical model of a network within GPe. Furthermore, it shows that the experimentally observed connectivity and response properties of GPe neurons fulfil the requirements necessary to support optimal action selection. This suggests the GPe neurons have properties that allow them to contribute to optimal action selection in the whole basal ganglia.
| A set of subcortical brain nuclei known as the basal ganglia are thought to be involved in action selection [1]. The external globus pallidus (GPe; sometimes referred to as simply the ‘globus pallidus’ in rodents) plays an important role within the basal ganglia, in part because it is an ‘integrative hub’ that is connected to all other nuclei in these circuits [2,3]. The function of the GPe within the basal ganglia has been conceptualized in many computational models [4–8]. A class of models [9–12] suggests that, during selection of the most appropriate action, cortico-basal ganglia circuits approximate a statistical procedure known as the Multihypothesis Sequential Probability Ratio Test (MSPRT) [13]. These models assume the basal ganglia continuously update the probabilities of different actions being appropriate given sensory signals, and that an action is initiated whenever its corresponding probability exceeds a threshold of confidence. Such a procedure for making decisions has been shown analytically to yield the fastest possible choices for a given accuracy level, when the accuracy level approaches 100% [14], and in simulations with lower accuracy, the MSPRT makes faster or equally fast choices compared to other decision algorithms [15]. For brevity, we will refer to this property as the ‘optimal action selection’.
The optimal action selection models [9–12] assume that the GPe together with the subthalamic nucleus (STN), another basal ganglia nucleus, compute the normalization term from the equation of Bayes’ theorem. This normalization ensures that the probabilities represented in the basal ganglia add up to 1 across all actions. Hence the normalization computed by the STN-GPe network mediates the competition between actions by ensuring that an action is only selected when there is high evidence for it relative to the other options (this normalization is also critical to implement the MSPRT procedure, where the actual probabilities are compared against the threshold; thus, to perform MSPRT, it is not sufficient to just know the relative probabilities, as proposed in other Bayesian models [16]). The optimal action selection models have predicted that, in order to compute the normalization, the GPe needs to send feedback to the STN that is proportional to a particular function of STN activity (we review this function in detail below). However, because of the complex form of this function, it is not clear whether GPe neurons could compute it, and thus whether the basal ganglia could approximate the optimal action selection. The goal of this paper is to refine the mapping of the Bayes’ equation on the basal ganglia anatomy by taking into account new insights into GPe cell types, and investigate whether the GPe could compute this function.
In previous models of action selection in basal ganglia, it has been widely assumed that GPe neurons are homogenous in form and function [4–6,9,11,12,17]. However, recent work shows that the GABAergic projection neurons of the GPe can be divided into two main cell types, namely arkypallidal neurons and prototypic neurons. These two cell types exhibit largely distinct firing rates and patterns in vivo, including divergent encoding of spontaneous movements [18] as well as selective temporal coupling with different phases of the oscillations present in the cortex of dopamine-intact and Parkinsonian rodents [19–21]. The physiological dichotomy is mirrored by a molecular dichotomy. Thus, arkypallidal neurons express the transcription factor forkhead box protein 2 (FoxP2), whereas prototypic neurons do not [18,19]. Conversely, most prototypic neurons express the calcium-binding protein parvalbumin (PV), whereas arkypallidal neurons do not [18–20]. Equally important, arkypallidal and prototypic neurons preferentially innervate distinct sets of basal ganglia neurons [19,20], which we review in more detail below. In summary, there is now compelling support for the idea that a dichotomous functional organization, as actioned by arkypallidal and prototypic neurons with specialized physiological, molecular, and structural properties, is fundamental to the operations of the GPe. In this paper, we extend this notion by examining how a GPe network composed of these two distinct types of neuron could compute the function required for optimal action selection.
In the next section, we review the optimal action selection model and its predictions concerning computations in GPe. In the subsequent section, we show that the observed connectivity of arkypallidal and prototypic neurons, as well as the relationships between firing rate and injected current (f-I curves) for the two populations of the GPe neurons, fulfil the requirements necessary to approximate optimal action selection. Finally, we discuss our results and consider future directions.
We first review the overall computation in the model, and then its implementation in the cortico-basal-ganglia circuit.
Let us first introduce a simple choice task, in the context of which we present the model. Consider a rat that has to press either a lever to the left or to the right on the basis of an auditory stimulus. On each trial, pressing only one of the levers will lead to the reward. The auditory stimulus consists of a sequence of short intervals during which a low- or high-pitched tone is presented, which provide probabilistic information on which lever is correct on a given trial. During trials in which pressing the left lever is rewarded, the low tone has 70% chance of occurring in each interval, while the high tone has only 30% probability of occurring. Conversely, on trials when pressing the right lever is rewarded, the low and high tones have 30% and 70% probabilities, respectively. Let us assume that the rat is well trained in this task. Please note that, in this hypothetical task, in order to maximize its reward, the rat needs to listen to the stimulus, accumulate information from successive beeps, and then only makes a choice (i.e. selects an action) once it reaches a certain level of confidence.
Let us denote different actions available in a given context by Ak, thus in the above example, the rat has two potentially rewarded actions, A1 and A2, corresponding to pressing the left and the right lever, respectively. The model suggests that, during action selection, the cortico-basal-ganglia circuit is evaluating the probabilities of alternative actions being appropriate in a given context, which we denote by P(Ak). Whenever any of the probabilities exceeds a threshold of confidence during the internalized process of action selection, the corresponding action is triggered.
In the model, the probabilities of actions P(Ak) are updated on the basis of sensory input. For simplicity, let us assume that the time during the action selection process is divided into discrete intervals, and during each interval a sensory input S in presented. The sensory input S could be used to update the probabilities of action P(Ak), because from past experience the animal could have learned how often S appeared on trials on which action Ak was rewarded. Let us denote this rate of occurrence by P(S|Ak). Thus, for example, in the task described above, if the low tone is presented at the current time step, then P(S|A1) = 0.7 and P(S|A2) = 0.3. Bayes’ theorem (see Eq 1) describes how to update the probabilities of actions according to the sensory input:
P(Ak|S)=P(Ak)P(S|Ak)P(S)
(1)
Bayes’ theorem says that in order to compute the updated or ‘posterior’ probability of action P(Ak|S), one needs to multiply the previous or prior probability P(Ak) by the learned probability of the sensory input S appearing on trials on which action Ak was correct, i.e. P(S|Ak). For example, when the low tone is presented, then P(A1) is multiplied by 0.7 and P(A2) is multiplied by 0.3. Additionally, to ensure that the posterior probabilities add up to 1, these products are divided by a normalization term P(S) equal to the sum of the products across all actions:
P(S)=∑k=1NP(Ak)P(S|Ak)
(2)
In Eq 2, N denotes the number of available actions. If for any action the posterior probability computed from Eq 1 exceeds a threshold of confidence, the corresponding action is chosen. Otherwise the integration of information continues and the posterior probability P(Ak|S) from the current time interval becomes the prior P(Ak) for the next one.
Eq 1 includes multiplication and division, which are not natural operations for neurons (as classical neural networks models rather assume that neurons add their inputs and potentially transform them through non-linear functions e.g. [22]), but this problem can be solved by taking the logarithm. Recall that the logarithm has the following properties: log a·b = log a + log b, and log a/b = log a–log b. Hence taking the logarithm of both sides of Eq 1 we get:
logP(Ak|S)=logP(Ak)+logP(S|Ak)−logP(S)
(3)
Thus, if in the context of neurons, they have firing rates proportional to the logarithms of probabilities, the update according to Bayes’ theorem can be performed just using addition and subtraction. The computation of the logarithm of the normalization term becomes only slightly more complex, as it needs to include nonlinear transformations:
logP(S)=log∑k=1Nexp(logP(Ak)+logP(S|Ak))
(4)
Fig 1 illustrates how Eq 3 could be mapped onto a subset of cortico-basal-ganglia-thalamic circuits [10]. The model assumes that within the circuit there exist populations of neurons selective for different actions (shown in different colours in Fig 1). The notion that different actions could be subserved by discrete neuronal populations within cortico-basal-ganglia-thalamic circuits is supported by anatomical data demonstrating that these circuits are composed of partially segregated (and topographically organized) ‘loops’ [23]. It has been demonstrated that, for certain assumptions, log P(S|Ak) is proportional to the activity of the sensory neurons selective for stimuli associated with action Ak being correct [9,24–26], so the term log P(S|Ak) could be encoded directly in the activity of cortical sensory neurons. In the model framework in Fig 1, the neurons in the frontal cortex add the input from sensory neurons to the logarithm of the prior probability, which is provided by a feedback from the thalamus, thus they perform the addition in Eq 3.
The logarithm of the normalization term is computed in the model in a circuit of reciprocally connected STN and GPe neurons, and this computation is described in detail in the next subsection. The output nuclei of the basal ganglia receive excitation from STN (which, in the model, is proportional to log P(S)) and inhibition from the cortex via the striatum (which, in the model, is proportional to log P(Ak) + log P(S|Ak)), and subtract these two inputs; thus, according to Eq 3, their activity is proportional to –log P(Ak|S). The output nuclei send inhibitory signals to the thalamus, so the activity in the thalamus is proportional to the logarithm of the posterior probability, i.e. log P(Ak|S). Finally, the logarithm of the posterior probability is sent back from the thalamus to the frontal cortex as it becomes the basis of the computation (or prior log P(Ak)) in the next time step.
The model described so far assumes that certain neural populations have activity proportional to the logarithms of probabilities, but these quantities are negative (as probabilities are smaller than one). This problem can be solved by assuming that the firing rates are proportional to the logarithms of probabilities increased by a constant c (we discuss the required value of this constant in the Results section). Eqs 5–9 below describe computations performed by each of the nuclei in the model:
SENk=logP(S|Ak)+c
(5)
CTXk={logP(Ak)+c+SENkat the first intervalTHk(t−1)+SENkat subsequent intervals
(6)
STN=log∑k=1NexpCTXk
(7)
OUTk=–CTXk+STN
(8)
THk=c–OUTk
(9)
In the above equations SENk, CTXk, OUTk and THk respectively denote the firing rates of populations of sensory cortical neurons, frontal cortical neurons, basal ganglia output nuclei neurons, and thalamic neurons, selective for alternative k. At the start of the trial, the frontal cortical neurons are initialized to the logarithms of initial prior probabilities of actions, and subsequently they receive feedback equal to thalamic activity in the previous time step i.e. THk(t–1). The STN term denotes the sum of activities across all STN neurons, while we denote the activity of STN the neurons selective for action Ak as STNk, i.e.
Thus, according to Eq 8, each neural population in the output nuclei in the model receives input from all populations in the STN, in agreement with experimental data suggesting that the axonal projections of STN neurons are relatively diffuse [27].
We will describe in the next subsection how the activity described by Eq 7 could arise in the STN, but first let us show that the model described in Eqs 5–9 correctly updates probabilities. At the first time-step, the activity of frontal cortical neurons is equal to (according to Eqs 5 and 6):
CTXk=logP(Ak)+logP(S|Ak)+2c
(11)
According to Eqs 7, 11 and 2, the activity in the STN is:
STN=log∑k=1Nexp(logP(Ak)+logP(S|Ak)+2c)==log∑k=1NP(Ak)P(S|Ak)exp(2c)==logP(S)+2c
(12)
Constants 2c then cancel while computing the activity in the output nuclei (using Eqs 8, 11, 12 and 3):
OUTk=−logP(Ak)−logP(S|Ak)−2c+logP(S)+2c=−logP(Ak|S)
(13)
We have shown that at the end of the first time interval, the model computes the posterior probabilities of actions. Since the posterior probabilities are then fed back to frontal cortical neurons as a prior for the next interval, it can be shown using analogous calculations that the network correctly computes the posterior probability in every subsequent interval.
We now describe the conditions under which the activity in STN is proportional to the logarithm of the normalization term. Bogacz and Gurney [9] have shown that the STN-GPe circuit with the architecture shown in Fig 1 would produce activity of STN that is given in Eq 7, if and when the neural populations in STN and GPe had the following relationships between their inputs and their firing rates:
STNk=exp(CTXk−GPk)
(14)
GPk=STN−logSTN
(15)
In the above equations, GPk denotes the firing rates of GPe neurons selective for action Ak (in the original model [9] the GPe neurons were assumed to belong to a single cell type). The STN neurons receive excitation from cortex and inhibition from GPe, so their total input is CTXk−GPk, thus Eq 14 implies that the STN neurons in the model have exponential input-output relationships (often termed ‘f-I curves’ in empirical studies). The GPe neurons receive input from STN, but this input is coming in Fig 1 from STN neurons selective for all actions which we denote by STN without a subscript (see Eq 10). Eq 15 implies that the GPe provides inhibition proportional to STN–log STN.
Before giving a mathematical proof for how the STN-GPe circuit computes Eq 7 in the model, let us first provide an intuition. Starting from the right end of Eq 7, the frontal cortical activity CTXk is provided to the STN in the model by the cortico-subthalamic pathway (see Fig 1). The exponentiation is performed by the STN neurons (cf. Eq 14). The summation is achieved due to the diffuse projections from the STN: in the model each neural population in the output nuclei receives input from all populations in the STN, hence the neurons in the output nuclei can sum the activity of STN populations. The only non-intuitive element of the computation of Eq 7 is the logarithm–this comes from the interactions between STN and GPe, as shown below.
We now present a sequence of simple mathematical operations that show that Eqs 14 and 15 imply Eq 7. Substituting Eq 15 into 14 gives:
STNk=exp(CTXk−STN+logSTN)
(16)
Using the property of exponentiation ea+b = eaeb we obtain:
STNk=exp(CTXk)exp(−STN)STN
(17)
Summing over k and using the definition of STN (given in Eq 10) we obtain:
STN=∑k=1Nexp(CTXk)exp(−STN)STN
(18)
Taking the logarithm of Eq 18 we get:
logSTN=log∑k=1NexpCTXk−STN+logSTN
(19)
log STN cancel on both sides in Eq 19, and moving STN to the left side we see that the sum of activities of all STN neural populations is equal to the required value of Eq 7:
STN=log∑k=1NexpCTXk
(20)
Eqs 14 and 15 thus describe the predictions of this model on the response properties of STN and GPe neurons, respectively. Bogacz and Gurney [9] have shown that the published f-I curves of STN neurons [28,29] indeed follow the exponential function precisely up to the firing rate of 135 spikes per second (STN neurons are unlikely to fire at higher rates in vivo). In the next Section, we investigate whether the properties of GPe neurons match those required to compute Eq 15.
We start by analysing in detail the computations in GPe required by the optimal action selection model, and by giving further insight into our initial hypotheses on how this computation could be performed [9]. We then show how the key aspects of our hypothesis match experimental data. Using simulations of an optimised computer model embedding realistic f-I curves of neurons for each sub-population, we demonstrate that GPe can perform computations required for optimal action selection. We finally report the dynamical properties of such network for varying cortical inputs, and discuss how the connections between GPe and striatum can be incorporated into the model.
The model described in the previous section predicts that the GPe neurons send an inhibitory signal to the STN that is proportional to STN–log STN. The black curve in Fig 2A illustrates the shape of this function. It is worth clarifying that the axes in Fig 2A are expressed in the units of log of probability [25], which are related to the units of firing rate and input current through scaling factors (discussed later) because the model assumes that the firing rates are proportional to the probabilistic quantities they represent. As shown in Fig 2A, the function STN–log STN diverges to infinity for very low STN input. However, such very low values of STN input are not biologically relevant because STN neurons are autonomously active [30,31]; GPe will thus receive input from STN even when STN itself does not receive any organised excitatory input. The lower bound on STN is provided by Eq 20 –please note that CTXk cannot be negative thus STN ≥ log N. The lowest possible number of choice alternatives is 2, thus STN ≥ log 2 ≈ 0.7.
The upper bound on STN is provided by Eq 12; please note that log P(S) cannot be positive thus STN ≤ 2c. Because the upper bound depends on constant c we now consider its value. Recall that constant c was added in the model to the activity of the neurons representing logarithms of probabilities to ensure their firing rates are not negative. Nevertheless for any value of c, for sufficiently low probability p, the value of log p + c will be negative, so such low probabilities will not be represented in the model. Thus the value of constant c determines the lowest probability of action that can be represented in the model. It has been demonstrated that humans can represent prior probabilities of actions as low as 0.05 [32]. If we wish the model to represent the probability of 0.05, then c needs to be around 3, as log 0.05 + 3 ≈ 0. If we set c to 3, then the upper bound on STN considered above becomes STN ≤ 2c = 6. In summary, we will consider the relevant range of STN for which GPe needs to compute STN–log STN to be from around 0.7 to 6. The black curve in Fig 2A is solid in this range, while it is dashed outside it.
In the relevant range STN ϵ (0.7, 6), function STN–log STN is non-monotonic, i.e. it initially decreases (for STN < 1) and then increases (for STN > 1), so it is very unlikely that neurons with such an f-I curve would exist, and thus that this function could be computed by single neurons. Even ignoring the non-monotonicity, which occurs on only a small part of the relevant range, the function STN–log STN is convex on its whole range, i.e. the larger the input, the larger is its rate of growth. By contrast, the previously published f-I curves of GPe neurons were only convex in a very narrow range of small input currents, while on the majority of their range they were linear or concave, i.e. they were decreasing their rate of growth for larger inputs [33–35]. Hence, the published data suggest that it is unlikely that individual GPe neurons, or a single type of GPe neuron, could compute even the monotonic part of function STN–log STN. Theoretically, however, this function could be computed in a microcircuit composed of two populations of GPe neurons with distinct activities and connections. Fig 2A shows how the function STN–log STN could be represented by a difference of two functions: a 'Prototype' (P) function, shown by a blue line, and an 'Adjustment' (A) function, shown by a red striped area. Therefore, the function STN–log STN could be computed by two neural populations, ‘P’ and ‘A’, with the connectivity shown in Fig 2B, and with f-I curves corresponding, respectively, to the blue line and the height of red striped area (i.e. the difference between blue and black lines) in Fig 2A (such f-I curves are monotonic and non-convex). In this architecture, both populations P and A receive input from STN. Population A transforms this input via its non-linear f-I curve and sends inhibition to neurons P thus adjusting their response. In this architecture only the neurons in population P have activity proportional to STN–log STN, so only they project back to STN.
GABAergic projection neurons in the GPe can be divided into two major cell types, termed prototypic and arkypallidal, on the basis of their distinct firing in vivo, molecular profiles and structure [18–20]. These sub-populations exhibit clearly distinct connectivity within the STN-GPe network (Fig 2C), as derived from empirical studies in rodents [19,20] and previous computational modelling of empirical data [36]. Interestingly, this pattern of connectivity resembles that of the model computing function STN–log STN (cf. Fig 2B). In particular, only one of the GPe populations, i.e. the prototypic neurons (but not arkypallidal neurons) send projections back to STN [19,20]. Additionally, recent computational modelling of effective connectivity in the STN-GPe network suggest that both prototypic and arkypallidal neurons receive input from STN [36]. Thus, in summary, the observed pattern of GPe connectivity (Fig 2C) includes all the connections present in the model computing function STN–log STN (Fig 2B), but it also includes two additional connections (i.e. that between prototypic neurons and that from prototypic to arkypallidal neurons) and we will address their roles below.
We investigated whether the f-I curves of GPe neurons have characteristics required for computation of function STN–log STN. The first characteristic we expected was that f-I curves of GPe neurons should have linear or logarithmic shape, allowing them to jointly compute the function STN—log STN. Then we sought to explore if the populations 'P' and 'A' introduced in our theoretical model could correspond to the prototypic and arkypallidal neurons reported experimentally. Such a correspondence would require further two characteristics in their f-I curves. For low levels of input from STN, the average firing rate of prototypic neurons should be higher than that of arkypallidal neurons. This requirement arises because the “adjustment” to the firing rate of the 'P' sub-population (red area in Fig 2A) is only necessary for high inputs, so one could expect arkypallidal neurons to have a firing rate closer to zero for low STN. Furthermore, one would expect that the responses of prototypic neurons to be more linear than those of arkypallidal neurons (as the two populations in Fig 2B compute the linear and logarithmic functions in Fig 2A).
The f-I curves of molecularly-identified prototypic and arkypallidal neurons have been recently measured experimentally by Abdi et al. [19] using perforated patch-clamp recordings in rat brain slices. Experimental procedures in that study were conducted either in Oxford in accordance with the Animals (Scientific Procedures) Act, 1986 (United Kingdom), or in Bordeaux according to institutional guidelines and the European Communities Council Directive 86/609/EEC and its successor 2010/63/EU. During these recordings, the slices were perfused continuously with oxygenated artificial cerebrospinal fluid at 35°C–37°C. Firing rates were recorded as a function of injected current for 18 prototypic neurons (expressed PV but not FoxP2) and 18 arkypallidal neurons (expressed FoxP2 but not PV) (Fig 3A). In these measurements, the depolarising current injection was gradually increased in magnitude until a point at which the neuron was unable to follow (by firing well-defined action potentials). To avoid excluding any neurons from the analysis, the average f-I curve for a given type of neuron was computed for the range in which all the studied neurons of this type were able to respond. All prototypic and arkypallidal neurons responded to currents up to 150 pA and 225 pA respectively, and hence the average f-I curves were computed up to these values. For each value of current, the firing rate was measured over 2 s interval during which the current was injected (Fig 3C). Additionally, the rate of autonomous firing was measured (Fig 3B), that is, the firing present with 0 current injection in the presence of receptor antagonists [19].
The average f-I curves of all prototypic and all arkypallidal neurons are shown in Fig 3D, and the f-I curves of individual neurons are shown in Fig 4B (the data is provided in S1 Table). One characteristic we expected was that, for low values of excitatory input (nominally from STN) the activity of prototypic neurons should be higher than that of the arkypallidal neurons. This expectation agrees with the experimental data (see Fig 3D), as the average rate of the autonomous firing of prototypic neurons (18.0 spikes/s) was higher than that of arkypallidal neurons (5.1 spikes/s), and this difference was highly significant (p = 0.0001, unpaired t-test).
To test the linear or logarithmic behaviour of GPe neurons, we compared the fitting quality of different functions to the f-I curves of each neuron. We obtained the best fitting parameters a, b and c for different functions (linear f = a + bI, logarithmic f = a + blogI, a combination of both functions f = a + bI + clogI, or a power function f = a(I + b)c) by minimizing the root mean squared error (RMSE) between the actual and the predicted firing rates over the range of positive input currents for each neuron. To help visually asses the shape of f-I curves in Fig 4B, we ordered the neurons according to a bias for linear fit, defined as a ratio of RMSE between linear and logarithmic fits. We found that all f-I curves can be well explained by a combination of linear and logarithmic functions (shown as solid curves in Fig 4B; average r2 for arkypallidal and prototypic neurons were 0.998 and 0.994 with standard deviations 0.0041 and 0.013). Furthermore, the f-I curves ranged from almost fully linear to clearly logarithmic (i.e. very low values of linear coefficient b in function f = a + bI + clogI). The power function could describe the f-I curves equally well (Fig 4A) as both a power function and a combination of linear and logarithmic functions can take a similar, concave shape. To verify whether the diversity of the shapes of f-I curves is not an artefact stemming from differences in the recording quality of individual neurons, we computed the correlation between the bias for linear fit and the series resistances during perforated-patch recordings, which was not significant.
Contrary to our expectations, we did not find statistically significant differences in the linearity of the f-I curves of prototypic and arkypallidal neurons (the linearity was quantified as the ratio of RMSE between linear and logarithmic fits). Indeed, both populations included neurons that lied within a continuum ranging from highly linear to highly logarithmic (Fig 4B). Even though this third expected characteristic was not present, it is striking that the f-I curves of the GPe neurons range from the linear to the logarithmic, which are the two components of Eq 15 that describe the predicted computation in GPe.
Although not all prototypic neurons had linear response curves, their reciprocal connections may contribute to adjusting the shape of their response profiles (these connections have been experimentally observed, but were not initially included in our theoretical network shown in Fig 2B). We now show that the mutual inhibitory connections within the population of prototypic neurons linearize their response profile. The intuition for this effect is provided in Fig 5A, which shows a hypothetical concave (or saturating) f-I curve that qualitatively resembles the average empirical response of prototypic neurons to depolarizing current injections (Fig 3D). Let us consider two cases of excitatory input currents. A small input current I1 without the mutual inhibition would produce firing f1. However, with mutual inhibition, the overall input is reduced proportionally to f1 to a smaller value I1s, which gives a lower firing f1s. Analogously, for a higher input current I2 the mutual inhibition reduces the firing from f2 to f2s. Please note that the reduction in firing due to mutual inhibition in the case of small input, i.e. f1 –f1s, is more pronounced than for the case of large input, i.e. f2–f2s. Since the reduction is more pronounced in the less linear part of the curve than the more linear one, the mutual inhibition linearizes the response profile.
To quantify the effect of mutual inhibition on response properties of prototypic neurons, we modelled the responses of a population of prototypic neurons receiving an external excitatory input I and inhibiting the neurons within the population (Fig 5B). Throughout the paper, we denote the strength of connections between population X and Y by wXY, where X and Y can be P for prototypic, A for arkypallidal or S for STN. Thus, the dynamics of this population of prototypic neurons with mutual inhibition are described by the following equation:
PR˙O=fP(I−wPPPRO)−PRO
(21)
In the above equation, function fP(x) is based on the average f-I curve of prototypic neurons (see Fig 3D), and was defined in the following way. If x was equal to a current used in the in vitro experiment for all prototypic neurons, fP(x) was simply equal to the average firing rate for this current. If x was between two currents tested in experiment, fP(x) was found using linear interpolation. If x was below the lowest current tested in experiment, fP(x) was set to 0. Finally, if x was above the range on which the average f-I curves were computed (see above), fP(x) was set to the firing at the maximum current used with all prototypic neurons, but in all analyses below, we ensure that we do not present results relying on this value, as in this case the f-I curve is undetermined.
The last term (–PRO) in Eq 21 is a decay term. The value to which the variable PRO converges can be found by setting the left hand side of Eq 21 to 0, because at convergence, the value of PRO does not change. Thus we find that at convergence PRO = fP(I − wPP PRO). So in this model, when wPP = 0, the activity PRO is equal to the activity determined by the f-I curve for given input I, while when we increase wPP, we can investigate the effect of mutual inhibition on the response profile.
The activity of the prototypic neurons at convergence as a function of external input is shown in Fig 5C for different values of mutual inhibition. As the strength of the connection between prototypic neurons (wPP) increases, their response becomes more linear. To quantify it, we computed squared correlation (R2) between I and PRO for inputs I>0. The R2 was increasing with wPP, namely it was equal to 0.988, 0.994, 0.996 and 0.997 for wPP equal to 0, 1, 2 and 3 respectively.
To test our main prediction, namely that a network of GPe neurons is capable of sending inhibition to STN to instantiate the function STN–log STN, we next investigated the behaviour of a realistic, multi-dimensional network including all the different neurons recorded in the arkypallidal and prototypic populations. We built a computational network embedding the 18 arkypallidal and 18 prototypic neurons with f-I curves based on experimental data, and all corresponding connections (Fig 6A). This included 36 excitatory connections from STN to each neuron in these two GPe sub-populations (wiSA and wiSP), and 18 inhibitory connections from each prototypic neuron back to STN (wiPS; these connections influence the cost function that will be introduced later). The model also included 182 inhibitory connections from arkypallidal to prototypic neurons, 182 inhibitory connections from prototypic to arkypallidal neurons, and 182 inhibitory connections between prototypic neurons, but to reduce the high dimensionality of this system, we assumed similar levels of connectivity, and constrained the weights in each group to a single value (wAP, wPA and wPP). Thus the dynamics of the simulated neurons was described by:
In the above equations fAi and fPi denote the f-I curves based on experimental data for individual neurons (computed analogously to fP described above). In addition, to prevent bias during the optimisation process, the STN input sent to GPe was assumed to stem from 18 identical neurons, which ensured a homogeneous number of neurons in all three populations, and normalised the weights between them to comparable ranges. For a given value of STN, these equations were solved until convergence.
In Eqs 22 and 23 the activity of STN is scaled by a constant αS to translate it to appropriate units. As mentioned before, the variables of the model are expressed in the units of the logarithm of probability that are proportional to firing rates, and we denote the proportionality constant for STN by αS. Thus, the firing rate of STN neurons in [spikes/s], is αSSTN. The value of αS can be estimated on the basis of published f-I curves of STN neurons. In particular, for zero input, STN = exp(0) = 1, so αS is equal to the autonomous firing rate of STN neurons. Hence we set αS = 17 on the basis of average autonomous firing rate of 7 STN neurons for which the f-I curves have been published [28,29] and have been fitted with exponential functions [9].
We found the weights for the 57 connection parameter through constrained nonlinear optimisation, which minimised the difference between the overall inhibition sent back to STN from GPe and the desired function STN–log STN in the input range considered [0.7–6] (a.u.).
Similarly as above, the feedback from prototypic neurons is divided by αS in Eq 24 to bring it back from [spikes/s] to the units of the model. We note that the precise value of αS does not affect the ability to find weights of connections, because it just scales other free parameters (wiSA, wiSP, wiPS in Eqs 22, 23 and 24, respectively), so changing αS would only result in rescaling these parameters. We only include αS to give the weight parameters the same scale, so their values could be compared.
We completed 100 runs of this optimisation process for weights initialised at random values ranging between 0 and 0.3. For all runs, the optimiser converged to set of weights that exhibited a behaviour following our desired output function (average Cost was 0.0214 with standard deviation 0.017). For each run, however, the parameters found showed clear differences, indicating that distinct set of weights are capable of producing similarly optimal output functions (i.e. close approximation of STN–log STN) in such a high dimensional system. Two examples of synaptic weights and resulting outputs from the network are shown in Fig 6B.
Regardless of differences between individual runs, we sought to further characterise whether any structure underlies the weights found through global optimisation. For example, one could expect the connections from STN to arkypallidal neurons to be stronger for neurons with logarithmic f-I curves, in accordance with our initial suggestion on how the function STN–log STN is computed (Fig 2B). The means and standard deviations of parameters found in all 100 optimisation runs is shown in Fig 6B (bottom panel), where the parameters in each group are sorted by how linear the f-I curve of each neuron is. Thus the preferential STN projections to logarithmic arkypallidal neurons would manifest in the figure in higher weights on the right of the first grey area. Fig 6B did not highlight any specific patterns, neither for linear nor for logarithmic neurons in any of the two populations of GPe neurons. This indicates that the network considered is so flexible that it contributes to optimal action selection without requiring clear patterns in the way linear and logarithmic neurons are interconnected.
Among the estimated weights, the lowest value was on average assigned to a connection from prototypic to arkypallidal neurons (Fig 6B). The low weight value was obtained probably because this connection was not necessary to compute function STN–log STN (it was not included in the hypothetical network of Fig 2B).
Finally, we sought to evaluate the dynamics of the entire STN-GPe network in response to varying cortical input. We evaluated the transient behaviour of the network (weights were set to one solution found during the optimization) when considering realistic time constants for firing rates (τ) and realistic transmission delays (Δt), as used in a previous model of this circuit [36]. In particular, we used the following time constant τS = 10 ms [37,38], τA = τP = 15 ms [39] and transmission delays ΔtSA = ΔtSP = 2.8 ms, ΔtPS = 1.3 ms and ΔtPA = ΔtPP = ΔtAP = 1 ms [37,39]. The dynamics of the model are described by:
τSST˙N=exp(CTX−∑i=118wiPSPROi(t−ΔtPS)/αS)−STN
(25)
τPPR˙Oi=fPi(18⋅wiSPαSSTN(t−ΔtSP)−∑k=118wAPARKk(t−ΔtAP)−∑j=118wPPPROj(t−ΔtPP))−PROi
(26)
τAAR˙Ki=fAi(18⋅wiSAαSSTN(t−ΔtSA)−∑j=118wPAPROj(t−ΔtPA))−ARKi
(27)
We run multiple simulations and studied the behaviour of the system for varying cortical input values in the range CTX ϵ [0.7–6]. Since the above network includes only a single group of STN neurons (selective for a single action), the desired value to which STN should converge according to Eq 7 for N = 1 is simply STN = CTX. We found that the network properly converged to the expected value after a transient time period (the curves in the second display from the top in Fig 7A converge to the values in the top display), except for a very high cortical input (CTX = 6). For such high input the network started oscillating and did not converge. Oscillations are known to emerge in networks of mutually connected excitatory and inhibitory neural populations like the STN-GPe network under certain conditions [40,41]. One of necessary conditions is a strong reaction of neurons in one population to changes in activity of the other [40,41], and the oscillations emerged in Fig 7A for the high cortical input, as then the STN neurons were operating in steeper range of their exponential f-I curve, and thus were more reactive to their inputs from GPe.
In contrast, the same network always converged to the appropriate values in the absence of delays, as the delays are necessary for oscillations in the model of STN-GPe network [40,41]. We further analysed the influence of delays in this behaviour (Fig 7B). We compared the output of the network for delays values as reported in literature, without delays, and for different values in between, when cortical input values followed a step function (steps of 2 cortical units) starting at different input values (2, 3, 4 and 5). We found that the level of cortical input for which the system became unstable progressively decreased as delays increased.
Beyond the simulated demonstration of how a GPe network, comprised of interconnected prototypic and arkypallidal neurons with experimentally observed f-I curves, is able to compute the function required for optimal action selection, it is also useful to show it analytically, as it will allow us to see in the next subsection how the connections between GPe and striatum can be incorporated into the model.
In order to demonstrate the computation analytically we consider a simplified model of the GPe circuit. We assume that the arkypallidal neurons have f-I curves described by a combination of linear and logarithmic terms (as Fig 4B shows that this function well describes experimentally observed curves):
fA(I)=aA+bAI+cAlogI
(28)
For simplicity, we will assume the response of prototypic neurons could be described as a linear function of their input, as we demonstrated that the mutual inhibitory connections among the prototypic neurons linearize their response profile (Fig 5).
We consider a network with the simplified connectivity shown in Fig 2B. Thus we do not consider connections among the prototypic neurons as their role has already been incorporated in assuming linear fP, and we do not consider connections from prototypic to arkypallidal neurons, as Fig 6B shows that this connection had a relatively low weight when the GPe network computed STN–log STN. The activity of prototypic neurons becomes:
PRO=aP+bP(wSPSTN−wAPARK)=aP+bP(wSPSTN−wAPaA−wAPbAwSASTN−wAPcAlogwSASTN)
(30)
Rearranging the term we obtain:
PRO=(bPwSP−bPbAwAPwSA)STN−bPcAwAPlogwSASTN+aP−bPaAwAP
(31)
It is easy to see that the output of prototypic neurons wPSPRO will become equal to STN–log STN, when the parameters satisfy the following constraints:
{bPwSPwPS−bPbAwSAwAPwPS=1bPcAwAPwPS=1aP=bPaAwAPwSA=1
(32)
This analysis shows that even if the f-I curves in GPe have a generalized form described by Eqs 28 and 29, by setting the weights to the values satisfying the above conditions, the GPe can compute the function STN–log STN.
The model presented so far describes the computations only in a subset of nuclei in the larger cortico-basal-ganglia network that, when embodied in Fig 1, does not include the connections between GPe and striatum. However, there are prominent projections from the striatum to the GPe [2,3], and arkypallidal neurons densely innervate the striatum [20]. We now discuss how these bidirectional connections can be included in the model.
The striatal projection neurons can be divided into those expressing type 1 dopamine receptors or type 2 dopamine receptors (D1 and D2), and they are thought to be involved in initiation and inhibition of actions, respectively. Let us first discuss the D2 striatal neurons for which GPe is the main target of their projections [3]. The connection from the D2 neurons to GPe is a critical part of the so-called indirect pathway (Fig 8A) involved in inhibiting the actions that in the past resulted in negative feedback [4]. Although there are no anatomical data that definitively describe to which GPe cell type(s) the D2 striatal neurons project, Fig 8A includes connections from D2 neurons to prototypic neurons, as only such connections would allow D2 striatal neurons to inhibit actions (if the D2 neurons projected to arkypallidal neurons, their activity would facilitate rather than inhibit action selection, because the pathway D2-ARK-PRO-STN-OUT-TH would include 4 inhibitory connections, so it would be effectively excitatory). The striatal D2 neurons are not included in the optimal action selection model (see Fig 1), because the MSPRT framework assumes that one of the actions is correct and should be taken on a given trial. By contrast, in certain situations, any action may lead to negative consequences, and the D2 striatal neurons are likely to ensure that no action is then taken. To include the D2 striatal neurons and their projections to GPe, the model has to be extended beyond the MSPRT framework, which is the subject of ongoing work.
It has been also reported that GPe receives input from most of the striatal neurons that target the output nuclei of the basal ganglia [42], i.e. the D1 striatal neurons. These striatal neurons are included in the model shown in Fig 1, but their projection to GPe is not. This pathway may form an alternative route by which the evidence for actions represented in cortex and striatum can be provided to STN. Please note that if the D1 striatal neurons project to prototypic neurons, the pathway CTX-D1-PRO-STN involves 2 inhibitory connections so it is effectively excitatory, similarly as the hyperdirect pathway CTX-STN (Fig 8B). Note that this would not be the case if striatal D1 neurons projected to arkypallidal neurons, as connection D1-ARK-PRO-STN would be effectively inhibitory, thus for simplicity we only consider the connections from striatal D1 neurons to prototypic GPe neurons.
We now show that adding the connection from striatal D1 neurons to prototypic GPe neurons, while also appropriately reducing the connection from cortex to STN, would not change the activity in the STN in the model, and hence the model would still implement optimal action selection.
Let us consider a modified version of the model in which we allow arbitrary weights of connections from cortex to STN, which we denote by wCS, and from striatal D1 neurons to prototypic neurons, which we denote by wXP. Since we assume that the activity of striatal D1 neurons reflects the cortical inputs, the activity of STN and GPe populations in this model becomes:
STNk=exp(wCSCTXk−wPSPROk)
(33)
PROk=fP(wSPSTN−wAPARKk−wXPCTXk)
(34)
ARKk=fA(wSASTN)
(35)
The computation in this modified model and the one described in the Models Section will remain the same if the activity in STN described in Eq 33 is the same as that described by Eq 16. Equating the arguments of exponentials in the two equations, one obtains the following condition:
wCSCTXk−wPSPROk=CTXk−STN+logSTN
(36)
Substituting Eqs 34 and 35, 28 and 29, and the conditions of Eq 32, we get:
(wCS+wXPbPwPS)CTXk−STN+logSTN=CTXk−STN+logSTN
(37)
It is evident that the above condition is satisfied when:
wCS+bPwXPwPS=1
(38)
Thus, if the GPe neurons respond to their inputs as described by Eqs 28 and 29, the computations in the model do not change if a portion of cortical input is delivered to the prototypic neurons (via striatum) rather than to STN.
What could be the benefit of the existence of two alternative routes by which the evidence for actions can be provided to the STN? The version of the model presented in this paper describes decision making in highly practiced tasks. It assumes that the mapping between stimuli and actions is already consolidated in cortico-cortical connections, and that information from sensory cortex is integrated with prior evidence in the frontal cortex (see Fig 1). Thus the direct connection from frontal cortex to STN is the fastest route to provide the integrated evidence for actions to STN. However, experimental data [43] suggest that while acquiring a task the stimulus-response mapping is initially learnt in striatum, and many computational models of this learning process have been proposed [4,44]. Consequently, a version of the optimal action selection model describing newly learnt tasks [10] assumes that the information brought by stimulus is integrated with prior evidence in the striatum. Thus in this case the connection D1-PRO-STN is a fast route to provide integrated evidence to STN (otherwise it would have to go via D1-OUT-TH-CTX-STN). Thus in summary, the two alternative routes to STN could allow fast information transfer to STN in different phases of task acquisition.
It has been also shown that arkypallidal neurons innervate striatum to a much larger extent than prototypic neurons [20]. Moreover, arkypallidal neurons are the only type of GPe neuron that has been shown so far to innervate striatal projection neurons [20]. This connection could provide a quick route for the STN feedback to reach the striatum (Fig 8C) and thus normalize the probabilities of actions represented in striatum. Without this connection, the feedback from STN would need to take a long route via thalamus and cortex (Fig 8C). We show below that adding the connection from arkypallidal neurons to striatal D1 neurons and appropriately reducing the connections from arkypallidal neurons to STN (via prototypic neurons) would not change the activity in the output nuclei in the model, and hence the model would still implement optimal action selection (it is not known whether the arkypallidal neurons project to D1 and/or D2 striatal neurons, but we only consider here the projections to D1 neurons, as the D2 neurons are not included in the model in the MSPRT framework–see above).
Let us consider a modified version of the model, in which we allow arbitrary strengths of connections from arkypallidal neurons to D1 striatal neurons, which we denote by wAX (for simplicity we no longer consider connections from striatum to GPe). The activities of neural population in STN, GPe and output nuclei in this model become:
STNk=exp(CTXk−wPSPROk)
(39)
PROk=fP(wSPSTN−wAPARKk)
(40)
ARKk=fA(wSASTN)
(41)
OUTk=−CTXk+wAXARKk+STN
(42)
Let us now recall that in the original model in the Models Section, the activity in the output nuclei was the following function of cortical activity:
OUTk=−CTXk+log∑k=1NexpCTXk
(43)
In order for the two models to select action in the same way, they need to have the same levels of activity in the output nuclei, thus comparing Eqs 42 and 43 we obtain the following condition:
wAXARKk+STN=log∑k=1NexpCTXk
(44)
We will now compute how the right hand side of Eq 44 depends on parameters of GPe neurons in the modified model. Substituting Eqs 40 and 41 into 39, we obtain:
STNk=exp(CTXk−wPSfP(wSPSTN−wAPfA(wSASTN)))
(45)
Summing over alternatives, taking logarithm and re-arranging, we obtain:
log∑k=1NexpCTXk=logSTN+wPSfP(wSPSTN−wAPfA(wSASTN))
(46)
Substituting Eqs 46 into 44 and rearranging terms we obtain a general condition that needs to be satisfied for the modified model to be equivalent to that in the Model Section:
STN−logSTN=wPSfP(wSPSTN−wAPfA(wSASTN))−wAXfA(wSASTN)
(47)
Substituting Eqs 28 and 29, we observe that the above condition becomes satisfied when:
{bPwPSwSP−bPbAwPSwSAwAP−bAwSAwAX=1bPcAwAPwPS+cAwAX=1aPwPS=bPaAwAPwPS+aAwAXwSA=1
(48)
The above conditions are generalizations of those in Eq 32. In particular, the second condition implies, that the computations in the model will not change if a proportion of output from arkypallidal neurons is sent to striatum instead of to STN via the prototypic neurons. This rebalancing of the network does not change the computation performed by the model, but could stabilize the dynamics in the striatum-GPe network (as pathway D1-PRO-STN-ARK-D1 is effectively inhibitory so forms a negative feedback loop).
This paper presents a model of the microcircuits within STN-GPe elucidating how this network could compute the logarithm of the normalization term in Bayes’ theorem that underpins the procedure for optimal action selection. The model is based on recently reported connectivity of two types of GPe neurons, and the shapes of their f-I curves. Our results suggest that GPe neurons have a diversity of structural and electrophysiological properties necessary for contributing to optimal action selection in the cortico-basal-ganglia circuit.
In the optimal action selection model, the input from the STN to the output nuclei plays a very similar function as in other models of action selection [45,46]. In particular, the surround inhibition model [45] assumes that when one of the actions is selected, the input from the STN ensures that others are inhibited. In the optimal action selection model, the input from STN normalizes the neural representation of probabilities so that they add up to 1. Hence it ensures that if estimated probability of one action increases, the probabilities of other action decrease (as in [45]). The conflict model [46] proposes that when neural populations selective for two different actions are both active, the input from the STN postpones the action initiation. In the optimal action selection model, the input from the STN ensures that if neurons representing two actions both receive equally high inputs, the probabilities of these two actions will not exceed 0.5, thus none of the actions will be initiated until the conflict is resolved (as in [46]). Thus the input from the STN in the optimal action selection model fulfils the roles assigned to it by previous theories, but the model additionally proposes that this input is a particular function of cortical activity (Eq 7) which allows optimal action selection. To compute this function the STN needs to interact with the neurons in the GPe.
The range of values of the weights in Fig 6B should not be treated as a precise prediction on the connectivity of GPe neurons, mostly because other key inputs to GPe neurons (e.g. from striatum and thalamus) were not considered to avoid introducing additional free parameters in the optimization procedure. Some of the fitted weights in Fig 6B did not match previous estimates of the effective connectivity of the STN-GPe network in Parkinsonism [36] (the difference occurred in the connection weights between the two types of GPe neuron as well as from STN to GPe). However, please note that the two sets of weights were based on different fitting procedures applied to different data sets recorded from different animals in different labs, and most importantly, in vitro data of neurons from dopamine-intact animals were used here, while in vivo data from animal models of Parkinson’s disease were used in the previous study [36]. Dopamine depletion may change the connectivity of prototypic and arkypallidal neurons, as it changes the relationship between their in vivo firing rates, such that the prototypic neurons, which in intact animals are more active than arkypallidal [18,19], become less active [19,21]. This parallels the differences in estimated connectivity from STN to prototypic and arkypallidal neurons, as in the present study both populations were estimated to receive similar excitation from STN (Fig 6B) whereas, in the previous study of the Parkinsonian STN-GPe network, prototypic neurons were estimated to receive less input from STN than arkypallidal neurons [36]. The previous estimates of effective connectivity [36] are consistent with our analysis of connections from striatum to GPe as they both suggest that prototypic neurons should receive more input from striatum than arkypallidal neurons.
While finding the connections of GPe neurons for which they compute function STN–log STN, we accounted for differences in the shapes of f-I curves for each neuron. Such diversity contributed to making the system high dimensional, thus enabling it to find multiple solutions to the optimisation problem. Contrary to our initial expectation, the different solutions did not exhibit a specific pattern of connectivity. Indeed, one might predict that linear prototypic neurons project to STN more strongly than logarithmic prototypic neurons project to STN. This was not the case, which highlights the flexibility and robustness of the network considered to find solutions that support optimal decision making.
Simulations investigating the dynamics of STN-GPe circuit (Fig 7) revealed that a circuit with realistic time constants and synaptic transmission delays converged to a state in which the activity of STN neurons encoded log P(S), except for high cortical input when the network produced sustained oscillations. Nevertheless, it has to be noted that weight parameters were estimated using a model without delays, so it is possible that other combinations of weights can be found for which the STN-GPe circuit performs the computation required for optimal action selection, but is more stable. Furthermore, even when the oscillations are present in Fig 7, the firing rate averaged over time is close to that required for the optimal action selection. It is also possible that log P(S) is encoded in the power of oscillations in STN in addition to the firing rate. The frequency of oscillations observed in Fig 7 is close to the beta range, and the beta oscillations are thought to be involved in inhibition of movement [47] and their power in STN is increased in Parkinson’s disease [48]. Given that the feedback from STN also slows down action initiation in the optimal action selection model (e.g. in the presence of conflict–see earlier Discussion), its representation in the power of beta oscillations would be consistent with the akinetic properties of these oscillations.
In this paper we have shown that the connections of GPe neurons required for optimal action selection can be found. However, we have not shown how these connections arise from a self-organization process employing local plasticity rules, in which the change in a synaptic weight depends on the activities of presynaptic neurons, postsynaptic neurons and levels of neuromodulators. Developing such plasticity rules would be an important direction for future work.
The model presented in this paper for simplicity assumed that each GPe neuron received input from all STN neurons. Although the projections from STN to GPe are widespread [27], single STN neurons do not contact all GPe neurons. It will be important to investigate in the future if a model with a more realistic pattern of connectivity between STN and GPe can also approximate optimal action selection.
So far the main experimental support for the optimal action selection model comes from in vitro studies of properties of isolated basal ganglia neurons: in the original paper introducing the model [9] it was already pointed out that the f-I curves of STN neurons have exponential shape for a wide range of firing rates, and here we show that the f-I curves of two types of GPe neuron can support the computation of function STN–log STN. To provide further support for the model, it is necessary to show that neurons embedded in the circuit in vivo also behave in the predicted way. For example, one could investigate how the activity of GPe neurons in vivo depends on the activity in STN (which could, for example, be parametrically manipulated using optogenetic actuators). The model predicts that the activity of prototypic GPe neurons should be a convex function of STN activity, proportional to STN–log STN, while the activity of arkypallidal neurons should be a concave function of STN activity. As far as we are aware, this prediction is specific to the model proposed in this paper, and not made by any other model of action selection in the basal ganglia.
A crucial test for the optimal action selection model would involve measuring whether the average activity in the STN during decision making encodes log P(S) (see Fig 1). For example, human participants could be asked to perform a task equivalent to that described at the start of Models section, and the STN activity after each cue could be measured either with high resolution fMRI or via deep brain stimulation electrodes, if the study were performed with Parkinson’s patients. In that case the model would predict higher amplitude of oscillations in STN after presenting stimuli with higher P(S), because the simulations in Fig 7 suggest that more oscillatory activity is produced with higher STN activity.
The framework of the model we present here could be employed in the future to investigate how the computations in the STN-GPe circuit change in cases of dysfunction, such as in Parkinson’s disease where the chronic loss of dopamine grossly disturbs circuit dynamics to impede the initiation and performance of actions. Our model emphasises that the f-I curve shapes and connection strengths of STN and GPe neurons are important for action selection. It is thus noteworthy that the shapes of STN neuron f-I curves depend on postsynaptic dopamine receptors [49], that the strengths of glutamatergic and GABAergic inputs to STN and GPe neurons are tuned by presynaptic dopamine receptors [50–52], and that the impact and structure of pallidosubthalamic connections are altered after dopamine depletion [53] as a result of homeostatic plasticity [54]. Our model also predicts that the autonomous firing rates of prototypic and arkypallidal neurons are necessarily different, with the former cell type exhibiting substantially higher rates of activity. With this in mind, it is notable that the autonomous firing of GPe neurons is grossly disturbed (and lost in a subset of neurons) in experimental Parkinsonism [55,56]. In conclusion, current evidence suggests that there are multiple ways by which dopamine loss in Parkinson’s disease could affect the input-output functions and connection strengths of STN and GPe neurons, such that, ultimately, prototypic GPe neurons might not efficiently compute the function STN–log STN that supports the optimal action selection.
|
10.1371/journal.pgen.1002739 | Histone H3 Localizes to the Centromeric DNA in Budding Yeast | During cell division, segregation of sister chromatids to daughter cells is achieved by the poleward pulling force of microtubules, which attach to the chromatids by means of a multiprotein complex, the kinetochore. Kinetochores assemble at the centromeric DNA organized by specialized centromeric nucleosomes. In contrast to other eukaryotes, which typically have large repetitive centromeric regions, budding yeast CEN DNA is defined by a 125 bp sequence and assembles a single centromeric nucleosome. In budding yeast, as well as in other eukaryotes, the Cse4 histone variant (known in vertebrates as CENP-A) is believed to substitute for histone H3 at the centromeric nucleosome. However, the exact composition of the CEN nucleosome remains a subject of debate. We report the use of a novel ChIP approach to reveal the composition of the centromeric nucleosome and its localization on CEN DNA in budding yeast. Surprisingly, we observed a strong interaction of H3, as well as Cse4, H4, H2A, and H2B, but not histone chaperone Scm3 (HJURP in human) with the centromeric DNA. H3 localizes to centromeric DNA at all stages of the cell cycle. Using a sequential ChIP approach, we could demonstrate the co-occupancy of H3 and Cse4 at the CEN DNA. Our results favor a H3-Cse4 heterotypic octamer at the budding yeast centromere. Whether or not our model is correct, any future model will have to account for the stable association of histone H3 with the centromeric DNA.
| During cell division, replicated DNA molecules are pulled to daughter cells by microtubules, which originate at the spindle poles and attach to a multiprotein complex, the kinetochore. The kinetochore assembles at a special region of the chromosome, termed the centromere. The kinetochore is comprised of more than 50 different proteins whose precise functions are far from being fully understood. The kinetochore assembles on the foundation of a specialized centromeric nucleosome. A nucleosome is a complex of eight subunits, termed histones, which compacts the DNA by wrapping it around itself in 1.7 turns of a superhelix. The centromeric nucleosome is very special, and its stoichiometry and structure are a subject of intense debate. It is believed that the centromeric nucleosome is devoid of histone H3 and instead contains its variant, termed CENP-A in vertebrates or Cse4 in budding yeast. Here we report that in budding yeast both CENP-A and histone H3 localize to a small centromeric DNA fragment that, due to its size, cannot accommodate more than a single nucleosome. Our results necessitate a revision of what is known about the structure of the inner kinetochore and the role of CENP-A in its assembly.
| During eukaryotic cell division sister chromatids, containing identical copies of genetic information, are pulled apart and driven towards opposite spindle poles by the microtubules of the mitotic spindle, which attach to the centromeric DNA sequences of the sisters via kinetochore protein complexes. It is imperative for proper chromosomal segregation that each chromosome assembles the kinetochore only at one site. The sites of kinetochore assembly are marked by specialized nucleosomes. Budding yeast represents the simplest case in which a single microtubule attaches to the so-called “point” kinetochore assembled around a single centromeric nucleosome. More complicated “regional” centromeres of most other eukaryotes are composed of arrays of specialized centromeric nucleosomes interspersed with conventional nucleosomes [1] and support the assembly of several microtubule attachment sites.
Centromeric nucleosomes were reported to have histone H3 substituted by a histone variant, CENP-A, called Cse4 in budding yeast [2]. It displays more than 60% similarity with the conventional histone H3 within the histone fold domain and has an additional N-terminal extension [3]. CENP-A has been demonstrated to co-purify with a subset of kinetochore proteins and is likely to provide interaction surfaces for kinetochore assembly [4], [5]. Recruitment of CENP-A to centromeric DNA requires the CENP-A targeting domain (CATD), comprised of loop1 and the α2-helix [6], [7], and is regulated by a number of other proteins [8]. One example is the non histone protein Scm3 (HJURP in human [9]), which is believed to be a histone chaperone required for recruitment of CENP-A to centromeres [10]–[18]. CENP-A overexpression in metazoans [19] and budding yeast [20] leads to its mislocalization. In budding yeast mislocalized Cse4 is very unstable [21]. Although budding yeast [22] and fission yeast [14], [23], [24] appear to be an exception, in several organisms CENP-A is loaded on the DNA outside of S phase, in anaphase of mitosis or the following G1 [25], [26], when it is proposed to replace histone H3.
Despite a significant progress in the field, the exact function of CENP-A at the centromere remains a mystery. CENP-A and H4 were reported to form a more compact and conformationally more rigid heterotetramer compared to the heterotetramer of histones H3 and H4 [6], [27]. However, the significance of the structural differences between H3 and CENP-A to their function is unknown. Even the question of the exact composition and localization of centromeric nucleosomes has not been resolved to date and remains the subject of controversy [28]. Besides an octamer composed of two molecules each of CENP-A, H2A, H2B and H4, a hexamer model in which Scm3 replaces H2A and H2B [11], [17] and a hemisome model which proposes a tetramer consisting of one copy each of Cse4, H4, H2A and H2B [29]–[32] were also proposed. Regional centromeres of higher eukaryotes can accommodate different versions of CENP-A-containing nucleosomes. While budding yeast with their point centromeres is an appealing model system to study the centromeric nucleosome, it is possible that the yeast centromeric nucleosome might also possess unique features.
Here we report the results of our analysis of the yeast centromeric nucleosome using a novel chromatin immunoprecipitation technique and discuss them in the context of the previously proposed models of the CENP-A containing nucleosome.
The composition of the centromeric nucleosome was previously analyzed by means of chromatin immunoprecipitation (ChIP) [11], [12] in yeast. In a conventional ChIP approach proteins are chemically cross-linked to DNA, the chromatin is fragmented by sonication to about 500 bp size, and immunoprecipitated fragments are identified in PCR or microarray hybridization assays. This approach suffers certain drawbacks when applied to the centromere. The DNA fragment size is much larger than the region accommodated by a conventional nucleosome (146 bp), which limits the resolution. This problem can in principle be overcome by the treatment of chromatin with micrococcal nuclease, which specifically digests the internucleosomal linker DNA. However the size of kinetochore footprint is highly variable depending on the digest conditions [33], [34] and apparently poses an accessibility problem for antibodies since the efficiency of the co-immunoprecipitation of the CEN DNA with canonical histones is very low compared to pericentric regions [11], [12], [35]. In addition, PCR with a specific pair of primers or microarray hybridization detect larger DNA fragments without identifying them by size, which imposes further limits on resolution.
We developed new versions of ChIP to reveal the composition of the centromeric nucleosome in budding yeast. There are three main differences from conventional ChIP. First, we performed our experiments with and without the chromatin cross-linking. We reasoned that omitting cross-linking improves the accessibility of the centromeric nucleosome to antibodies and prevents potential artifacts due to the cross-linking of loosely associated proteins. However, because cross-linking prevents local re-arrangements due to nucleosomal sliding along the DNA, we also included cross-linked samples in our analysis. Second, we flanked CEN DNA by restriction sites and excised it by a specific endonuclease similar to earlier studies by [36]. Finally, analysis of the immunoprecipitated DNA was performed using methods that identify the isolated fragments by size, initially by a Southern blot with specific probes hybridizing to the excised CEN fragment. In experiments where qPCR with a specific pair of primers was used, the immunoprecipitated DNA was size-fractionated prior to PCR to preclude the detection of uncut DNA. The Biggins's laboratory recently employed a similar approach [37]. In this study, micrococcal-nuclease digested chromatin was immunoprecipitated with an anti-Cse4 antibody and analyzed by Southern blot. The results demonstrated a single Cse4 nucleosome positioned at the budding yeast centromere but did not address its composition further.
In our initial experiments we used a small minichromosome that contained the CEN region of chromosome IV (Figure S1A). We utilized strains with HA-tagged versions of H3 and Cse4 and found that the minichromosome can be specifically co-immunoprecipitated with an anti-HA antibody even in the absence of cross-linking (Figure 1A). This result demonstrates that the minichromosome assembles conventional nucleosomes as well as a centromeric nucleosome. Next, we tested whether it is possible to digest the minichromosome in yeast cell lysate and subsequently immunoprecipitate the fragments. We constructed minichromosomes with BglII sites at different positions with respect to CEN. The digest efficiency was highly variable depending on the position of the BglII site (Figure S1B). It was previously reported that the centromeric DNA is inaccessible for the nuclease digest [33], [34]. However, under our conditions it was possible to excise CEN DNA and even to cut it between CDEII and CDEIII in agreement with the previous results by [38], [39].
In subsequent ChIP experiments we used a minichromosome with BglII restriction sites 50 bp upstream and downstream of CEN4 boundaries flanking a 214 bp CEN fragment. The chromatin was digested with the endonuclease BglII and immunoprecipitated with an anti-HA antibody (Figure 1B). A probe hybridizing to the TRP1 gene located on the minichromosome outside of CEN was used for the Southern blot. Due to an incomplete chromatin digest, a linearized full-length minichromosome and a CEN-less fragment could be detected. Only the full-length linearized minichromosome co-immunoprecipitated with Cse4-HA6 while both the full-length linearized minichromosome and the CEN-less fragment were recovered with HA-tagged histones H4, H2A, H2B and H3 (Figure 1C). Therefore, although the minichromosomes assemble conventional nucleosomes along their entire length, only CEN DNA is associated with Cse4, which is in agreement with [37]. Since it was proposed recently that the Scm3 histone chaperone might replace H2A/H2B dimers in the centromeric nucleosome [11], [17] we performed the minichromosome ChIP with the Scm3-HA6 strain. We could not co-immunoprecipitate the minichromosome with HA-tagged Scm3 under our conditions indicating that Scm3 is unlikely to be a part of the centromeric nucleosome (Figure 1C).
The observation that no CEN-less fragment was recovered in the Cse4-HA6 immunoprecipitation rules out lateral sliding of Cse4 nucleosome during the course of the immunoprecipitation as well as tethering of DNA fragments via protein-protein interactions, e.g., between centromeric and conventional nucleosomes in our assay. The efficiency of immunoprecipitation of the minichromosome fragments of approximately 1000 bp and longer was exceptionally high and close to 100%. When a 930 bp fragment from ARS1 until position +50 downstream of CDEIII was excised, it could be depleted from yeast cell lysate with anti-HA antibodies recognizing Cse4-HA6 while virtually none of the remaining CEN-less fragment of the minichromosome could be detected on the beads (Figure S2). Considering the immediate proximity of the +50 cutting site to the centromere it is highly unlikely that there was a significant local rearrangement of nucleosomes and/or tethering of the CEN fragment to the rest of the minichromosome under our experimental conditions.
The detection of the small 214 bp CEN fragment was very inefficient using the 32P-labelled probe. Therefore we employed a digoxygenin (DIG)-labeled locked nucleic acid (LNA) oligonucleotide (Figure 1D) with improved hybridization properties [40]. Using the LNA probe it was possible to detect the 214 bp fragment released from 6 pg of the minichromosome which corresponds to about 0.1% efficiency of immunoprecipitation starting with 150 ml of yeast culture in the early log phase (Figure S3). We could detect the 214 bp CEN fragment in the immunoprecipitates with Cse4, H4, H2A and H2B. Surprisingly, we reproducibly observed an interaction of H3 with the 214 bp CEN fragment using this method (Figure 1D). This was in contrast with previous studies proposing that H3 is replaced by Cse4 at the centromere [2].
We next tested whether the interaction of H3 with CEN is dependent on the cell cycle stage as it is possible that Cse4 replaces H3 at a specific point in the cell cycle. The notion that the composition of the centromeric nucleosome might vary through the cell cycle was proposed earlier [17], [28]. Yeast cultures were arrested in G1-phase with alpha-factor and in G2-phase with nocodazole/benomyl (Figure S4B), and chromatin was digested with BglII to release the 214 bp CEN fragment prior to immunoprecipitation. Both H3 and Cse4, as well as H2B, were found to be associated with CEN in G1-phase and in G2-phase (Figure 2A).
Although nearly a 100% efficiency of co-immunoprecipitation of the minichromosomes with Cse4-HA6 (Figure 1A) indicated that it is unlikely to be the case, it is possible that a fraction of minichromosomes assemble a conventional nucleosome at the centromere and this would explain the association of H3 with CEN DNA in the above experiments. To address this possibility we adapted our ChIP approach to the native centromeres on the chromosomes and introduced BglII restriction sites 50 bp upstream and downstream of CEN on chromosome IV. The excised “native” 214 bp CEN4 fragment could be efficiently co-immunoprecipitated with H3-HA3 and Cse4-HA6 (Figure 2B). We conclude that both histones H3 and Cse4 localize to centromeric DNA in budding yeast.
In order to rule out the possibility that Cse4 is replaced by H3 during our immunoprecipitation procedure, we mixed yeast cell lysate of an H3-HA3 strain that does not carry minichromosomes with lysate of an untagged H3 strain carrying the minichromosomes. We could not observe any immunoprecipitation of the minichromosome with anti-HA antibody from those mixed lysates (Figure 2C). Thus there is little or no turnover of minichromosome-associated H3 in our cell lysates.
However, this experiment could not rule out local rearrangement of nucleosomes such as lateral sliding in the course of our experimental procedure, which included long incubations. Therefore we cross-linked proteins to DNA with formaldehyde prior to immunoprecipitation. Adding formaldehyde to the spheroplasts dramatically reduced the efficiency of centromeric DNA co-immunoprecipitation with either Cse4 or H3. This was partially due to the low yield of the minichromosome in the cleared lysate after centrifugation presumably because the minichromosomes were cross-linked to larger structures. However, when formaldehyde was added directly to yeast lysate the immunoprecipitation was not impeded. In order to minimize the potential rearrangement of nucleosomes after cell lysis, the duration of the restriction digest of the minichromosomes was limited to 5 minutes followed by formaldehyde addition and immunoprecipitation. We were able to efficiently co-immunoprecipitate the 214 bp CEN fragment with both Cse4 and H3 after cross-linking (Figure S5A). Therefore, it is unlikely that the detection of H3 at the CEN DNA is due to nucleosomal sliding during our experimental procedure.
A qPCR-based approach was employed to compare the efficiencies of co-immunoprecipitation of the CEN DNA with H3-HA3 and Cse4-HA6. After excision of the 214 bp CEN fragment CEN DNA was co-immunoprecipitated with Cse4-HA or H3-HA using anti-HA antibodies, eluted off the beads using SDS, size-fractionated via agarose gel-electrophoresis to separate it from full-length minichromosome and quantified using a quantitative PCR reaction. Using this procedure, we ensured that the 214 bp CEN fragment was exclusively detected since no PCR product was obtained when the restriction digest step was omitted (Figure 3A). We did not observe any significant differences in ChIP efficiencies with H3 and Cse4 when the same anti-HA antibody was used. Similar IP/input ratios were observed with and without crosslink (Figure 3B) with the CEN DNA located on a minichromosome and on the native chromosome IV flanked by restriction sites (Figure 3C). Thus we have no indication that only some centromeres are associated with H3.
The association of H3 and Cse4 with yeast centromeres can be mutually exclusive, i.e., a fraction of the centromeres are occupied by the Cse4 nucleosome while a different fraction assembles a conventional nucleosome containing H3. Alternatively, H3 and Cse4 are co-occupying the centromeric DNA at the same time. In order to distinguish between these two possibilities we performed a sequential ChIP experiment. After excision of the 214 bp CEN fragment and formaldehyde cross-linking CEN DNA was co-immunoprecipitated with Cse4-Myc using anti-Myc antibodies covalently coupled to the beads (Figure S5B and S5C), eluted off the beads using SDS, and re-immunoprecipitated with anti-HA antibodies recognizing H3-HA. The CEN DNA fragment eluted off the beads was decross-linked, size-fractionated via agarose gel-electrophoresis to separate it from uncut DNA, and quantified using a quantitative PCR reaction (Figure 3D–3F). The efficiency of the second immunoprecipitation step in this experiment was approximately 100 fold higher than the “mock” immunoprecipitation from a strain in which only Cse4 was tagged and was comparable to that of H3-HA re-immunoprecipitation in the experiment where both the first and the second steps were performed with anti-HA antibodies. Similar results were obtained when CEN DNA was excised from the minichromosome (Figure 3E) or native chromosome (Figure 3F). We conclude that H3 and Cse4 co-exist at least at some centromeres. Unfortunately, we could not perform the reverse experiment, i.e., to immunoprecipitate the CEN DNA via HA-tagged histone H3 and then re-precipitate via Myc-tagged Cse4, since we could not re-precipitate CEN DNA from Cse4-Myc strain with anti-Myc antibody in 0.1% SDS. Switching the tags was also unsuccessful since the H3-Myc6 strain was not viable.
Because the length of our excised centromeric fragment (214 bp) is much shorter than would be necessary to accommodate two conventional nucleosomes (292 bp assuming no linker DNA in-between) or a conventional nucleosome and a Cse4 nucleosome (268 bp if the Cse4 nucleosome organizes only 121 bp of DNA [41]), it is plausible that the centromeric nucleosome is a heterotypic octamer with one molecule of H3 and one molecule of Cse4. If the structure of this hypothetical heterotypic nucleosome is similar to the structure of the conventional nucleosome and the CENP-A containing nucleosome [41], [42], histones H3 and Cse4 are expected to form a four-helix bundle with parts of their α2 and α3 helices. In vertebrates and many other organisms the α2 helix of H3 contains a cysteine residue, C110. These cysteine residues from two histones H3 within the same nucleosome are within 6.2 Å from each other [42] and were reported to form a disulfide bond under oxidizing conditions in vitro [43]. In human CENP-A the corresponding residue is a leucine, L112, although CENP-A proteins from some other mammals, such as platypus, as well as birds and amphibians have a cysteine in this position. In the recently reported crystal structures of human CENP-A nucleosome the two leucines 112 are 4.8–5.7 Å apart [27], [41], which should allow cross-linking if they are mutated to cysteines. (Figure S6A). In order to test whether a cross-link between two Cse4 molecules or between Cse4 and H3 is at all possible we co-expressed the histone fold domain of Cse4-Cys and the full-length H3-Cys in bacteria. We could observe the formation of spontaneous covalently cross-linked H3 homodimers, Cse4 homodimers and some H3/Cse4 heterodimers. The dimers were detected after denaturing SDS-electrophoresis and could be resolved by β-mercaptoethanol treatment indicating that they indeed resulted from the formation of the disulfide bond between the cysteine residues (Figure S6B).
We reasoned that disulfide bond formation between the two α2 helix cysteines would only be possible if the two histones form a four helix bundle and the ability to cross-link Cse4 and H3 would be a test of a heterotypic octamer model. Since in budding yeast neither H3 nor Cse4 contain cysteine residues, we mutated the corresponding alanine 111 and leucine 204 to cysteines. We were able to cross-link homodimers of H3-Cys in crude lysates and on isolated chromatin in the presence of 5,5′-dithiobis-(2-nitrobenzoic acid) (DTNB, Ellman's reagent), which has been reported to facilitate intermolecular disulfide bond formation between H3 histones in chicken nucleosomes [44] (Figure S7A). We could also cross-link H3-Cys histones using cysteine-specific cross-linkers, bBBr and BMOE. However, we did not observe a reproducible cross-link either between two Cse4-Cys molecules or between Cse4-Cys and H3-Cys (Figure S7B) in crude yeast lysate or isolated chromatin.
Thus we currently have no direct evidence for the presence of the heterotypic octamer at budding yeast centromeres. It is possible that the heterotypic nucleosome has a very unusual structure compared to the conventional H3-H3 nucleosome [42] or the human CENP-A-CENP-A octamer that were recently reported [27], [41] and that this structure does not allow for the cysteine cross-link. It remains to be confirmed whether the cysteines can be cross-linked in the context of the fully assembled octamers.
An alternative to the octamer is the hemisome model, which proposes a tetramer consisting of Cse4, H4, H2A and H2B histones [30], [31]. Our refinement of this model will imply that in budding yeast in the immediate vicinity of the Cse4 hemisome there is either a conventional nucleosome or, possibly, an H3-containing hemisome. According to the recently reported structure, the human CENP-A-containing octamer assembled in vitro organizes 121 bp of DNA [41] while a conventional nucleosome wraps 147 bp of DNA. Thus, a Cse4 hemisome and a conventional nucleosome without any linker in-between would require approximately 207 bp which would fit with the size of our excised centromeric fragment of 214 bp. An important and testable prediction of this model is that Cse4 and histone H3 are incorporated into distinct structures, which can be potentially mapped to different stretches of DNA.
The budding yeast centromere is defined by a 125 bp sequence [45] consisting of three elements. CDEI is a non-essential 8 bp palindrome, CDEII is 78–86 bp long and is composed of 87–98% A/T, and CDEIII is a highly conserved 25 bp sequence which binds the CBF3 protein complex [46]. We conducted a series of experiments in which we tested whether Cse4 and histone H3 associate with distinct elements within CEN DNA. It was reported earlier that CSE4 genetically interacts with CDEI and CDEII but not with CDEIII [47] suggesting that the Cse4-containing nucleosome is localized upstream of the CDEII/CDEIII boundary. Since we were able to cut the minichromosome between CDEII and CDEIII we hoped to gain further insights in the exact localization of Cse4 with regard to CEN by using our ChIP approach. We created a minichromosome with a restriction site between CDEII and CDEIII and a restriction site outside of the CEN DNA, in ARS1. Using our ChIP approach we were able to co-immunoprecipitate Cse4-HA6 with both the CDEI/CDEII and the CDEIII-containing fragments (Figure 4A) suggesting that the centromeric nucleosome straddles the boundary between CDEII and CDEIII. However, an interaction with the CDEIII fragment appeared less efficient, indicating that the Cse4-containing nucleosome interacts mostly with the CDEI/CDEII region of the CEN DNA. An important corollary from this observation is that in our assay the Cse4-containing nucleosome (or hemisome) is not displaced from the CEN DNA to the edge of the 214 bp fragment.
To gain further insight into spatial distribution of H3 and Cse4-containing nucleosomes on CEN DNA we next excised a 139 bp fragment from position −50 upstream of CDEI until the CDEII/CDEIII boundary. When cross-linked, this fragment could be co-immunoprecipitated with both H3 and Cse4 (Figure 4B). This result demonstrates that H3 is present at the CDEI/II region of the centromere and/or at the preceding 50 bp of the non-centromeric DNA. Since the detection of a fragment containing CDEIII and 50 bp of DNA downstream of the CEN DNA with the LNA probe was not possible, we followed the association of histone H3 and Cse4 with CDEI/II and CDEIII elements using qPCR. Both the fragment containing CDEI/II region with upstream 50 bp and the fragment containing CDEIII region with the downstream 50 bp could be co-immunoprecipitated with HA-tagged Cse4 and histone H3 with and without crosslinking (Figure 4C). Therefore histone H3 and Cse4 appeared to be inseparable when associated with the CEN DNA implying that they are likely to be a part of one and the same structure. We would like to note that since Cse4 is capable of tethering CDEII and CDEIII fragments together (Figure 4A), the co-immunoprecipitation of the small CDEI/II and CDEIII fragments with H3 might be due to the small CDE-containing fragments maintaining the association with the large CDE-less fragment of the minichromosome throughout co-immunoprecipitation. No such tethering was observed when the complete 214 bp CEN DNA containing fragment was excised from the minichromosome (Figure 1C and Figure S2).
Three models of the centromeric nucleosome are proposed in the literature. In the first model the centromeric nucleosome is an octamer, where Cse4/CENP-A replaces histone H3. While octameric nucleosomes with two copies of budding yeast Cse4 [48], [49] or human CENP-A [41] were assembled in vitro, whether only one or both copies of H3 are replaced in vivo is not known. There is evidence from different organisms for and against either of these possibilities. In HeLa cells CENP-A released from chromatin by micrococcal nuclease digestion is still associated with histone H3 even after 2M NaCl treatment resulting in dissociation of H2A and H2B, implying heterotypic tetramers with two histones H4, one H3 and one CENP-A [4]. In contrast, in Drosophila S2 and Kc cells when chromatin is digested with micrococcal nuclease and CENP-A/CID is immunoprecipitated, no H3 co-purifies with CENP-A [1]. It was recently reported that Drosophila CENP-A/CID forms homodimers in vivo, which are unexpectedly very salt-sensitive but could be crosslinked via cysteines in the four-helix bundle after a prolonged incubation [50]. The authors did not exclude the formation of H3-CENP-A/CID heterodimers in addition to CENP-A/CID homodimers and it remains possible that different forms of CENP-A/CID nucleosomes are simultaneously present at the regional centromeres of Drosophila and possibly other higher eukaryotes.
In this study we demonstrate that a budding yeast centromeric DNA fragment of only 214 bp is associated in vivo with both H3 and Cse4. We can exclude a homotypic octamer with two copies of Cse4. Our experiments suggest a very intimate spatial association between the conventional histone H3 and centromeric Cse4. This association cannot be explained if the Cse4-containing centromeric nucleosome is separated from the neighboring conventional H3 nucleosomes by spacer DNA as was proposed recently [51] but rather suggests that H3 and Cse4 co-occupy the CEN DNA fragment of only 214 bp in length. We favor the Cse4-H3 heterotypic octamer model (Figure 5, model 1). This octamer appears to be resistant to cysteine cross-linking, which might be due to the reduced stability of the four-helix bundle similar to the Drosophila CENP-A/CID [50].
The hexamer model postulates that in budding yeast the non-histone protein Scm3 replaces H2A and H2B and the nucleosome is composed of two copies each of Scm3, CENP-A and H4 [11], [17]. Although it was initially proposed that the Scm3 dimer constitutes an integral part of the centromeric hexasome [11], the recent structures of budding yeast Scm3 associated with Cse4/H4 [16], [18] and human HJURP in complex with CENP-A/H4 [52], [53] revealed that binding of DNA as well as the (Cse4/H4)2 heterotetramer formation are incompatible with Scm3 binding. In the experiments in vitro it was demonstrated that Scm3 association with the reconstituted (Cse4/H4)2 nucleosome-like particles depends on a DNA binding domain within Scm3 [17]. Our results are compatible with the view that Scm3 does not form a part of the centromeric nucleosome. Under our experimental conditions we were able to co-immunoprecipitate minichromosomes with Cse4, H4, H2A, H2B and H3 but not with Scm3, which most likely dissociated from the centromere in yeast lysate.
Finally, the hemisome model proposes a tetramer consisting of Cse4, H4, H2A and H2B histones [30]–[32]. According to this model, the Cse4 hemisome is positioned mostly at CDEII [20] and is expected to occupy approximately 60 bp of DNA [41]. This scenario leaves approximately 77 bp on each side of our 214 bp fragment available to accommodate the H3-containing nucleosome(s). We can speculate that a hemisome with Cse4 might, for example, be incorporated into a DNA loop between the two halves of an H3-containing octamer (Figure 5, model 3). This model might explain the tripartite organization of the budding yeast centromere that was observed in the micrococcal nuclease protection pattern [20]. Although it is technically possible that 77 bp upstream and downstream of the hypothetical centromeric hemisome are wrapped around ½ of the flanking conventional nucleosomes (Figure 5, model 4), this model will result in tethering of the excised 214 bp fragment to the rest of the minichromosome which we did not observe (Figure 1C and Figure S2) and therefore can be excluded.
More exotic models can be also considered. Two recent studies compared Cse4-GFP fluorescence in vivo to independent standards and found 3.5–6.0 [54] or even 7.6 [55] Cse4-GFP molecules per budding yeast centromere in anaphase. Even more surprisingly, in prolonged G1 arrest Cse4-GFP fluorescence was reduced more than two-fold [55]. These observations are inconsistent with the notion of a single Cse4 nucleosome at the budding yeast centromere [37]. It was proposed that the budding yeast centromere is in fact a regional centromere with additional Cse4s associated with the flanking DNA similar to the much larger centromeres of higher eukaryotes [54]. However, we could not observe any Cse4 associated with the non-centromeric part of the 2.4 kb minichromosome, which is expected to assemble 10 conventional nucleosomes. Therefore no additional Cse4 nucleosomes assemble, at least at these relatively short flanking sequences. Our results are consistent with those of [20], [56] who did not detect additional Cse4 nucleosomes in centromere-flanking regions by high-resolution mapping of yeast genome. The additional Cse4 molecules at the centromere could result from Cse4 mis-incorporation which is observed in strains overexpressing Cse4 [20] and could potentially be caused by GFP-tagging. Alternatively, additional Cse4 molecules may not be incorporated into the centromeric nucleosome but are rather associated with it via protein-protein and/or protein-DNA interactions (Figure 5, model 2). In this scenario the centromeric nucleosome can be a Cse4-H3 heterotypic octamer to which more Cse4 molecules are bound. Intriguingly, when (Cse4/H4)2 heterotetramers were reconstituted in the presence of Scm3 into nucleosome-like particles on a 207 bp-long high affinity nucleosome positioning DNA sequence in vitro, high molecular weight complexes possibly representing additional Cse4/H4 in loose association with the Cse4/H4/DNA complex were detected [49]. Similar complexes were reported to be assembled in vitro on a 148 bp CEN3 DNA [17].
It is more than a decade now since it was proposed that H3 is replaced by the histone variant Cse4 [2]. Our results appear to contradict this well-established dogma. If Cse4 and H3 indeed co-localize to the centromeric DNA why wasn't it noticed before? We can offer the following explanation. We have noticed that in most publications reporting ChIP experiments at the budding yeast centromere, the absolute efficiency of ChIP of the CEN DNA with H3 and Cse4 is very similar and typically in the range of 1% [11], [35]. The claim that only Cse4 is associated with the CEN DNA is then based on an observation that non-centromeric DNA is co-immunoprecipitated with H3 at about 5 to 10-fold higher rate than CEN DNA while almost no non-CEN DNA is found associated with Cse4 (Figure S8). We suggest that if CEN DNA were generally difficult to immunoprecipitate, for example due to cross-linking of the large number of kinetochore proteins during the in vivo cross-linking, this would explain the reduced efficiency of H3 ChIP at the centromere compared to the chromosomal arms.
Our results appear to contradict those of [35]. This group could co-immunoprecipitate differentially tagged versions of Cse4 from budding yeast but did not observe co-immunoprecipitation of tagged Cse4 and H3. However, one of the tagged Cse4s was expressed from a plasmid and Cse4 overexpression was reported to result in its ectopic incorporation genome-wide into octameric nucleosomes that were not observed in the wild type strain [20]. It remains possible that even in budding yeast there is a degree of heterogeneity in the composition of the centromeric nucleosomes among different chromosomes and that either a homotypic Cse4/Cse4 octamer or a heterotypic Cse4/H3 octamer can provide the essential function.
At this time we can only speculate at the function of H3 at the budding yeast point centromere. It is possible that the presence of two different nucleosomes (or hemisomes), one with Cse4 and one with H3 provides structural asymmetry which might form the basis for two separate surfaces, one facing the sister centromere and another providing the attachment site for the spindle microtubule.
Generation of the minichromosome containing a 850 bp long sequence from chromosome IV encompassing CEN4 was described earlier [57], [58]. A version without Tet operators was used to introduce BglII restriction sites using QuikChange Site-Directed Mutagenesis Kit (Stratagene). A SalI digest and religation was used to remove the pUC19 sequence from the final construct prior to transformation into yeast.
To introduce BglII restriction sites flanking the CEN DNA into the native chromosome IV, the region of CEN4 +/− 200 bp was cloned into the PvuII site of pOM10 (courtesy of Anne Spang) and BglII sites were introduced by mutagenesis. A yeast strain was transformed with a PCR product containing CEN4 DNA with BglII sites, marker, and a CEN flanking sequence. The BglII flanked CEN4 DNA was recombined into the endogenous locus and the marker cassette was removed with Cre recombinase [59] leaving 85 bp of the pOM10/loxP sequence 200 bp downstream of CDEIII (Figure 2B). The whole CEN4 region was sequenced.
Cse4 was tagged with HA6, Myc6 or Myc3 at an internal XbaI site as described in [2]. All other histones were tagged at the C-terminus and the second gene was either left untagged (H4) or deleted (H2A, H2B, H3). The strains are described in Table S1.
Yeast strains transformed with the minichromosome were grown overnight in synthetic medium without tryptophan at 30°C, were inoculated into fresh medium to a final OD600 of 0.2, and grown until the OD600 reached 1.6. For G1 arrest, yeast culture was grown from an OD600 of 0.05 until an OD600 of 0.2 and then arrested with 2 µg/ml alpha factor for 1 hour. After 1 hour, additional 1.5 µg/ml alpha factor was added followed by an additional hour of incubation. For G2/M arrest, 15 µg/ml nocodazole and 10 µg/ml benomyl were added to a yeast culture at an OD600 of 0.65 in YEPD medium, and cells were incubated for 1.5 hours.
Spheroplasting was carried out with lyticase (Sigma, L2524) as described in [60]. Spheroplasts were lysed for 30 min on ice in 2.5 ml of lysis buffer (25 mM HEPES/KOH [pH 8.0], 50 mM KCl, 10 mM MgSO4, 10 mM Na citrate, 25 mM Na sulfite, 0.25% TritonX-100, 1 mM PMSF, 3 mM DTT, 1× complete EDTA-free protease inhibitors (Roche) and 100 µg/ml RNase A). The lysate was cleared by centrifugation at 10,000 rpm for 5 min in an Eppendorf microcentrifuge. For DNA cleavage, lysate was incubated with 1 unit/µl of BglII (NEB) for 2 hours with rotation at 4°C before adding NaCl to a final concentration of 300 mM to stop the digest. For strains with BglII sites on chromosome IV the crude lysate was incubated with BglII and cleared after 2 hours of digestion. Pre-cleared lysate (2 ml) was incubated with 25 µg of anti-HA (12CA5) antibody and 0.5 ml suspension of protein A Dynabeads (Invitrogen) overnight. Beads were washed 3 times with 1.5 ml of the lysis buffer with 300 mM NaCl. Isolated DNA was eluted off the beads two times with 250 µl of 50 mM Tris [pH 8.0], 10 mM EDTA and 1% SDS at 65°C. For cross-linked chromatin the DNA digest with BglII was performed for 5 min at 37°C, the digest was stopped by adding 300 mM NaCl and chromatin was cross-linked by adding 0.1% formaldehyde for 30 min and 125 mM glycine for 15 min on ice. The cross-linked lysate was incubated with protein A Dynabeads covalently coupled to either anti-HA (12CA5) or anti-Myc (9E11) antibody with DMP (dimethyl pimelimidate) according to the manufacturer's guidelines. For the sequential immunoprecipitation the chromatin was eluted off the beads as described above, diluted to 0.1% SDS with lysis buffer with 300 mM NaCl and immunoprecipitated with protein A Dynabeads covalently coupled to anti-HA (12CA5). The DNA was eluted off the beads as above. All the samples were adjusted to 1% SDS final concentration, extracted twice with phenol/chloroform/isoamyl alcohol (25∶24∶1), ethanol precipitated in the presence of 20 µg glycogen (Roche) and samples were dissolved in 20–40 µl TE. For the Southern blots detected with a 32P-labelled probe specific for TRP1 or CEN4, samples were separated on a 1% agarose gel with ethidium bromide and a capillary transfer to Hybond-N+ (GE) was carried out under neutral conditions. Blots were scanned on Personal Molecular Imager (Bio-Rad) and bands quantified with QuantityOne 4.6.7. For Southern blots detected with double-DIG labeled LNA probe (AAAGTTGATTATAAGCATGTGAC, Exiqon) samples were separated on a denaturing 6% TBE polyacrylamide gel followed by an electrophoretic transfer to Hybond-N+ at 80 V for 1 hr in 1× TBE in the Trans-Blot System (Biorad). Hybridization with DIG labeled LNA probe was performed according to instructions of DIG High Prime DNA Labeling and Detection Starter Kit II (Roche). For qPCR the samples were size fractionated on a 2% agarose gel (Certified Low Range Ultra Agarose, Bio-Rad), gel excised to separate from uncut and linear minichromosome and subjected to qPCR with the primers AGTAACTTTTGCCTAAATCAC and TAGGTAGTGCTTTTTTTCCA for the 214 bp CEN4, TAGTAACTTTTGCCTAAATC and TAATAAATAAATTATTTCATTTATGTTT for the 139 bp CDEI/II fragment, and TGTTTATGATTACCGAAACA and TTAGGTAGTGCTTTTTTTCC for the 77 bp CDEIII fragment, qPCR analysis was performed using LightCycler 480 SYBR Green I Master (Roche) according to the manufacturer's manual.
Spheroplasting was carried out using the same procedure as for ChIP. Spheroplasts were washed in 1 M sorbitol and lysed in cold reaction buffer (25 mM Sodium Phosphate [pH 7.0], 100 mM KCl, 2.5 mM MgCl2, 0.25% TritonX-100) for 15 min on ice. Chromatin was pelleted using a low-speed centrifugation (4,000 rpm, 1 min) and the supernatant was discarded. The chromatin pellet was then resuspended in the reaction buffer with varying concentrations of the cross-linker. DTNB (5,5′-dithiobis-(2-nitrobenzoic acid), Sigma) was prepared as a 50 mM stock in DMSO and diluted into the reaction mixture as appropriate. Cross-linking was allowed to proceed for 1 hour on ice. The chromatin was pelleted by centrifugation and resuspended in SDS-PAGE loading dye without DTT or β-mercaptoethanol.
Codon optimized sequences of yeast histone H3-Cys, N-terminally tagged with Avitag (Avidity), and the histone fold domain of Cse4-Cys (D150-end), N-terminally tagged with 6xHis, were cloned either together into pRSFDuet1 (Novagen) or separately, Cse4 in pETDuet1 and H3 in pRSFDuet1, transformed and expressed in BL21 (DE3) according to the manufacturer's instructions. Aliquots of bacterial culture were harvested and resuspended in SDS-PAGE loading buffer with and without β-mercaptoethanol. Samples were separated on a 15% SDS-PAGE and Western blots were analyzed with Streptavidin-HRP (Pierce) for H3-Cys and with anti-Penta-His antibody (Qiagen) for Cse4-Cys.
|
10.1371/journal.pmed.1002879 | Expectations of healthcare quality: A cross-sectional study of internet users in 12 low- and middle-income countries | High satisfaction with healthcare is common in low- and middle-income countries (LMICs), despite widespread quality deficits. This may be due to low expectations because people lack knowledge about what constitutes good quality or are resigned about the quality of available services.
We fielded an internet survey in Argentina, China, Ghana, India, Indonesia, Kenya, Lebanon, Mexico, Morocco, Nigeria, Senegal, and South Africa in 2017 (N = 17,996). It included vignettes describing poor-quality services—inadequate technical or interpersonal care—for 2 conditions. After applying population weights, most of our respondents lived in urban areas (59%), had finished primary school (55%), and were under the age of 50 (75%). Just over half were men (51%), and the vast majority reported that they were in good health (73%). Over half (53%) of our study population rated the quality of vignettes describing poor-quality services as good or better. We used multilevel logistic regression and found that good ratings were associated with less education (no formal schooling versus university education; adjusted odds ratio [AOR] 2.22, 95% CI 1.90–2.59, P < 0.001), better self-reported health (excellent versus poor health; AOR 5.19, 95% CI 4.33–6.21, P < 0.001), history of discrimination in healthcare (AOR 1.47, 95% CI 1.36–1.57, P < 0.001), and male gender (AOR 1.32, 95% CI 1.23–1.41, P < 0.001). The survey did not reach nonusers of the internet thus only representing the internet-using population.
Majorities of the internet-using public in 12 LMICs have low expectations of healthcare quality as evidenced by high ratings given to poor-quality care. Low expectations of health services likely dampen demand for quality, reduce pressure on systems to deliver quality care, and inflate satisfaction ratings. Policies and interventions to raise people’s expectations of the quality of healthcare they receive should be considered in health system quality reforms.
| Expectations of healthcare quality are believed to influence how patients experience and rate healthcare; however, little is known about expectations in low- and middle-income countries (LMICs).
High satisfaction with poor-quality care is common in LMICs; one possible explanation is that expectations of care are low.
This study was conducted to better understand expectations of healthcare quality in LMICs.
We used an innovative internet sampling methodology to collect information from 17,996 individuals in 12 LMICs.
Vignettes describing poor-quality care were used to elicit expectations of healthcare; good ratings for poor-quality care approximate low expectations.
Over half of respondents (53%) rated the poor-quality care vignettes as good quality or better; low expectations were more likely if respondents were less educated, were male, reported good personal health, or had experienced discrimination during a healthcare visit in the past.
Internet users in LMICs are a relatively privileged subgroup in our study countries, limiting the generalizability of our findings.
Populations with low expectations are more likely to be satisfied with poor-quality care, reducing the demand-side pressure that health systems experience to deliver high-quality care.
Raising expectations of quality may be one approach to improving the quality of healthcare in LMICs.
| A growing body of literature describes systematically poor quality of healthcare in low- and middle-income countries (LMICs) today [1–4]. For example, only 21% of providers correctly managed tuberculosis in a study using standardized patients in India [5]. Health workers in 18 LMICs performed on average less than half of recommended reproductive, maternal, newborn, and child health actions during a visit, and a patient in Africa is twice as likely to die after surgery than the global average [2,6]. An analysis of global data estimated that 8.6 million lives lost in LMICs in 2016 could have been prevented by high-quality healthcare; whereas 40% did not have access to care, 60% made it to a facility but did not receive the high-quality care needed to avert death [7]. Nonhealth outcomes such as confidence in the health system and cost of care also suffer in settings of low quality [8].
Despite this, satisfaction with care has been generally high [2,9]. People’s satisfaction with care is related to their expectations of quality, and these expectations can be lowered by information asymmetry (i.e., not knowing what elements of care delivery are optimal) or lack of experience with high-quality services in their environment [10–12]. Low agency and disempowerment may further depress expectations: a range of studies have found that poor and less educated respondents are more likely to rate care as satisfactory [9,11,13,14].
Low expectations are problematic for several reasons. One, if people expect poor-quality care, either because they do not know what high-quality care is or because they have become accustomed to poor-quality care, they are less likely to hold health systems accountable for poor performance. This is a missed opportunity to improve healthcare through feedback. In addition, people with low expectations are less effective in seeking better care. A growing literature in health economics and health services research has found that “active” patients, those who make strategic decisions about where to access care in an effort to receive higher quality services, are able to extract higher quality care from the system [15,16]. They select, bypass, and abandon care based on whether or not a facility is perceived to be able to meet their expectations of quality [15,17,18]. Thus, raising expectations may result in more people obtaining better care and provide feedback to health systems for improvement. Finally, measures of health quality expectations can be used as anchoring vignettes to permit better comparison of self-reported service quality and satisfaction across countries [19–21].
Despite the importance of understanding expectations of healthcare quality, this concept is undertheorized and has been little researched in LMICs [9,11,22]. To address this gap, we assessed the ratings of quality for standardized healthcare vignettes designed to portray poor-quality care among internet users in LMICs. These ratings are considered a measure of expectations of quality. We explored associations between good-quality ratings and user and healthcare factors.
We fielded an internet survey to explore healthcare expectations in 12 LMICs: Argentina, China, Ghana, India, Indonesia, Kenya, Lebanon, Mexico, Morocco, Nigeria, Senegal, and South Africa (see S1 Appendix and S2 Appendix for survey instrument). The study countries were selected because they represent a variety of world regions and comprise a large share of populations living in LMICs. All study countries had internet penetrance over 20% (S3 Appendix). Our original analysis plan for this data set included the research question pertinent to the current analysis, “What are expectations of healthcare in LMICs across socio-demographic and contextual factors?” (S4 Appendix).
We asked web users aged 18 or older about demographics, healthcare utilization, perceptions of healthcare quality, and healthcare vignettes describing poor quality of care. Respondent location was determined using internet protocol (IP) addresses. We collected data during August and September of 2017. Translators translated the survey into local languages, and native speakers then back-translated into English to check for accuracy (S5 Appendix).
Internet surveys allow for collection of a large number of responses across countries while minimizing social desirability bias [23]. We used Random Domain Intercept Technology (RDIT) to reach a wide population of internet users. RDIT “intercepts” internet users who have entered the name of a site that does not exist or has expired and invites them to complete the survey. RDIT produces a sample that is highly representative of the internet-using population [24]. The method has been found to produce stable findings over time. For example, a mental health survey repeatedly conducted in India every month for over 21 months produced consistent estimates with low standard errors as did a vaccine belief survey in Ontario [25–27].
We used several strategies to ensure validity of the responses. IP addresses were monitored to avoid duplicate responses, and proprietary code prevented automated entries by “bots.” In order to hold the respondent’s attention and to reduce thoughtless clicking, we kept the survey short, randomly varied the order of categorical responses, and moved the location of the questions and responses on the screen (see S6 Appendix for a further discussion of our internet survey methodology).
The main study outcome is the respondent’s rating of vignettes illustrating poor quality of healthcare in the domains of technical quality (competence) or interpersonal quality (communication, respectful treatment). A good, very good, or excellent rating is considered a measure of low expectation of service quality. All respondents were shown a vignette about a routine clinic visit for hypertension management (Fig 1). In this visit, the nurse does not check the patient’s blood pressure or ask about his symptoms but changes his medicine and is courteous (“A. blood pressure visit; poor technical quality”). Respondents also received 1 of 3 additional vignettes. The second vignette also describes poor technical quality of care; a patient is seen for an arm injury caused by an accident. The patient’s arm is not examined, and he is not asked about his symptoms (“B. accident visit; poor technical quality”). This vignette was included to explore the effect of the specific health condition on quality ratings. The last 2 vignettes use the same clinical conditions as above but describe poor interpersonal quality of care (“C. blood pressure visit; poor interpersonal quality” and “D. accident visit; poor interpersonal quality”). The last 2 vignettes were included to test the impact of interpersonal quality deficits on overall ratings.
The 4 vignettes were intentionally designed to describe poor quality of care that is discernable to the respondent. Although lay people are unlikely to be able to judge provider adherence to clinical guidelines or choice of correct treatment, the literature has shown that they understand the importance of clinician assessment, including thoroughness of history and physical exam [28–30]. To further assist the respondents in identifying the nature of clinical assessment that is indicated in a visit for hypertension or injury, we identified items that were not completed in the visit. Two of our 4 vignettes describe poor interpersonal quality of care (C. and D.); lay people are expert judges of this, and surveys of patient perceptions of quality, such as the Consumer Assessment of Healthcare Providers and Systems, routinely ask about interpersonal quality of care. Respondents were asked to rate the overall quality of care described in the vignettes using a 5-point categorical response scale (excellent, very good, good, fair, and poor). We defined low expectations as a rating of excellent, very good, or good.
Individual and health system variables believed to influence expectations of healthcare were selected based on the literature on the determination of satisfaction and quality ratings (S7 Appendix). These include demographics, health, nature of local health system, past care experience, and attitudinal positivity [9,11,14,31,32]. We included gender, age, urban or rural residence, and educational attainment. Self-reported general and mental health status were rated using a 5-point categorical response scale (excellent, very good, good, fair, and poor). Educational attainment response options were as follows: completed college or university, some college or university, secondary or high school completed, some secondary or high school, primary school completed, some primary school, or no formal schooling.
Experience with the health system included frequency of use (number of outpatient visits over the last year) and past treatment (having “ever been discriminated against, hassled, or made to feel inferior by a health provider/staff”). Although it is hypothesized that information about healthcare and healthcare quality is likely to influence expectations, our survey did not allow us to explore this factor.
We calculated descriptive statistics for survey respondent characteristics across and within countries for the variables of interest. Weights using age, gender, urban or rural residence, and educational attainment were created to approximate national populations (see S8 Appendix and S9 Appendix). We tested the association between the predictors and good or better quality ratings using multilevel random intercept logistic regression. We include an ordinal logistic model in the appendix for reference (S10 Appendix), although this model was rejected because the assumption of proportional odds was violated. The psychological literature shows that individual personality and tendency towards positivity or negativity affect both survey reporting behavior and expectations [14,33]. Previous studies of patient health preferences have used mental health status to approximate positivity or negativity [14]. Given this literature, we performed an additional analysis in which we added self-reported mental health status to our model. Though self-reported mental health is not a direct proxy for positivity or negativity, personality is likely to play some role in reporting mental health status. To explore potential patterns in the influence of country income or health system performance on expectations, we conducted several supplementary analyses in which we regressed ratings of quality on country income and the Healthcare Access and Quality (HAQ) index of health system performance [34]. All regressions used unweighted data. Data analysis was completed in Stata/SE version 14.2 (StataCorp, College Station, Texas). This study (protocol number IRB17-0907) was reviewed and determined to be exempt by the Harvard University Human Research Protection Program.
Of the 57,786 respondents who opted to take the survey, 17,996 respondents (31%) completed the survey questions for this analysis; this formed our analytic sample (Table 1). The completion rate is similar to rates reported for similar studies [35–38]. After weighting the data with population census weights, approximately half of respondents (51%) were male, 35% were between the ages of 18 and 29, 39% were between the ages of 30 and 49, and 25% were over the age of 50. Rural residents made up 41% of the sample. Nearly half of respondents (45%) had a primary education or less. The vast majority (73%) reported good general health with a mean number of outpatient healthcare visits over the last year of 2.5 (SD 3.0). One-third of all respondents (34%) reported that they experienced discrimination in the healthcare system in the past.
Over half (53%) of respondents across countries rated the quality of care described in vignette A (blood pressure visit; poor technical quality) as good or better (Fig 2). This rate was similar when varying the health condition; 55% rated vignette B (accident visit; poor technical quality) as good or better. In the variations of the 2 vignettes in which interpersonal quality is poor, 54% (C. high blood pressure; poor interpersonal quality) and 57% (D. accident: poor interpersonal quality) gave a rating of good or better. Respondents from Senegal were most likely to rate vignette A (blood pressure visit; poor technical) as good or better.
Men had lower expectations than women in our sample (Table 2). The following variables increased the odds of rating vignette A(blood pressure visit; poor technical quality) as good or better: male gender (adjusted odds ratio [AOR] 1.32, 95% CI 1.23–1.41, P < 0.001), no formal schooling (AOR 2.22, 95% CI 1.90–2.59, P < 0.001), excellent self-reported health (AOR 5.19, 95% CI 4.33–6.21, P < 0.001), and history of discrimination in healthcare (AOR 1.47, 95% CI 1.36–1.57, P < 0.001). There were small differences in these predictors across the 4 vignettes. Results of the multilevel models confirmed significant differences across countries. Supplementary analyses exploring effects of country income or health system performance on expectations did not suggest a clear pattern. We believe that heterogeneity of preferences across populations is influenced by a large range of factors, including prevailing models of care, utilization patterns, media, and political factors. The supplementary analysis, which included self-reported mental health in our multilevel model for vignette A, found that those with excellent self-reported mental health had nearly 4 times higher odds of having low expectations of quality (S11 Appendix).
Responses from nearly 18,000 internet users in 12 LMICs show that good ratings for poor quality are common: over 50% of respondents indicated that objectively poor quality of care described in vignettes was good, very good, or excellent. Vignettes highlighting poor technical quality and poor interpersonal quality yielded similar results, supporting the hypothesis that low expectations, not lack of technical knowledge, drives these ratings. The prevalence of good ratings is especially notable given that internet users are likely to be more affluent and educated than the general population and thus more likely to access better quality care and have higher expectations of quality [40,41]. Our finding that a majority of people in the study countries have low expectations of healthcare quality points to a lost opportunity to keep health systems accountable for the quality of care that they deliver. This work may help explain current high satisfaction ratings and inform efforts to better measure people’s assessment of health system performance.
We found significant variation in expectations across countries both in our analytic models and in unadjusted comparisons. Greater differences between countries were noted for the vignette describing poor interpersonal quality, perhaps because of the more subjective and socio-culturally influenced nature of interpersonal quality of care. Health system factors probably play an important role in this variation. For example, of all country respondents, the Senegalese were most likely to rate both hypertension vignettes as good or better; their ratings of both accident vignettes were below the full sample average (Fig 2). Healthcare utilization for injury is higher than that for (diagnosed) hypertension in Senegal [42,43]. Could the Senegalese pattern be due to respondent familiarity with the 2 conditions? High ratings for poor quality were also frequent in India but in only 3 of 4 vignettes without a clear explanation for the pattern—Indian respondents appear to be most sensitive to poor technical quality during a visit for an injury. Relatively low health system investment and documented poor quality of care in both public and private sectors in India would have predicted chronic exposure to poor-quality care and lower expectations across all 4 vignettes [44–46]. Moroccan respondents consistently had some of the lowest ratings for the poor-quality vignettes. This may be driven by remarkable recent improvements in health outcomes in Morocco [47,48]. However, Kenyan respondents also had lower ratings despite less progress on achieving broad-based health gains [49]. Higher expectations in Kenya may have been shaped by recent countrywide healthcare strikes [50]. A complex interplay between political, social, economic, and health system factors is likely to explain these finding, and national differences require further research.
Male gender, low education, experience of discrimination in the health system, and good health status were associated with low expectations of care in adjusted models for all versions of the vignette. Respondents with no primary education were 2 times as likely to have low expectations than their more educated peers. This is consistent with existing literature on expectations and quality ratings [22,51]. Satisfaction with maternity services and quality ratings, for example, are lower in respondents with higher educational levels in various LMIC settings [52–54], whereas expectations of good patient-doctor communication are higher [22,51]. Low education both represents an information deficit and is likely to be linked to low socioeconomic status. Respondents with low social status may have lower expectations because of chronic and repeated exposure to low-quality services as well as less access to accurate information about the health system [2,55]. Low expectations may also stem from a more general experience of disempowerment and lack of entitlement to other government services [56].
We found that women had higher expectations of quality than men. Their higher expectations may be due to frequency and nature of interactions with the health system: women are often the primary caregivers in families and more likely to interact with the healthcare system for maternal and child health services [57,58]. Women may place a higher value on healthcare because they are often responsible for sick children, particularly vulnerable newborns, and personally experience the potentially life-threatening event of childbirth. Though our model included the number of outpatient visits in the past year, we did not have information on the nature of care received, and it is possible that this was systematically different for men and women. Another possible explanation is that women have better access to information because they are frequently targeted for health education. The lack of association with age is consistent with a study of pre- and post-visit expectations in the United Kingdom that showed that older adults did not have lower expectations of care but that they were more likely to believe that their expectations had been met [11]. This last point is supported by relative agreement in the literature on the role of age on patient satisfaction; satisfaction increases with age [31].
Regarding experience with the health system, more outpatient visits in the last year is associated with good ratings for poor-quality care, but the increased odds are small. Without knowing the quality of those visits, it is difficult to interpret this result. However, there is a growing body of evidence that describes substantial deficits in the quality of care in our study countries, making it likely that additional experience with healthcare is experience with poor-quality healthcare [2,59–61]. More directly, our results showed that a history of discrimination raised the likelihood of good ratings for poor quality by nearly 50% for the main vignette, suggesting that a history of poor-quality care is associated with lower expectations. This is consistent with the literature showing that a person’s experience with healthcare is the most important source of information about quality and a strong driver of perceptions of that care [54,62]. It is possible that experience of poor-quality care may stifle the demand for high-quality care and create a vicious cycle of low expectations and poor-quality care (Fig 3). Conversely, improvements in quality of care or raising people’s expectations may break this cycle, possibilities that require further study.
Better self-reported health status was strongly associated with low expectations of care. This is consistent with the hypothesis that health status would be inversely related to expectations of care. Healthier people need less from the health system and may have a more positive outlook. In contrast, sicker patients may place a higher value on healthcare, making them less tolerant of poor quality. Research from a variety of different countries has shown that patients who actively navigate the health system are more likely to have a serious health condition [15,16,62]. In our subanalysis including self-reported mental health status, the variable was strongly associated with low expectations, suggesting that the role of individual outlook (overall positivity or negativity) in shaping expectations warrants further assessment.
Further research is needed to fully understand how expectations of quality can be increased and whether these higher expectations can contribute to health system quality improvement in LMICs. Studies of expectations focusing on equity and specifically targeting marginalized groups will be especially important for understanding the role of disempowerment on expectations of care.
The report of the Lancet Global Health Commission on High Quality Health Systems recommends that governments consider “igniting demand” for high-quality care as 1 of 4 universal actions to improve quality at scale [2]. Understanding people’s expectations of quality and how to raise these expectations will be important in raising demand for quality. We were unable to explore the role that information about the health system—i.e., educating people about their entitlements or information about what good care consists of—plays in shaping expectations. From an improvement perspective, information about the health system is potentially the most malleable of the factors influencing expectations. It has also been shown to drive decisions about care seeking [63]. Information as a lever for raising expectations of quality may be especially important in poorly functioning health systems, in which experience with poor quality is likely to be prevalent and to lower expectations.
Evidence for specific interventions that raise healthcare expectations is still sparse [2]. Several intervention types show the potential to raise expectations and suggest that higher expectations through information sharing are involved in improving quality. For example, participatory women’s groups are associated with better provider practices during childbirth, partially because women are learning to expect proper hygienic care and demanding it [64]. Consumer quality reporting, patients-rights charters, and mass media campaigns are also promising approaches that warrant further exploration. Increasing quality and transparency of information is an especially timely area for further research and potential action because of the recent growth in mechanisms for information sharing and patient involvement created by the digital revolution. People are increasingly using the internet and mobile phones to share and receive information on health and healthcare quality [65,66]. Program planners are also leveraging the widespread use of mobile technology to engage people [67].
This study has several limitations. Internet surveys in countries with generally low internet penetration are not representative of the full population because internet users are likely to be male, wealthy, young, and more urban than the general population [36]. Based on our conceptual framework, we believe that our sample is likely to have higher expectations than the general population in our survey countries. Population weights have been applied to our descriptive statistics, but this cannot compensate for the absence of entire demographic groups, and thus our inference is limited to the internet-using population. Internet surveys are also known to have lower response rates than face-to-face surveys, leading to concerns of nonresponse bias [37]. We addressed this by surveying a large sample, limiting the length of the questions, and structuring the survey for ease of response. On the other hand, internet surveys are useful for exploring sensitive topics due to low social desirability and acquiescence biases, which may lead to more honest responses about healthcare [39]. We were unable to directly assess respondent wealth, which is likely to play a role in expectations of quality that may be distinct from education. Additionally, social, community, and healthcare context was only approximated in our model by country, and we were unable to assess the respondents’ access to information about quality healthcare. Finally, concerns about cognitive overload for the respondent meant that we elected not to include a “control” vignette describing only high-quality care.
Our findings suggest that expectations of quality are low and that caution is needed when interpreting satisfaction and other health system ratings because they are likely to be biased upward. Vignettes such as ours that establish people’s expectations, values, and preferences for healthcare can be used as anchoring vignettes to rescale satisfaction ratings. Anchoring vignettes allow researchers to control for people’s internal standards of quality and make more accurate comparisons and interpretations of satisfaction and quality ratings [20,21]. These vignettes can also help policy makers accurately gauge the impact of new policies and interventions on the quality of care in their health systems.
To our knowledge, this is the first multicountry study of expectations of healthcare quality in LMICs. Our results show that people’s ratings of poor quality are remarkably high in LMIC settings, with over 50% of respondents rating vignettes describing poor-quality care as good or better. This points to an opportunity for future efforts to improve health systems. Increasing expectations of good care from the public should exert much needed pressure on health systems to provide competent and respectful care. Raising expectations should be part of the broader health system improvement agenda as countries adopt universal health coverage.
|
10.1371/journal.pcbi.1004242 | Cell-Specific Cardiac Electrophysiology Models | The traditional cardiac model-building paradigm involves constructing a composite model using data collected from many cells. Equations are derived for each relevant cellular component (e.g., ion channel, exchanger) independently. After the equations for all components are combined to form the composite model, a subset of parameters is tuned, often arbitrarily and by hand, until the model output matches a target objective, such as an action potential. Unfortunately, such models often fail to accurately simulate behavior that is dynamically dissimilar (e.g., arrhythmia) to the simple target objective to which the model was fit. In this study, we develop a new approach in which data are collected via a series of complex electrophysiology protocols from single cardiac myocytes and then used to tune model parameters via a parallel fitting method known as a genetic algorithm (GA). The dynamical complexity of the electrophysiological data, which can only be fit by an automated method such as a GA, leads to more accurately parameterized models that can simulate rich cardiac dynamics. The feasibility of the method is first validated computationally, after which it is used to develop models of isolated guinea pig ventricular myocytes that simulate the electrophysiological dynamics significantly better than does a standard guinea pig model. In addition to improving model fidelity generally, this approach can be used to generate a cell-specific model. By so doing, the approach may be useful in applications ranging from studying the implications of cell-to-cell variability to the prediction of intersubject differences in response to pharmacological treatment.
| Mathematical models of cardiac cell electrophysiology are widely used as predictive and illuminatory tools, but have been developed for decades using a suboptimal process. The models are typically constructed by manual adjustment of parameters to fit simple data and therefore often underperform when used to predict complex behavior such as arrhythmias. We present a novel method of model parameterization using automated optimization and dynamically rich fitting data and then demonstrate that this approach is better at finding the “real” model of a cell. Application of the method to cardiac myocytes leads to cell-specific models, which may enable well-controlled studies of both cellular- and subject-level population heterogeneity in disease propensity and response to therapies.
| Mathematical models of cardiac electrophysiology trace their roots to Hodgkin and Huxley’s seminal work from 1952 [1]. Since then many models have been developed describing cardiac electrophysiology in a number of species and cell types helping to make modeling an integral part of cardiac research [2–5].
The typical method for model development and parameterization is a bottom-up approach. Individual ionic membrane currents are characterized using voltage-clamp experiments from which mathematical equations are derived [6,7]. Although it has led to many advances, this traditional approach to model development has several limitations, including:
Several studies have looked into how to improve model parameterization. Approaches in cardiac myocyte modeling have included the parameterization of individual channel dynamics, typically when making more complex Markov models of ionic currents [6,15–17]. Whole-cell optimization approaches have focused on generating models that can match action potentials from different types of cardiomyocytes, using both simple models consisting of a few generic currents [13,18–20] and more physiologically detailed ionic models [21–25]. A resulting synthesis is that optimization results are improved when models are fit to data beyond a single action potential, e.g., action potentials from multiple pacing rates [13,19,22] or voltage waveforms during varying current injection [20,21,24]. In particular, using a global search heuristic applied to an ionic model, Syed et al. demonstrated that it is feasible to estimate conductance parameters for experimental data and showed that the fits improved when using data recorded during multiple periodic pacing frequencies [22]. Sarkar and Sobie presented a much simpler, but more approximate, linear regression strategy to estimate model conductances based on biomarkers from simulated model output and have used it to investigate how specific conductances relate to particular model outputs [26]. In neuroscience, considerably more research has been carried out on parameter estimation problems (e.g., [27–30]) and a few studies have developed protocols that allow parameterization of cell-specific models [31,32]. However, these protocols are not directly applicable to cardiac myocytes, due to intrinsic differences in electrophysiological behavior between neurons and cardiomyocytes.
Here, we present a novel strategy to develop cardiac models by pairing dynamically rich electrophysiology protocols with powerful computational parameter fitting methods. We first developed novel electrophysiology protocols that probe the dynamics of a subset of ionic currents in an intact cardiac myocyte without ion channel inhibitors, agonists, or unphysiological ion concentrations. The protocol consists of (1) stochastic current-clamp stimulation and (2) multi-step voltage clamping. As will be discussed, stochastic stimulation represents a quick method to sample the rate-dependent cardiac dynamics, while the multi-step voltage-clamp protocol is designed to highlight individual currents using a tailored sequence of holding potentials. Based on the assumption that ion-channel kinetics are preserved among (healthy) subjects while maximal conductances vary as a result of differences in expression levels, the resulting data are used to estimate maximal conductance values of several ionic currents and maximal flux of calcium ion transporters in the model. Because of the complexity of the data, hand parameter tuning is not feasible; thus, we utilized a genetic algorithm (GA), which is an efficient method for such complex optimization applications [33]. The approach was first developed and validated computationally. It was then used to develop cell-specific models of isolated guinea pig ventricular myocytes that simulate the electrophysiological dynamics significantly better than does a standard guinea pig model.
We first developed our parameter estimation strategy using a guinea pig ventricular myocyte model (Faber and Rudy [34], the “FR” model) and tested the ability of the optimization procedure to return the original parameter values. Traditionally, one of the main criteria for cardiac electrophysiology model quality is the ability of a model to describe the cardiac action potential. Therefore, we first ran the parameter estimation using a single FR model action potential as the target objective. Nine model parameters, describing maximal conductances of ionic currents [the sodium current (INa), the L-type calcium current (ICaL), the T-type calcium current (ICaT), the inwardly rectifying potassium (IK1), the rapid and slow delayed rectifier potassium currents (IKr and IKs), the plateau potassium current (IKp), and the sarcolemmal calcium pump current (IpCa)] and the maximal flux of the sarcoplasmic reticulum Ca2+-ATPase (JSERCA) were estimated using a GA technique.
A GA optimization is initialized by a population of models with different parameter values. We used a population size of 500 model instantiations generated by randomly drawing values for the nine parameters from a range of 0.01–299% of the published value. The GA methodology uses ideas from evolution [23,33]. In GA terminology, the initial state is referred to as generation 0 and the 500 models as individuals. Fig 1 A–1 C shows three different individuals in generation 0. Most of these generate action potentials that are very different from the target action potential (Fig 1A and 1C). However, by chance, a few of the 500 individuals provide reasonably good fits to the action potential, even though the parameter values are very different from those of the baseline model (parameter scaling of 1, Fig 1B).
The optimization proceeds in generations (steps), for which the GA applies crossover (parameter swapping), mutation (parameter variation), and selection (discarding poorly performing models) to increase the fitness of the population (reduce the error between model output and target objective). We ran the GA for 100 generations, as this was sufficient for the error of both the population average and the best individual to reach a minimum (Fig 1 F and 1G). Although the optimized model action potential matches the optimization objective to a very high degree, the estimated parameter set does not match that of the FR model (Fig 1E). This is consistent with previous results showing that if only a single action potential is used for parameterization, cardiac models may be overdetermined and more than a single set of parameter values can describe that action potential [20,21,26].
The duration of the action potential, and to a lesser extent its morphology, varies with stimulation interval and history. Models tuned to single action potentials during periodic pacing (as in Fig 1) would not be expected to accurately reproduce such dynamics. A more complete method to probe cellular behavior is the restitution portrait [35], which is a systematic, but prohibitively long, mapping of this rate dependence. An alternative approach that would significantly reduce the protocol duration, while maintaining some dynamic information, is random sampling of the rate dependence. To accomplish this, we utilized a stochastic stimulation protocol containing 11 randomly timed stimuli delivered over 5 s. When applied to the FR model, the stochastic stimulation results in considerable action potential variability (Fig 2 A and 2B).
We used this stochastic stimulation protocol and resulting voltage response as an optimization sequence to test the extent to which dynamic stimulus timing would improve the parameter estimation. Because the GA parameter estimation is a stochastic method, it was run 10 times with 10 different initial populations. For each run, we selected the best individual, i.e., the model instantiation with the best fit to the target voltage trace. All 10 best individuals matched the voltage trace very closely (Fig 2 A and 2B show the best one) as was the case for the single action potential fitting. Compared to the single action potential, the stochastic stimulation leads to a modest overall improvement of the parameter estimation, but it did notably better in determining the maximal conductances of IKr, ICaL, and IKs (Fig 2C). However, as shown in Fig 2C, a few parameters remain incorrectly estimated (maximal conductances of ICaL and IKs), and some are estimated with a large spread (maximal conductance values of IKp, IpCa, ICaT, and IKr). Sensitivity and correlation analyses illuminate why these parameters are less well determined—they have low sensitivity (in which case they have minimal effect on the fitting objective, making them difficult to probe) and/or they have inter-correlations (in which case two or more parameters lack independent contributions to the fitting objective and are therefore difficult to discriminate) (S1–S4 Figs and S1 Text).
To measure the predictive capabilities of the optimized models, we presented the 10 best individuals from the GA runs with a novel stochastic stimulation sequence (S6 Fig). When subjected to this new prediction sequence, both the individuals trained with the stochastic stimulation sequence and those trained to the single action potential matched the baseline FR model response well (S6 Fig), but the stochastic stimulation protocol lead to the better match, as shown by the smaller prediction error (error between optimized model and target during the prediction sequence) in Fig 2D. Thus, although the parameter recovery seem only modestly improved for the stochastic stimulation compared to the single action potential target, the stochastic stimulation outperforms the single action potential in that it results in models that are significantly better at predicting the response to a novel set of stimuli.
To improve parameter estimation accuracy, more improvement is typically gained from adding measurements of a different state variable than adding additional measurements of the same state variable [36]. This suggests that a longer stochastic stimulation protocol is unlikely to yield much improvement. This idea is in line with our finding above (Figs 2D and S6) that models optimized to the 5s stochastic stimulation protocol matched well when subjected to a novel stochastic stimulation sequence.
Therefore, to improve the parameter estimation accuracy, we added a multi-step voltage clamp protocol to the objective function. Traditionally, voltage clamp is applied to a cell as a set of holding potentials varied systematically either in its timing or its potential to characterize one particular current. We developed a voltage clamp protocol consisting of a sequence of holding potentials, with each step designed to emphasize specific currents relative to the others (Fig 3). The rationale is that if a particular conductance contributes most of the total current for a particular holding potential, then only models fit with a correct value of that conductance will reproduce the current target for that phase of the protocol and therefore have a low fitting error. Our 6s long voltage clamp protocol effectively isolates IK1, ICaL, and IKs as shown by the disproportionally large contributions of these currents in step -120 mV, +20 mV, and +40 and -30 mV, respectively (Fig 3 B–3F). Therefore, we hypothesize that this protocol will directly improve the conductance estimation for these currents.
Indeed, using the voltage clamp protocol as the objective during an optimization recovers the conductances for IK1 and IKs very accurately (Fig 4B). The estimation of the ICaL conductance is very close to 1, but is slightly overestimated in all runs. Less predictively, IpCa and IKp were also estimated more precisely than during stochastic pacing alone (Fig 4B). However, a few conductances were estimated poorly (in particular JSERCA and ICaT) and models optimized based on voltage clamp data alone were, not surprisingly, inferior at predicting complex action potential dynamics during stochastic pacing (Fig 4C).
The extension of the target objective to include the multi-step voltage clamp protocol results in a joined unitless error function (Eq 3 in the Methods). As both the stochastic pacing recording and the voltage clamp data is fit increasingly well during optimization, the error contribution from each sequence decreases (Fig 4A, left). Although the main contribution to the total error comes from the stochastic pacing segment, the error from voltage clamp segment drops more during the optimization process, suggesting that both protocols help constrain the parameters.
Running the optimization with the combined objective does indeed lead to improved accuracy of the parameter estimation, with all nine current parameters being recovered to within one standard deviation (orange symbols, Fig 4B). For six of the nine current parameters, the combined protocol results in parameters whose mean estimates are closer to 1 and/or have less variational spread than either of the individual protocols alone (Fig 4B). Only currents that were estimated very accurately by the voltage clamp protocol alone (IKs, IK1, and IpCa) did not show improvement with the combined protocol.
For some of the currents, one protocol segment is clearly better than the other in terms of parameter recovery (e.g., stochastic pacing for INa and voltage clamp for IKs). However, for other currents, in which both individual protocol segments result in off-target outcomes, the combined protocol produces estimates spanning 1 (e.g., IKp). Such improvement is consistent with the combined protocol restraining parameter space and avoiding local minima.
Again, we tested the ability of the 10 best individuals to predict the response to a stochastic stimulation sequence to which they were not fit. The prediction error of the 10 individuals from the combined objective function runs was significantly lower than the error for the individuals that were estimated using only the stochastic stimulation protocol (Fig 4C). Hence, the combined stochastic pacing and voltage clamp protocol improves both parameter recovery and prediction performance.
To further improve the estimation results, the results of the first 10 GA runs were used as the new parameter bounds for a second set of runs (see Methods), e.g. 0.01% to 299% changed to 92.8–114.9% for INa. Note that this method only works when the fits of the first 10 runs span the correct solution as is the case with the combined protocol. During this second, local, iteration, better fits are generated causing the error for both the voltage clamp and the current segments, as well as the total error for the best individual, to drop (Fig 4A, right). Thus, using this iterative approach, the error bounds around the estimated parameter values decreased (magenta symbols, Fig 4B) and the prediction error reduced markedly (Fig 4C).
In summary, the combined protocol, consisting of stochastic stimulation and multi-step voltage clamp, allows accurate parallel estimation of eight maximal conductance values and maximal pump rate of SERCA for the FR guinea pig ventricular model. Such validation simulations laid the groundwork for using the method to fit computational models to real cardiac cell data.
The parameter estimation method was next applied to four guinea pig left basal ventricular myocytes from four different animals. Each cell was subjected to the stochastic stimulation protocol in current clamp mode, followed by the multi-step voltage clamp protocol, using the perforated patch clamp technique.
All four myocytes exhibited action potentials and membrane current responses that were very different from the baseline FR model (Fig 5 shows output from one cell, S7–S9 Figs presents the results from the remaining three cells). In particular, their action potentials were substantially longer than those of the FR model and their current response to prolonged depolarization was substantially smaller.
For each cell, the GA estimate from the experimental data fit much better than did the FR model (Fig 5 and S7–S9 Figs). In particular, the optimization leads to very accurate voltage dynamics, which is important for arrhythmogenesis prediction. The total current is fit less well, potentially due to mismatch in ion channel kinetics (see Discussion). Overall, the optimization results in more accurate predictions, with the prediction error being an order of magnitude lower for the fitted models than for the FR model (Fig 5C and 5D).
The dissimilarities between the original FR model and the experimental data led to considerable changes in the estimated values for the model parameters for all four cells (Fig 6). Interestingly, these changes were qualitatively similar between all four myocytes for most of the parameters, indicating conserved differences between our experimental data and the FR model. In particular, IKs and IKp are scaled down significantly and JSERCA is slightly reduced. In contrast, for all four cells, maximal conductance of IKr and IK1 are increased around 2-fold compared to the FR model, while ICaL is slightly increased. The results for INa show variation among cells, with a significant increase for three out of four cells and a small decrease for one cell.
In summary, the optimized models show a much closer match to the experimental data as reflected in the individual voltage and current traces as well as in the prediction error. In addition, the optimization identified similar trends in the underlying channel conductance values for different cells from a particular region in the heart. Considered together with the demonstration that the approach accurately identifies model parameters (Figs 2–4), these findings suggest that the approach significantly improves the fidelity of the model for cellular data, relative to the published generic model.
To overcome the limitations inherent to traditional cardiac model construction (most notably manual parameter adjustment and use of averaged, dynamically limited data), we developed a novel approach for parameter estimation that combines an electrophysiology protocol that is rich in dynamic information, short in duration, experimentally feasible, and does not require the use of drugs or special solutions, with a parallel computational parameter tuning algorithm (GA). The protocol was first tested computationally, which showed reliable parameter estimation. We then applied the protocol in vitro to guinea pig basal left ventricular myocytes. Compared to the baseline guinea pig FR model, the optimized cell-specific models showed a significantly improved fit to the experimental data. Furthermore, because our model enables validation on data from the same cell for which a model was optimized, we were able to demonstrate that the cell-specific models are markedly better at predicting the response to novel stimulation sequences than was the generic model.
In cardiac modeling, a single action potential or biomarkers derived from it such as amplitude and duration, is often used as a minimal objective for model parameterization. Ionic models can be optimized to fit single action potentials using, e.g., global search heuristics [22,23], but because the optimization problem is overdetermined, fits may be improved when adding more data, such as data recorded at multiple pacing rates [19,22]. In fact, relative to a single action potential, more complex driving protocols have the potential to dramatically improve parameterization by creating target objectives that are richer in information. On the other hand, to be experimentally feasible, protocols have to be relatively short in duration due to the inevitable current rundown that occurs in patched myocytes, even when using perforated patch. As a compromise, we utilized a stochastic stimulation protocol because it rapidly samples the rate-dependence of the action potential. In addition, irregular excitation patterns are present in many cardiac arrhythmia; thus models tuned to aperiodic excitation patterns are inherently better suited for modeling irregular arrhythmia. In our simulations, we found that the estimates for IKr and IKs were improved the most by the stochastic stimulation objective. During a single action potential, IKr and IKs have similar and compensatory effects (S4 Fig), which impedes estimation of their conductances. In contrast, stochastic stimulation more thoroughly explores their kinetics, thereby revealing small differences throughout the protocol, resulting in a more accurate estimation. Although the stochastic stimulation protocol led to at most a modest improvement in the parameter estimation for the remaining parameters, the prediction error was reduced by an order of magnitude, compared to using a single action potential (Fig 2). Thus, significant model improvement is obtained through the use of a dynamically rich objective, as this helps the optimization avoid the false alternatives that can appear to fit well when dynamically sparse data are used for fitting.
In addition to such current-clamp experiments, currents recorded during voltage clamping add additional data to improve fitting and optimization [21,31,32]. While our multi-step voltage clamp protocol alone is very useful for estimating many of the parameter values (Fig 4B), it tends to generate models that fail to predict novel stochastic pacing data well (Fig 4C), which is unsurprising given that it does not train the models according to membrane potential. In our simulations, the addition of the multi-step voltage clamp objective to the stochastic current-clamp stimulation objective enhanced the quality of the parameter estimation compared to using only stochastic stimulation (Fig 4B). This improvement was the result of: (i) some parameters being estimated accurately by the voltage-clamp protocol and (ii) information on two, rather than a single, state variable putting more constraints on the parameter values [36]. In particular, estimates for all nine parameters became centered on their baseline values and the prediction error dropped by another order of magnitude relative to that of stochastic pacing alone. Finally, the iterative optimization approach [31] refined our in silico parameter estimation by decreasing the spread of the returned parameter sets, which caused the prediction error to again decrease by an order of magnitude.
Generic models have the advantage that direct comparisons can be made among different simulation studies. However, when comparing a generic model such as the out-of-the-box FR model to our experimental data, there are substantial differences, which likely would cause inaccurate predictions if simulating, e.g., effects of pharmacological agents or genetic variations. For one, there are clear distinctions in action potential morphology, e.g., in the plateau phase (Fig 5). This difference in plateau phase most likely explains the method’s downscaling of the IKp conductance. Our recorded action potentials are also of considerably longer duration, which is consistent with the finding of a much reduced IKs in the voltage-clamp experiments. The step to -120 mV in the voltage-clamp protocol induced a much larger current in the experiments than in the FR model and IK1 conductance was increased accordingly in all four cells. These consistent changes in voltage traces and currents between our cells and the FR model may be due to lab-to-lab variability and to the fact that the FR model is not region-specific.
Despite such consistent changes, the parameterization also points to important cell-to-cell variability, in particular for the INa conductance, which is increased in three cells and decreased in one. In neuronal modeling, it has become clear that different combinations of conductance parameter sets can give rise to the same activity pattern and that using average values of the conductances may fail to generate that pattern [10,11,37]. The differences in cell-to-cell variation in current densities have been linked to mRNA expression differences or post-translational modifications [38,39]. The extent to which such variation occurs in healthy cardiomyocytes remains to be seen, but some examples of functional coupling among ionic currents in perturbed systems have been described [40–42]. This failure-of-averaging concept may also extend to cardiac tissue: although intrinsic cellular heterogeneity tends to be smoothed out when myocytes are electrically coupled, coupled cells do not necessarily behave like their average. For example, a myocyte with intrinsically shorter action potential duration may promote repolarization in a cell pair [43]. Also, a range of synchronization patterns have been described in coupled pacemaker cells [44]. Thus, there may be important utility to developing cell-specific models.
Indeed, cardiac cell-specific models have a range of potential application areas. First, the models can obviously be used to study cell-to-cell variability [45]. Second, in the clinic, inter-subject variability can lead to response differences among patients to pharmaceutical treatment. A dramatic example of this variation is the response to IKr-block, which can vary from minor changes in the electrocardiogram to ventricular tachyarrhythmias [46]. Understanding and predicting this variability is an important step towards patient-specific treatment. In turn, model optimizations such as those developed here represent an advancement towards patient-specific prediction. Finally, multiple models could be grouped into a heterogeneous population and used to generate more realistic responses than those of a randomly-generated population [47,48].
The developed protocols allow accurate estimation of nine conductance/flux parameters. To characterize a single cell more thoroughly, additional flux parameters could be included (e.g., those describing the sodium/calcium exchanger and the sodium/potassium pump), but as inclusion of more parameters makes the optimization problem harder, this may necessitate tweaking of the methods described here. As detailed below, possible strategies for improvement of the parameter estimations are: 1) improving the stochastic stimulation and voltage-clamp protocols; 2) adding measurements of different state variables during the same protocols (e.g., intracellular calcium or membrane resistance); 3) incorporating altered solutions and/or ion channel blockers to improve isolation of individual currents [49]; 4) including ion channel kinetic parameters in the optimization; or 5) including relative weights for the current and the voltage contributions to the summed error.
Our multi-step voltage-clamp protocol effectively isolates IKs, ICaL, and IK1. An improved voltage-clamp sequence that isolates the remaining currents could improve estimation of their conductance/flux parameters. We designed the voltage clamp steps based on a priori knowledge of the current-voltage (IV) relations in guinea pig ventricular myocytes. As a way to design better protocols, an automated optimization approach may be feasible, i.e., an optimization of the optimization protocol.
Further, differences in structure, channel kinetics and IV-relationships between model and experiment are likely to result in less accurate parameter estimations [31] and may underlie the deviations between fit and experimental data during voltage clamp (Fig 5 and S7–S9 Figs). Adding parameters describing ion channel kinetics to the optimization process would likely improve the fits and predictability, but would almost certainly necessitate longer voltage clamp protocols [6,16]. As channel kinetics are not expected to vary substantially among cells of the same type, a possible strategy is to first parameterize average channel kinetics in a cell population, then apply our method to derive cell specific models.
Additionally, improvement could likely be gained by simultaneously recording calcium fluorescence and adding that to the objective function [36], a strategy with merits illustrated by Fig 1 of Ref. [26]. As expected, local sensitivity analysis on simulated calcium traces demonstrates that they are most sensitive to changes in ICaL, IpCa, and JSERCA (S5 Fig), which leads to the speculation that the estimation of these parameters could improve. Inclusion of calcium data may also allow determination of INaCa, which depends on and influences both intracellular calcium and transmembrane potential. Incorporation of membrane resistance in the objective function would also be expected to improve the fitting, as shown in recent work by Kaur et al. [25]. A potential caveat in such multi-objective optimization is that simultaneous good fits are not always achievable, necessitating trade-offs between the different objectives. In that case, balancing which objective(s) to prioritize would be application dependent.
Although we allow a generous range for the conductance parameters (0.01–299% of baseline), some parameters did reach the bounds when fitting the experimental data (Fig 6). Increasing the range will likely require running the GA optimization with a larger population size or for more generations, as will including additional parameters. The main computational cost of the GA is that of simulating the individual models. As this process is inherently parallel, it is straightforward to take advantage of parallel computing. Future implementations could decrease run time by utilizing a GPU, on which optimization for neuronal data has been shown to be feasible [32].
Finally, although the four cells tested in this study provide a strong proof-of-concept for the approach, to further develop the method, it could be applied to a larger number of cells.
In the novel approach developed here, cell-specific cardiac models are developed by coupling complex electrophysiology protocols with genetic algorithm parameter fitting. Neither the electrophysiological data (which are too complex to fit by hand), nor the fitting algorithm, would offer much advantage alone. However, merging the two enables markedly improved models that can more accurately simulate dynamically rich cardiac dynamics than can models developed using traditional approaches. Given the widespread use of ionic cardiomyocyte models in investigating arrhythmogenesis, there is utility in models that are better at reproducing such rich electrophysiological dynamics, which are more representative of the complex dynamics that are often inherent to arrhythmias. In addition to improving model fidelity generally, because this approach can be used to generate a model from a single cardiac myocyte, it may be useful in applications ranging from studying the implications of cell-to-cell variability to the prediction of intersubject differences in response to pharmacological treatment.
All animal care and handling for 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 was approved by the Institutional Animal Care and Use Committee of Weill Cornell Medical College (protocol number: 0701-571A).
The cardiac guinea pig model developed by Faber and Rudy ("FR model") [34] was used. Model intracellular and extracellular ionic concentrations were set to the values used in our in vitro experiments (see below), after which the model was simulated to a steady state in current clamp mode for 1800 beats at a pacing cycle length of 500 ms. Stimuli were square pulses of 1 ms duration and -40 A/F amplitude.
Stochastic stimulation sequences were 5 s in duration. Stimulation times were randomly drawn from a uniform distribution with a range of 100–700 ms. Stimulation times for the optimization sequence were: 216.48, 623.36, 764.64, 1101.12, 1790.16, 2073.10, 2642.28, 3183.10, 3786.07, 3959.02, and 4579.72 ms (Fig 2A). Prediction sequence stimulation times were: 247.40, 705.30, 1020.60, 1347.90, 1462.60, 1705.60, 2038.30, 2546.70, 3150.30, 3706.70, 3939.70, 4077.80, and 4645.50 ms (S6 Fig). After each parameter change during the GA optimization (details below), current clamp was simulated for 9 beats at a static pacing interval of 500 ms to dampen transients.
The voltage clamp protocol was designed on general, a priori, IV relations for the individual channels (e.g., IK1 is the predominately active current at a holding potential of -120 mV). The 6000 ms protocol was composed of the following steps: 50 ms at -80 mV, 50 ms at -120 mV, 500 ms at -57 mV, 25 ms at -40 mV, 75 ms at +20 mV, 25 ms at -80 mV, 250 ms at +40 mV, 1900 ms at -30 mV, 750 ms at +40 mV, 1725 ms at -30 mV and 650 ms at -80 mV (Fig 3). The contribution of individual membrane currents to the voltage clamp protocol was evaluated using the FR model and the following equation:
Contribution=100%⋅∑t=t1t2|Ix(t)|∑j=1N∑t=t1t2|Ij(t)|
(1)
Eq 1 calculates the percentage contribution of the absolute individual current (Ix) relative to the absolute sum of all N currents for all time points during one of the holding potentials (t1 to t2). This calculation was done for all model currents at all holding potentials.
During GA optimization, the simulated multi-step voltage clamp is preceded by 5 s holding at -80 mV to allow the model to settle after parameter changes.
Multiple global search heuristics have been applied to electrophysiology models, including gradient-based descent [13,19–21], simulated annealing [50], particle swarms [18,24], and genetic algorithms [22,23,25]. We chose a genetic algorithm as it is effective for a range of the number of parameters [50], is computationally simple and readily parallelizable, and has been shown to be successful at optimizing sophisticated ionic models to experimental data [22,23,25].
The GA used in this study was originally developed by Sastry [33] and was used with the settings for selection, crossover, mutation, and elitism strategy as in Bot et al. [23]. Compared to the study of Bot et al., we increased the parameter search range to 0.01–299% of the baseline model values. This larger range, in combination with the increased number of parameters and the diminished requirements for computation speed relative to the Bot et al. study, caused us to enlarge the population size to 500 and raise the number of generations to 100, based on test runs showing consistent convergence when using these values. Because the GA is inherently stochastic, it was run 10 times per optimization problem. In addition, an iterative approach was implemented, based on the study of Hobbs and Hooper [31]. For each parameter, the iterative approach uses the span of the 10 best individuals from the first 10 GA runs as the search boundaries for a second set of GA runs. With these new search ranges, the GA was again run 10 times with a population of 500 individuals for 100 generations.
We used mean squared differences for the objective functions (errors) that the GA works to minimize:
E1=∑t=tIC,starttIC,end(Vtarget(t)−Vindividual(t))2
(2)
E2=∑t=tIC,starttIC,end(Vtarget(t)−Vindividual(t))2+∑t=tVC,starttVC,end(Itarget(t)−Iindividual(t))2
(3)
where E1 is the objective function when only current clamp data (i.e., stochastic stimulation or a single action potential) was fit, and E2 is the objective function for the combined stochastic stimulation and multi-step voltage clamp protocol. In both equations, Vtarget is the membrane potential during current clamp of the target (i.e., either the simulated nominal model or the experimental data), and Vindividual is the membrane potential of a simulated individual. Itarget and Iindividual are the current responses during voltage clamp of the target and a simulated individual, respectively. Errors are summed over the entire duration of the protocols.
Although of different units, the voltage clamp and the current clamp components to E2 were simply summed into a single objective (Eq 3) as we expect them to be minimal for the same range of parameters, rather than being competitive as in typical multi-objective optimization. E2 is therefore unitless.
The estimated model parameters are the maximal conductances of INa, ICaL, ICaT, IK1, IKr, IKs, IKp, and IpCa, and the maximal flux of JSERCA.
Optimizations were run on a 3.2Ghz Intel Xeon W3670 6-core, 6GB memory, machine and took approximately 8 hours per run for the iterative approach and then combined stochastic pacing and voltage clamp protocol.
Two sample t-tests were performed with a significance level of 0.05. Numbers and error bars indicate average ± standard deviation.
Guinea pigs (n = 4) were anesthetized using an intraperitoneal injection with Euthasol (Virbac Corporation, Fort Worth, TX), 550 mg/kg. Excised hearts were then Langendorff retrograde perfused, and myocytes were isolated from the base (top 1/3) of the left ventricle through enzymatic digestion. Myocytes were stored in Dulbecco’s Modified Eagle Medium (DMEM) with 5% fetal bovine serum (FBS).
Amphotericin-B (Sigma-Aldrich Corp., St. Louis, MO; 480 μg per 1 ml pipette solution) perforated patch clamp technique was used to record cellular action potentials. Bath solution contained (in mmol/l) 139.4 NaCl, 5.4 KCl, 1.0 MgSO4, 10.0 Hepes, 10.0 dextrose, 2.0 CaCl2, pH 7.35 with NaOH, osmolality 310 ± 3 mmol/kg. Intracellular solution contained (in mmol/l unless otherwise noted) 125 KCl, 10 NaCl, 5.5 dextrose, 0.5 MgCl2, 11 KOH, 10 Hepes, 10 μmol/l CaCl2, pH 7.1 with HCl, osmolality 295 ± 3 mmol/kg. Recordings were performed at 35°C.
Patch-clamp measurements were recorded using an Axopatch 200A amplifier (Molecular Devices, Sunnyvale, CA). The Real-Time eXperiment Interface [RTXI; rtxi.org; [51,52]] software platform was used to control the amplifier and record data. Cells were initially paced in current clamp mode at a BCL of 500 ms to steady state (500–1000) beats using suprathreshold square pulses of 1 ms duration. Next, the optimization and prediction stochastic stimulation sequences were applied. Amplifier mode was then switched to voltage clamp and series resistance measured (4–8 MΩ) and compensated for (70–90%). The multi-step voltage clamp protocol was then applied in triplet. Holding potentials were corrected for a liquid junction potential of -3 mV. The magnitude of the Donnan equilibrium was estimated to 0 mV using the IV-curve of INa and therefore not corrected for.
To remove stimulus artifacts from current-clamp traces, data from a 1.3 ms window following the start of each stimulus were excluded from the optimization. In addition, voltage-clamp data from a 1.2 ms window following each potential change were excluded from the GA optimization because of the capacitance transient.
From the set of three voltage-clamp trials, the current response trace with the shortest time to peak INa (step to -40 mV at 600 ms) was selected for each cell.
|
10.1371/journal.ppat.1004013 | Identification of OmpA, a Coxiella burnetii Protein Involved in Host Cell Invasion, by Multi-Phenotypic High-Content Screening | Coxiella burnetii is the agent of the emerging zoonosis Q fever. This pathogen invades phagocytic and non-phagocytic cells and uses a Dot/Icm secretion system to co-opt the endocytic pathway for the biogenesis of an acidic parasitophorous vacuole where Coxiella replicates in large numbers. The study of the cell biology of Coxiella infections has been severely hampered by the obligate intracellular nature of this microbe, and Coxiella factors involved in host/pathogen interactions remain to date largely uncharacterized. Here we focus on the large-scale identification of Coxiella virulence determinants using transposon mutagenesis coupled to high-content multi-phenotypic screening. We have isolated over 3000 Coxiella mutants, 1082 of which have been sequenced, annotated and screened. We have identified bacterial factors that regulate key steps of Coxiella infections: 1) internalization within host cells, 2) vacuole biogenesis/intracellular replication, and 3) protection of infected cells from apoptosis. Among these, we have investigated the role of Dot/Icm core proteins, determined the role of candidate Coxiella Dot/Icm substrates previously identified in silico and identified additional factors that play a relevant role in Coxiella pathogenesis. Importantly, we have identified CBU_1260 (OmpA) as the first Coxiella invasin. Mutations in ompA strongly decreased Coxiella internalization and replication within host cells; OmpA-coated beads adhered to and were internalized by non-phagocytic cells and the ectopic expression of OmpA in E. coli triggered its internalization within cells. Importantly, Coxiella internalization was efficiently inhibited by pretreating host cells with purified OmpA or by incubating Coxiella with a specific anti-OmpA antibody prior to host cell infection, suggesting the presence of a cognate receptor at the surface of host cells. In summary, we have developed multi-phenotypic assays for the study of host/pathogen interactions. By applying our methods to Coxiella burnetii, we have identified the first Coxiella protein involved in host cell invasion.
| Infectious diseases are among the major causes of mortality worldwide. Pathogens‚ invasion, colonization and persistence within their hosts depend on a tightly orchestrated cascade of events that are commonly referred to as host/pathogen interactions. These interactions are extremely diversified and every pathogen is characterized by its unique way of co-opting and manipulating host functions to its advantage. Understanding host/pathogen interactions is the key to face the threats imposed by infectious diseases and find alternative strategies to fight the emergence of multi-drug resistant pathogens. In this study, we have setup and validated a protocol for the rapid and unbiased identification of bacterial factors that regulate host/pathogen interactions. We have applied this method to the study of Coxiella burnetii, the etiological agent of the emerging zoonosis Q fever. We have isolated, sequenced and screened over 1000 bacterial mutations and identified genes important for Coxiella invasion and replication within host cells. Ultimately, we have characterized the first Coxiella invasin, which mediates bacterial internalization within non-phagocytic cells. Most importantly, our finding may lead to the development of a synthetic vaccine against Q fever.
| Coxiella burnetii is an obligate intracellular Gram-negative bacterium responsible of the worldwide neglected zoonosis Q fever [1], [2]. Acute forms of the disease are characterized by a febrile illness associated with severe headache, pneumonia and hepatitis. In a small percentage (2–5%) of cases, acute Q fever develops into a chronic infection that may lead to endocarditis and chronic fatigue syndrome [1], [3]. Coxiella resists environmental stress by generating small cell variants (SCVs) that facilitate its airborne dissemination; during infections, this pathogen converts into a metabolically active large cell variant (LCV) with a unique resistance to the degradative machinery of host cells [4], [5]. These factors contribute to the extreme infectivity of this microbe, making of Coxiella a serious health concern, especially in rural areas where outbreaks are likely to occur and are accompanied by heavy economic burdens [6], [7]. Moreover, the development of Coxiella as a potential bioweapon during and since World War II, has ascribed this pathogen among class B biothreats [7]. Coxiella has two antigenic phases: phase I organisms, isolated from natural sources of infection, are extremely virulent. Phase II bacteria originate from spontaneous mutations after several in vitro passages of phase I organisms and present a truncated lipopolysaccharide (LPS) [8]. These non-reversible mutations result in a strong attenuation of virulence in vivo [9], [10]. Phase II Coxiella organisms are internalized more efficiently than phase I organisms by both professional macrophages and non-phagocytic cells [9], [11], however, once internalized, both antigenic phases replicate within host cells with similar kinetics. A phase II clone (Nine Mile phase II clone 4 or NMIIC4), which has been authorized for biosafety level 2 (BSL-2) manipulation, represents therefore an optimal model to study Coxiella infections [2], [5]. In natural infections, Coxiella has a tropism for alveolar macrophages [1], [2], however, infection of epithelial and endothelial cells has also been reported [12], [13]. Indeed, in vitro, Coxiella invades and replicates in a wide variety of phagocytic and non-phagocytic cells [5]. Coxiella internalization within host cells is a passive, endocytic process, which involves the remodeling of the host cell actin cytoskeleton [14], [15] and αVβ3 integrins have been reported as Coxiella receptors in THP-1 cells [11]. However the Coxiella factors that mediate interactions with host cell surfaces, as well as the bacterial host receptor on epithelial cells remain unknown. During the first 48 hours following internalization, bacteria reside into tight-fitting vacuoles, positive for early endosomal and autophagosomal markers [16]. As Coxiella-containing vacuoles mature along the endocytic pathway, the drop in vacuolar pH triggers the translocation of bacterial proteins by a Dot/Icm type 4b secretion system (T4SS) [17]. Effector translocation is essential for the biogenesis of a large parasitophorous vacuole (PV) that occupies the majority of the host cytoplasm [18], [19]. Such large membranous structures are highly dynamic and fusogenic and the host endocytic SNARE Vamp7 is required for optimal PV development [20]. Importantly, mature Coxiella PVs are positive for lysosomal markers and contain active degradative enzymes [5], [16]. Coxiella infections are not lytic and bacteria-filled PVs persist within infected cells, which are protected from apoptosis by a Dot/Icm-dependent mechanism [19], [21]–[25]. Importantly, due to the obligate intracellular nature of this pathogen, the microbial factors involved in host/pathogen interactions remain to date largely unknown. The homology between the T4SS of C. burnetii and L. pneumophila allowed the in silico identification of 354 candidate Coxiella effectors based on the presence of a conserved Dot/Icm regulatory motif (PmrA) [26]–[28], C-terminal translocation signals (E-block) [26]–[28], and eukaryotic-like domains [29]–[31]. Dot/Icm–dependent secretion has been validated for 108 of these using either Coxiella or Legionella as a surrogate host [18], [26]–[31]. Recent advances in Coxiella axenic culture techniques [32] rendered this pathogen genetically tractable [33], allowing for the first time to couple bioinformatics analysis to morpho-functional assays and investigate the role of candidate Coxiella virulence determinants in intracellular replication [18], [19], [27]. To date, 20 Coxiella genes encoding Dot/Icm substrates have been mutated to investigate their role in Coxiella replication within the host [27]. Here we have set up new, integrative approaches that combine transposon mutagenesis with genomics, bioinformatics and fluorescence-based functional assays aiming at the large-scale identification of intracellular bacteria virulence factors. Our approach is designed for the simultaneous investigation of multiple key steps of Coxiella infections and is based on the identification and characterization of transposon-induced phenotypes. We have generated and isolated 3000 Coxiella transposon mutants, 1082 of which have been sequenced and screened in the present study. Our analysis revealed important insights into the functionality of the Coxiella Dot/Icm apparatus and revealed a variety of bacterial factors involved in 1) internalization within host cells, 2) PV biogenesis and intracellular replication, and 3) protection of the infected cell from apoptosis. By focusing our analysis on the early events of Coxiella infections we identified the first Coxiella invasin that plays an essential role in bacterial internalization by non-phagocytic cells.
To identify the Coxiella factors involved in host-pathogen interactions, we have undertaken the generation of a library of GFP-tagged bacterial mutants by transposon mutagenesis. We have modified the Himar1-based transposon system initially developed by Heinzen and colleagues [33], [34], by inserting the enhanced green fluorescent protein (egfp) gene under the regulation of the Coxiella promoter P311, upstream of the chloramphenicol resistance cassette, thus generating pITR-CAT-ColE1-P311-GFP. To obtain stable mutants, Nine Mile Phase II clone 4 (NMIIC4) Coxiella (hereafter referred to as wt Coxiella) were electroporated using a two-plasmid system, where the transposase is encoded by a suicide plasmid that is lost during bacterial replication [34]. The eGFP-tagged Coxiella mutants thus generated were isolated on ACCM-2 agar plates in the presence of chloramphenicol and further amplified for 7 days in liquid ACCM-2 supplemented with chloramphenicol. The final concentration of each bacterial culture was calculated using the Quant-iT PicoGreen dsDNA assay. Transposon insertion sites were identified by single-primer colony PCR followed by DNA sequencing. Using the primer SP3 we amplified DNA fragments including a 278 bp region upstream of the 3′ Inverted Terminal Repeat (ITR) of the inserted transposon (Fig. 1A). The amplified fragments were then sequenced using the transposon-specific primer P3, which recognizes a sequence in the 3′ region of the Chloramphenicol Acetyltransferase (CAT) gene (Fig. 1A). The obtained sequences were then aligned on the Coxiella burnetii RSA493 annotated genome using automated sequence analysis software. The genome of Coxiella burnetii RSA493 contains 1849 coding sequences (CDS), 1814 in the bacterial chromosome and 35 in the cryptic plasmid QpH1 [35]. To date we have isolated 3000 transposon mutants, 1082 of which have been sequenced, annotated and analyzed for this study (Fig. 1B). Transposon insertions were homogeneously distributed throughout the Coxiella chromosome and plasmid, with seven “hot spots” of preferential transposon insertion (identified and annotated from 1 to 7, Fig. 1B) and a large, 52 CDS region, upstream of hot spot n. 2, which remained non-mutated. Of note, region n.7 corresponds to the locus that hosts T4SS core genes (dot/icm genes) whereas the non-mutated region between CBU_0215 and CBU_0272 is enriched in genes encoding ribosomal proteins. Overall, 926 transposon insertions were found within Coxiella annotated CDS and 156 in intergenic regions of the Coxiella genome (excluding insertions within the first 100 bp upstream of a CDS; Fig. 1C). Frequency distribution analysis revealed that mutations occurred in 483 CDS on the Coxiella chromosome and 8 CDS on the QpH1 plasmid (corresponding to 26.6% and 22.8% of the total CDS present on chromosome and plasmid respectively; Fig. 1C). The mutated CDS were then clustered according to their predicted function based on the data available on the Pathosystems Resource Integration Center (PATRIC, www.patricbrc.org; Fig. 1D).
Isolated transposon mutants were used to infect Vero cells at comparable multiplicities of infection (MOI). Non-infected Vero cells were used as negative control and cells infected with GFP-NMIIC4 Coxiella (GFP-Coxiella) [33] were used as positive controls. Variations of GFP fluorescence associated with intracellular bacterial growth over 7 days of infection were monitored using a multimode micro-plate reader to obtain intracellular growth curves for all the Coxiella mutants screened (Fig. 2A). Seven days after infection, plates were imaged using an automated fluorescence microscope and images were analyzed using the automated image analysis software CellProfiler (www.cellprofiler.org). 9 morphological features were extrapolated from an average of 14000 cells for each condition (Fig. 2A). Nuclei features were used to assess the overall conditions of host cells and the potential cytotoxicity of Coxiella mutations. Coxiella features were used to score the intracellular replication of bacteria. Finally, the number of Coxiella colonies was divided by the number of host cell nuclei to estimate the efficiency of bacterial internalization within cells.
We validated our approach by comparing the growth curves and morphology of GFP-Coxiella to those of 6 mutants (Tn2384, Tn1948, Tn292, Tn2184, Tn514, Tn270) carrying independent transposon insertions in the gene CBU_1648, which encodes DotA, an essential component of the Coxiella Dot/Icm secretion system [36] (Fig. 2B). Axenic growth of the 6 dotA mutants was comparable to that of wt Coxiella (Fig. S1A). GFP signal analysis indicated that intracellular growth of GFP-Coxiella followed a typical growth curve, with bacterial replication clearly detectable from day 2 post-infection and increasing until day 7 post-infection (Fig. 2C). Morphological analysis of GFP-Coxiella-infected Vero cells reported bacterial colonies with an average area of 80.71±4.65 microns2 and an average number of Coxiella colonies/cell of 1.13±0.12 (n = 21656; Fig. 2D). As expected, the GFP signal originated by all the DotA transposon mutants failed to increase significantly during the 7 days of infection (Fig. 2C). Accordingly, morphological analysis indicated a strong reduction in the average area of intracellular Coxiella colonies, which ranged from 5.86±1.85 to 32.03±4.83 microns2 (n = 19066 and 14879 respectively, Fig. 2D). Interestingly, the majority of dotA mutations corresponded to an increased number of colonies/cell, which, in the case of mutant Tn514, reached 3.48±0.61 (Fig. 2D). This phenotype suggests that, in the absence of a functional secretion system, the Coxiella-containing vacuoles fail to coalesce to form the typical, large PV. Moreover, two independent transposon insertions in the Dot/Icm gene icmV (Tn2445 and Tn2214), which is located in the same operon as dotA, produced a comparable, strong replication phenotype (data not shown).
The data derived from the multi-phenotypic analysis of Coxiella infections of host cells were mined to identify mutations that perturbed: 1) host cell invasion, 2) intracellular replication and 3) host cell survival. To screen for invasion and replication phenotypes, we plotted the average area of Coxiella colonies against the average number of colonies/cell (Fig. 3A). Statistical analysis was used to define regions in the resulting scatter plot corresponding to mild (−4<Z-score≤−2) and severe (Z-score≤−4) phenotypes. The 1082 analyzed Coxiella mutants were found in 3 well-defined clusters: one included mutants whose phenotype did not vary significantly from that of GFP-Coxiella (Fig. 3A green dots). A second cluster was clearly shifted towards a reduction in the size of Coxiella colonies and an increase in the average number of Coxiella colonies/cell, representative of mutations that affect Coxiella intracellular replication without affecting host cell invasion (Fig. 3A light and dark red dots). A third cluster was shifted towards a reduced number of colonies/cell, indicating mutations that affect host cell invasion (Fig. 3A light and dark blue dots). In parallel, the average area of Coxiella colonies was plotted against the total number of host cells surviving the 7 days of infection, to identify Coxiella genes that are potentially involved in the protection of the host cell from apoptosis (Fig. 3B). As above, statistical analysis was used to define regions corresponding to mild (−4<Z-score≤−2) and severe (Z-score≤−4) phenotypes. The vast majority of the mutants analyzed did not affect cell survival, regardless of bacterial replication within host cells (Fig. 3B green dots). 37 mutations mildly affected host cell survival (Fig. 3B light red dots), and 7 mutations were particularly detrimental to host cell survival (Fig. 3B dark red dots). Next, the phenotypic data from mutations within CDS were integrated with the annotated transposon insertions in the Coxiella genome. We thus clustered the screened mutations according to the mutated CDS and assigned to each mutant a color-coded map based on the intensity of their replication (R), internalization (I) or cytotoxic (C) phenotype (Table S1). Intergenic transposon insertions were retained in a separate table (data not shown).Importantly, mutant Tn1832 carries an intergenic transposon insertion and phenocopies wt Coxiella and GFP-Coxiella (Fig. S2). This mutant has been used in our validation experiments as additional control.
Intracellular replication of Coxiella relies on the functionality of a Dot/Icm T4SS, which is highly homologous to that of L. pneumophila [2], [37]–[39]. The Coxiella genome contains 23 homologues to the 25 known dot/icm genes of the Legionella T4SS. Accordingly, Legionella has been used as a model organism to test the secretion of candidate Coxiella effector proteins. Recently, it has been reported that the Coxiella dot/icm genes icmD, dotA, dotB and icmL.1 are essential for Coxiella replication within host cells, proving for the first time the functionality of the Coxiella T4SS [18], [19], [36]. The enrichment of transposon insertions in dot/icm genes (Fig. 1B, region n. 7), prompted us to analyze the phenotype of 38 Coxiella mutants carrying single transposon insertions in 16 Dot/Icm genes (Fig. 4A; Fig. S1). First, we followed the axenic growth of the 38 Dot/Icm transposon mutants and found it to be indistinguishable from that of wt Coxiella and of the control transposon mutant Tn1832 (Fig. S1A). Multi-phenotypic analysis confirmed the previously reported observations that icmD, dotA and icmL.1 are essential for Coxiella replication within host cells [18], [19], [36] (Fig. 4B, C; Fig. S1B). Moreover, we could observe that 12 dot/icm genes (dotA, dotB, icmV, E, D, G, J, N, C, P, K, X, L.1) are essential for bacterial replication within the host, whereas mutations in icmB and icmS showed an intermediate phenotype, which corresponded to partial intracellular replication as assessed by morphological analysis (Fig. 4B, C; Fig. S1B). Of note, transposon insertions in dot/icm genes present in operons produced consistent phenotypes. Each mutation was then trans complemented by challenging host cells in combination with wt Coxiella as previously described [19]. The intergenic mutant Tn1832 was used as positive control. As illustrated in Figure 4C, mutations resulting in severe replication phenotypes were efficiently complemented by the presence of wt Coxiella in the PV occupied by Dot/Icm mutants (black bars).
We have also isolated and screened 63 transposon mutants in 31 CDS encoding predicted Coxiella Dot/Icm substrates (Fig. 3C). The products of 22 of these CDS have been previously reported to be positive for Dot/Icm secretion [18], [26], [27], [29], [30], [31] (Fig. 3C). Our analysis indicated that mutations in 17 of these genes produced replication phenotypes (Fig. 3C, CDS in red), further suggesting that these genes encode Coxiella effectors. Interestingly, a transposon insertion in CBU_1639 resulted in a strong cytotoxic phenotype (Fig. 3C CDS in blue), suggesting that this gene plays a role in the Coxiella-mediated inhibition of apoptosis.
The Coxiella genome encodes 4 two-component systems: PhoB-PhoR (CBU_0367-CBU_0366), GacA-GacS (CBU_0712-CBU_0760), QseB-QseC (CBU_1227-CBU_1228) and an RstB-like system (CBU_2005-CBU_2006) [2], [35]. In particular, the QseB-QseC system has been reported to be homologous to the L. pneumophila PmrA-PmrB system, a fundamental regulator of Dot/Icm secretion and its role has been indirectly confirmed [26], [40]. Consistent with these observations and with our analysis on Coxiella Dot/Icm genes, 6 independent transposon insertions in CBU_1227 (qseB) and one insertion in CBU_1228 (qseC), significantly impaired Coxiella replication within host cells (Table S1). A single transposon insertion in CBU_2006, part of the RstB-like system also produced a significant replication and entry phenotype (Table S1), however, the role of this two-component system in Coxiella remains to be defined.
The annotation of the Coxiella burnetii NMI RSA493 genome revealed the presence of 207 pseudogenes, which are not conserved among different Coxiella isolates [35]. Recent whole transcriptome analysis (RNA-Seq) has validated the previous annotation confirming that these sequences do not encode complete open reading frames (Prof. Howard Shuman personal communication). Interestingly, we have isolated 85 transposon mutants in 56 CDS annotated as pseudogenes and mutations in 32 of these resulted in a strong replication phenotype (Fig. S3). Finally, we could observe that 71 out of the 151 transposon insertions in intergenic regions of the Coxiella genome exhibited a significant replication phenotype (not shown). This may reveal small RNA-mediated regulation of Coxiella virulence. Indeed, 11 transposon insertions fall within intergenic regions where putative Coxiella sRNAs have been identified by RNA-Seq (Prof. Howard Shuman personal communication). Further investigations are currently aiming at an in-depth characterization of these new candidate factors that may regulate intracellular replication of Coxiella.
As mentioned above, our multi-phenotypic screen identified 7 transposon mutants that exhibited a strong cytotoxic phenotype when incubated with Vero cells (Fig. 3B, Fig. S4). To further analyze the phenotype of these mutants, we investigated their intrinsic capacity of triggering apoptosis and their potential of protecting infected cells from staurosporine-induced apoptosis. HeLa cells were preferred to Vero cells for this assay as they displayed a more consistent response to staurosporine and all Coxiella strains tested displayed similar replication phenotype in these cells (data not shown). Cells were either left unchallenged or incubated with wt Coxiella, the control transposon mutant Tn1832, the DotA mutant Tn270 and the 7 cytotoxic mutants (Tn881, Tn616, Tn946, Tn926, Tn1226, Tn1232, Tn1233). Three days post-inoculation, cells were either fixed in paraformaldehyde or incubated with 1 µM staurosporine for 4 h prior to fixation. The percentage of apoptotic cells was then evaluated for each condition by the TUNEL assay. Very few TUNEL-positive cells were observed among untreated cells, these were increased to 50% of the total cell population upon staurosporine treatment. As expected, incubation of cells with wt Coxiella did not increase the number of TUNEL-positive cells as compared to untreated cells and wt Coxiella-colonized cells were efficiently protected from staurosporine-induced apoptosis. Cells challenged with the control transposon mutant Tn1832 presented the same phenotype as cells incubated with wt Coxiella and conversely, the DotA mutant Tn207 failed to protect infected cells from induced apoptosis (Fig. S4). Mutant Tn881, which carries a transposon insertion in CBU_0485, exhibited a partial protection of infected cells from induced apoptosis whereas the remaining mutants failed to effectively protect cells from the effects of staurosporine (Fig. S4). Interestingly, incubation of HeLa cells with mutants carrying transposon insertions in CBU_1639, CBU_1366 and CBU_0307a significantly increased the number of TUNEL-positive cells also in the absence of staurosporine, indicating that these mutants may possess intrinsic cytotoxic properties (Fig. S4).
As for all obligate intracellular pathogens, Coxiella invasion of host cells is a priming step of the infection. However, since the bacterial factors that mediate Coxiella invasion of host cells remain unknown, we sought to identify bacterial factors whose mutations affect Coxiella internalization. High-content screening identified 48 mutations in 37 Coxiella CDS that resulted in a significant reduction in the number of infected cells after 7 days of infection (Fig. S5). Of these, 18 CDS involved in bacterial metabolism and transcription were excluded. Among the remaining 19 candidate CDS (Fig. S5, CDS boxed in red), we have identified 5 independent transposon insertions (Tn175, Tn208, Tn27, Tn907 and Tn749) in CBU_1260, all sharing a consistent, strong internalization phenotype (Fig. S5). CBU_1260 is a 747 bp CDS on the positive strand of the Coxiella chromosome encoding a hypothetical protein of a predicted size of 23 kDa. Importantly, the gene is not part of an operon, indicating that the phenotype observed for the 5 mutants analyzed in this study was indeed due to the inactivation of CBU_1260 alone (Fig. 5A). Axenic growth of the 5 transposon mutants was comparable to that of wt Coxiella and of the control transposon mutant Tn1832 (Fig. S6A). As expected, intracellular growth curve analysis of the 5 transposon mutants in CBU_1260 indicated that GFP fluorescence failed to increase during the 7 days of infection (Fig. 5B). Indeed, all 5 mutations in CBU_1260 reduced the number of cells presenting Coxiella colonies at 7 days of infection by 60–70% compared to cells challenged with GFP-Coxiella (Fig. 5C). We next used the online analysis software i-TASSER (http://zhanglab.ccmb.med.umich.edu/I-TASSER/) and Phyre2 (http://www.sbg.bio.ic.ac.uk/phyre2/) to predict the structure of the hypothetical protein encoded by CBU_1260. Bioinformatics analysis predicted 8 transmembrane beta sheets forming a beta barrel domain, an N-terminal alpha helix and 4 unstructured loops (L1 to L4), exposed at the cell surface (Fig. 5A, D). This prediction was confirmed by analyzing the sequence of the protein using TMpred (http://www.ch.embnet.org/software/TMPRED_form) and the BetAware software [41]. Sequence analysis indicated that transposon insertions occurred in the distal part of the CDS, within the 4th and 6th beta sheets, with mutants Tn27 and Tn907 presenting insertions at the same site (Fig. 5A). The predicted transmembrane domain of CBU_1260 is typical of Outer Membrane Protein A (OmpA) family of proteins, which are found in several bacteria and mediate adhesion and/or internalization within host cells [42]–[50]. We therefore named the product of CBU_1260 OmpA. Coxiella encodes 3 hypothetical proteins that contain predicted OmpA-like domains: CBU_0307, CBU_0936 and CBU_1260 (ompA). Sequence alignment of these three hypothetical proteins showed high degree of homology at the level of the transmembrane domains and the N-terminal alpha helix (Fig. S7). However, little homology was observed at the level of the 4 unstructured loops of OmpA (Fig. S7). Accordingly, a transposon insertion in CBU_0307 produced a cytotoxic phenotype whereas 4 transposon insertions in CBU_0936 produced a replication phenotype (Table S1). An OmpA-specific antibody was then raised against the predicted extracellular domain of the protein. To this aim, a 15 amino acid peptide in predicted loop 1 (KKSGTSKVNFTGVTL) was used for its immunogenic potential as compared to peptides in the other loops. When tested by Western blot on lysates from wt Coxiella and the control mutant Tn1832, the anti-OmpA antibody revealed a major band at the expected size of 23 kDa (Fig. 5E, arrow) and a faint, background band of lower molecular weight (Fig. 5E). When incubated on lysates from the 5 transposon mutants in CBU_1260, the anti-OmpA antibody only recognized the background band of lower molecular size. We next performed membrane fractionation assays on wt Coxiella and the OmpA mutant Tn208 to validate the outer membrane localization of OmpA. The protein was highly enriched in the outer membrane fraction of wt Coxiella and, as expected, absent in lysates from the OmpA mutant Tn208 (Fig. 5F). Taken together, our data indicate that CBU_1260 encodes an outer membrane protein with a predicted OmpA-like structure that plays a relevant role in host cell invasion.
To further investigate the role of OmpA in Coxiella invasion of host cells, we validated the transposon insertion in the OmpA mutant Tn208 by PCR and used a GFP-probe to confirm by Southern blot that Tn208 contained a single transposon insertion (Fig. S6B, C). Next, non-phagocytic epithelial cells (A431), THP-1 (PMA-differentiated), J774 and RAW 264.7 macrophages were challenged either with wt Coxiella, the OmpA mutant Tn208 or the control transposon mutant Tn1832. Differential labeling of extracellular and intracellular bacteria was used to assess the efficiency of Coxiella internalization within host cells at 15, 30, 45 and 60 minutes post-infection (Fig. 6A, E, top panels). Longer time points (5 and 6 days post-infection) were analyzed to investigate the intracellular development of internalized OmpA mutants (Fig. 6A, E, bottom panels). Automated image analysis was then used to analyze approximately 8000 bacteria per condition and quantify the efficiency of bacterial internalization as well as the area occupied by intracellular Coxiella colonies. In A431 non-phagocytic cells, the internalization of Tn208 was strongly reduced as compared to that of wt Coxiella or Tn1832, which shared similar kinetics (Fig. 6B). Interestingly, when the same internalization experiment was performed using macrophages, the three bacterial strains tested (wt Coxiella, Tn208 and Tn1832) were internalized with comparable efficiency (Fig. 6F; Fig. S8A, D), with a concomitant local rearrangement of the actin cytoskeleton (Fig. 6E, top panels). Of note, despite the strong inhibition of bacterial internalization in A431 cells, OmpA mutants retained the capacity to adhere to host cells (Fig. 6A). This suggests that if OmpA plays a role in bacterial adhesion, this may be masked by the presence of alternative factors involved in Coxiella adhesion to host cells. At longer time points of infection we observed a decrease in the number of Tn208 colonies per cells in A431 cells as compared to wt Coxiella and Tn1832 colonies, which confirmed our previous observations in Vero cells (Fig. 6C). On the contrary, the number of Coxiella-colonized cells was not affected when macrophages were challenged either with wt Coxiella, the control mutant Tn1832, or the OmpA mutant Tn208 (Fig. 6G, Fig. S8B, E). Remarkably however, the average area of OmpA mutant Coxiella colonies was significantly reduced as compared to wt Coxiella and Tn1832, regardless of the cell line used for the experiment (Fig. 6D, H; Fig. S8C, F). Interestingly, this is in agreement with previously reported roles of OmpA proteins in intracellular survival of bacterial pathogens [43], [45], [46], [51]–[53].
To further dissect the role of OmpA in Coxiella interaction with host cell surfaces, we 1) purified the recombinant protein to coat inert latex beads and 2) ectopically expressed OmpA in E. coli, to assess the capacity of OmpA to confer adhesiveness and invasiveness to inert particles and extracellular bacteria, respectively. Histidine-tagged, recombinant OmpA was produced by E. coli BL21-DE3 star pLysS transformed with the pET28a vector containing ompA32-248. The first 31 amino acids of OmpA corresponding to the intracellular N-terminal alpha helix were excluded to increase the solubility of the protein. Red fluorescent latex beads were then coated with 100 µg/ml His-OmpA32-248 or GST as control and used to challenge A431 cells for 1 hour at 37°C. Unbound beads were removed by rinsing cells in PBS and cells were fixed in paraformaldehyde. Cells were then probed with an anti-histidine antibody without permeabilization to differentially label adherent and internalized beads and with fluorescent phalloidin to define the cell perimeter and volume. Alternatively, after fixation samples were further processed for scanning electron microscopy. Confocal microscopy analysis of cross sections of cells incubated with His-OmpA32-248-coated beads revealed a fraction of beads adhering to the cell surface, hence positive for the anti-histidine staining, and another fraction within the cell volume (as defined by the actin labeling) and negative to the anti-histidine staining (Fig. 7A). Three-dimensional reconstruction of confocal sections coupled to surface rendering confirmed the presence of adhering and internalized beads (Fig. 7A). GST-coated beads failed to adhere to and invade A431 cells significantly (Fig. 7B). Scanning electron microscopy analysis corroborated these observations: several His-OmpA32-248-coated beads were adhering to the surface of A431 cells (Fig. 7C, green inset) whereas others were clearly covered by the cell plasma membrane (Fig. 7C, red inset). Very few GST-coated beads were observed at the surface of cells and none seemed to be internalized (data not shown), confirming our observations by fluorescence microscopy. We next assessed the capacity of OmpA to trigger the internalization of non-invasive bacteria by non-phagocytic cells using the gentamicin protection assay. To this aim E. coli BL21-DE3 star pLysS were transformed with pET27b-OmpA, which allowed the IPTG-regulated expression and periplasmic targeting of full-length OmpA. The expression and outer membrane localization of OmpA were verified by Western blot using the OmpA-specific antibody on transformed E. coli cultures induced overnight with IPTG and processed to separate the bacterial outer membranes from the inner membranes and cytoplasmic components (Fig. 7D). Non-induced, transformed bacteria were used as control. We could observe that IPTG-induced bacteria efficiently produced OmpA, which was enriched in the outer membrane fraction of E. coli lysates (Fig. 7D). Importantly, induction of OmpA expression conferred E. coli a 20-fold increase in invasiveness as compared to the non-induced bacteria (Fig. 7E). Next, we used the E. coli ectopic expression approach to dissect the role of the 4 predicted extracellular loops of OmpA. By replacing each loop with a myc tag, we generated 4 OmpA mutants (OmpAΔL1, OmpAΔL2, OmpAΔL3, OmpAΔL4) that were used to test their capacity to confer invasiveness to E. coli in a gentamicin protection assay. Similar to wt OmpA, all mutated proteins were detected in the outer membrane fraction of induced E. coli (not shown). Interestingly, only the exchange of loop 1 with a myc tag significantly reduced E. coli internalization by non-phagocytic cells (Fig. 7E). Collectively, these data suggest that OmpA is necessary and sufficient to mediate Coxiella internalization within non-phagocytic cells and that loop 1 is primarily involved in interacting with a potential cognate receptor at the surface of host cells.
To determine whether OmpA function requires the interaction with host cell surface factors, we sought to block candidate ligand/receptor interactions, either by saturating potential OmpA receptors at the surface of host cells or by masking OmpA at the surface of bacteria, prior to infection. In the first case, A431 cells were incubated with 100 µg/ml His-OmpA32-248for 1 hour at 4°C prior to challenging with wt Coxiella and the efficiency of bacterial internalization was determined by differential bacterial labeling. A431 cells incubated in the same conditions with GST or with buffer alone were used as controls. Indeed, pretreating A431 cells with His-OmpA32-248 effectively inhibited wt Coxiella internalization as compared to buffer- or GST-treated cells (Fig. 8A, B). Alternatively, wt Coxiella were incubated with increasing concentrations (0.1, 1 and 5 µg/ml) of either anti-OmpA antibody or naïve rabbit serum prior to infection and bacterial differential labeling was used to determine Coxiella invasiveness in A431 cells. Confirming our previous observations, the pre-treatment of bacteria with the anti-OmpA antibody, but not naïve rabbit serum, inhibited Coxiella internalization in a concentration dependent manner (Fig. 8C, D). These observations suggest the presence of a receptor for OmpA at the surface of host cells, which remains to be identified, and that OmpA/receptor interactions are essential to mediate Coxiella internalization within host cells.
Larvae of the wax moth Galleria mellonella are an emerging, efficient model for the study of host/pathogen interactions in vivo. Like other insects, Galleria larvae present essential aspects of the innate immune response to microbial infections, which are conserved in mammals. In particular, insects possess humoral and cellular defense responses, the first including antimicrobial peptides (galiomycin, gallerimycin and lysozyme in the case of Galleria) and the latter consisting of specialized phagocytic cells, known as hemocytes or granulocytes [54]–[57]. Importantly, in the case of several bacterial pathogens, typical phenotypes observed in mammalian infection models were efficiently reproduced using Galleria [58]–[61]. It has been recently demonstrated that the Coxiella closely-related pathogen Legionella pneumophila invades Galleria hemocytes and replicates within large membranous compartments that present the same morphology and characteristics of Legionella-containing vacuoles generated during infection of macrophages and amoeba [62]. Infections by L. pneumophila result in severe damage to insect organs, which is accompanied by an immune response, including larvae melanization and nodule formation [62]. Moreover, the role of bacterial virulence factors previously characterized in higher mammalian models is conserved during infections of Galleria mellonella [61], [62]. Importantly, Galleria mellonella larvae are also susceptible to phase II Coxiella infections (Norville et al. Unpublished data). We thus investigated the phenotype associated with the OmpA mutation carried by Tn208 in the context of Galleria infections. Larvae were exposed to Coxiella by injecting 106 bacteria (either wt Coxiella, Tn1832 or Tn208) in the upper right proleg and larvae were incubated at 37°C up to 300 h post-infection to determine survival rates. Alternatively, larvae were incubated up to 24 and 96 h prior to fixation in paraformaldehyde and processing for immunofluorescence. In all cases, larvae injected with PBS were used as a negative control. Larvae injected with PBS alone did not show any survival defect throughout the incubation time, whereas larvae infected with wt Coxiella or the control mutant Tn1832 died significantly faster compared to those infected with the OmpA mutant Tn208 (Fig. 9A). Immunofluorescence analysis revealed that at 96 h post-inoculation, wt Coxiella as well as the control mutant Tn1832 organisms triggered the formation of large, highly infected nodules of hemocytes (Fig. 9B). These were often juxtaposed to larval organs that also appeared severely infected and damaged (Fig. 9C, top panels). In contrast, when Galleria were challenged with the OmpA mutant Tn208 fewer nodules of smaller size were observed throughout the larvae and only a small fraction of these presented signs of infection (Fig. 9B). When these nodules were observed at higher magnification, we could detect small Tn208 colonies (Fig. 9C bottom panels) that were reminiscent of what we had previously observed in cultured macrophages. Our observations indicate that the OmpA-associated phenotypes observed in cultured cells can be reproduced during in vivo infections.
Infection by bacterial pathogens depends on the subversion of host functions, which is tightly orchestrated by an array of bacterial proteins referred to as virulence factors. In the last decade, cellular microbiology has stressed the importance of studying pathogens in relation to their host, however, the effective, global identification of bacterial virulence determinants and the characterization of their diverse mechanisms of action requires the development of new high-throughput and high-content screens (HTS and HCS respectively) [63]. Here, we have set up protocols for the multi-phenotypic screen of bacterial factors that are involved in host cell invasion and colonization. Our approach integrates transposon mutagenesis, genomics, bioinformatics and fluorescence-based functional assays that have been adapted for the large-scale identification of virulence factors from virtually any intracellular bacterium. The advantage of our screening technique lies in the possibility of analyzing every bacterial mutations for multiple phenotypes, such as 1) internalization within the host/cell, 2) intracellular replication and 3) cytotoxicity, simultaneously. This analysis, integrated with the map of genome mutations, allows a global overview of bacterial genes involved in host/pathogen interactions. The emerging bacterial pathogen Coxiella burnetii is an excellent model system to apply our strategy. To date, very little is known about the bacterial factors that regulate Coxiella interactions with the host; however, the recent development of axenic culture techniques now allows genetic manipulation of this microbe and in-depth analysis of its virulence factors [32], [64]. Moreover, previous in silico identification of putative Coxiella virulence factors [18], [26]–[31], provides an excellent database to cross-reference bioinformatics analysis to our functional assays. Importantly, our assay allows the identification of Coxiella virulence factors on a whole-genome scale, based on the phenotypes associated with the random mutagenesis of Coxiella CDS.
Our aim being the generation of the first bank of C. burnetii transposon mutants, we chose to sequence all isolated mutants independently of their phenotype during infections. This has provided a global survey of the distribution of transposon insertions and allowed us to pinpoint also those genes, suspected to encode virulence factors, which failed to produce a phenotype during Coxiella colonization of host cells. Mapping of transposon insertions revealed an overall homogeneous distribution of mutations, with regions of preferential transposon insertion as well as other regions that remained non-mutated. Of note, the lack of transposon insertions in the region between CBU_0678 and CBU_0698 is expected, having used the annotated genome of Coxiella NMI (RSA493) to map transposon insertions in Coxiella NMII (RSA439) in which this region is missing. Interestingly however, we failed to isolate transposon mutants from the large region between CBU_0215 and CBU_0272, enriched in essential genes encoding ribosomal proteins. This suggests that non-mutated regions may have been targeted by the transposon but gene disruption was lethal for the bacterium. Non-mutated regions may be thus exploited to map essential genes in the Coxiella chromosome and plasmid. When analyzing transposon mutants exhibiting a strong replication phenotype, we have occasionally observed internalization phenotypes that were not consistently reproduced in all transposon mutants isolated for a given gene. We believe that this is due to a technical limitation of our screening technique, imposed by the extremely different sizes and associated fluorescence of Coxiella colonies. In some cases the signal to noise ratio was very close to the threshold imposed to the image analysis software and resulted in colonies and/or bacteria that were not detected when images were segmented. Ongoing implementation of our automated analysis pipeline will allow the identification of these outlier phenotypes.
Independent studies have formally proven the essential role of the Coxiella Dot/Icm secretory apparatus by isolating mutants in 4 of the 23 dot/icm Coxiella genes (dotA, dotB, icmD and icmL.1) [18], [19], [36]. In all cases dot/icm mutants retained the capacity of invading host cells but failed to generate large PVs and replicate therein [18], [19], [36]. The enrichment of transposon insertions in the region of the Coxiella genome hosting dot/icm core genes allowed us to validate our assays exploiting existing data and, more importantly, provided a comprehensive overview of the role of the different components of the Coxiella T4SS. Interestingly, we have identified mutations resulting in intermediate phenotypes that allowed partial bacterial replication. Of particular interest is a mutation in icmS, which confers a multivacuolar phenotype to mutants. The icmS gene encodes a chaperone protein that mediates the secretion of a subclass of bacterial effectors [40]. The observation of a multivacuolar phenotype suggests that in Coxiella, IcmS may be involved in the secretion of effectors that mediate membrane fusion events required for the biogenesis of the PV. To facilitate the identification of putative IcmS substrates, a machine learning approach is currently being used to identify other transposon mutants that share the same multivacuolar phenotype.
To date, candidate Coxiella effectors have been identified by bioinformatics analysis based on conserved Dot/Icm regulatory motif (PmrA) [26], [27], C-terminal translocation signals (E-block) [27], [28], and eukaryotic-like domains. So far, 354 candidate effectors have been thus identified, however Dot/Icm-dependent translocation assays using either Coxiella or Legionella as a surrogate model, indicated that the majority of these might be false positives [18], [26]–[28]. In addition, the lack of efficient methods for the genetic manipulation of Coxiella severely hampered the functional study of these putative effectors. In a recent study, transposon mutants were obtained from 20 Coxiella candidate effectors with 10 exhibiting a significant replication phenotype [27]. Here we report the entry, replication and cytotoxic phenotype of 63 transposon mutants in 31 previously identified Coxiella candidate Dot/Icm substrates. Indeed, some of these candidates play a role during infection whereas some others fail to produce a phenotype, stressing the importance of coupling high-content screens to in silico analysis to identify bacterial effectors. Moreover, further studies of other Coxiella genes sharing none of the features of Dot/Icm substrates, can be exploited to enrich existing databases for the bioinformatics-based identification of Coxiella effectors, thus creating a feedback loop that would significantly improve and accelerate the study of Coxiella pathogenesis.
A considerable number of transposon insertions were mapped outside Coxiella CDS. By excluding mutations that affected the first 100 bp upstream of annotated genes (to exclude mutations that might affect promoter regions of genes), we obtained a list of 151 intergenic transposon insertions. Interestingly, 71 of these resulted in a significant replication phenotype, suggesting that these non-coding regions of the Coxiella genome may play a role in host/pathogen interactions. sRNAs are emerging as regulators that enable pathogens to adapt their metabolic needs during infection and timely express virulence factors [65], [66]. However, recent studies in other organisms revealed the existence of a number of putative sRNAs higher than initially expected, suggesting the presence of many non-functional sRNAs and complicating the identification of relevant sRNAs. The functional data obtained by our screening approach are being cross referenced with a list of putative Coxiella sRNAs identified by RNA-seq to facilitate the identification of Coxiella sRNAs that may coordinate host/pathogen interactions. Similarly, the interesting observation that a number of mutations in Coxiella CDS annotated as pseudogenes have an effect in host cell infection suggests that these genomic regions may have an important regulatory role.
Bacterial adhesion and invasion of host cells is a fundamental step of the infection by intracellular bacterial pathogens [67]. These processes can be active or passive depending on the nature of the pathogen. “Triggering” bacteria commonly use a type 3 secretion system (T3SS) to inject effectors across the host cell plasma membrane to trigger actin rearrangements and pathogen internalization by phagocytosis, whereas “zippering” bacteria use surface proteins that interact with cognate receptors at the surface of host cells [67]. This activates a ligand/receptor signaling cascade that leads to the internalization of large particles by an endocytosis-like mechanism [68]–[71]. The lack of a T3SS in Coxiella suggests that these organisms adhere to and invade host cells by a zippering mechanism. Indeed, it has been reported that Coxiella is passively internalized by host cells by a yet undefined mechanism, which is accompanied by the local rearrangement of the actin cytoskeleton [11], [14], [15]. αVβ3 integrins have been shown to mediate Coxiella adhesion to THP-1 cells [11], however, the lack of these integrins at the surface of epithelial cells, which are effectively colonized by Coxiella, suggest the presence of additional/alternative receptors. Similarly, the Coxiella surface determinants for host cell adhesion and invasion remain to be defined. Here we have identified the product of CBU_1260 as the first Coxiella invasin. Predictive analysis on the primary sequence of CBU_1260 revealed the presence of 8 beta sheets forming an OmpA-like domain highly homologous to that identified and characterized in several other bacterial pathogens [49], [50]. Examples are the OmpA proteins encoded by E. coli K1 [44], [47], Yersinia pestis [45], Francisella tularensis [46], Klebsiella pneumoniae [48] and Shigella flexneri [72]. These outer membrane proteins are involved in bacterial adhesion and/or internalization within host cells, as well as in the NF-κB-mediated modulation of the immune response to infection, which is required for intracellular bacterial development. Importantly, OmpA-like proteins with similar functions in bacterial adhesion and internalization have been reported in other bacterial pathogens such as Rickettsia conorii, Anaplasma phagocytophilum and Ehrlichia chaffeensis [73]–[75], however these proteins share no structural homology with the OmpA proteins described above. In agreement with in silico predictions, membrane fractionation experiments performed in Coxiella as well as in E. coli ectopically expressing OmpA, showed that the protein is indeed enriched in the outer membrane fraction of bacterial lysates. Our multi-phenotypic analysis revealed that five independent transposon insertions that disrupted CBU_1260 sequence severely affected Coxiella internalization and replication within host cells. The internalization phenotype was specific of non-phagocytic cells, whereas OmpA mutants were still internalized by phagocytic cells. This observation indicated that, in the absence of an active phagocytic process, OmpA is able to actively trigger Coxiella internalization by means of ligand/receptor interactions. Importantly however, the intracellular replication of OmpA mutants was severely affected in both epithelial and macrophage cell lines. This phenotype is in line with a reported role of OmpA proteins in facilitating bacterial survival within host cells [46], [48], [51], [53]. Importantly, OmpA-like proteins share conserved transmembrane domains but are characterized by extremely variable extracellular domains, which are unique to each pathogen, and confer specific functions [50]. OmpA was predicted to have 4 unstructured loops exposed at the bacterial surface. By replacing each loop with a myc tag, we have generated 4 OmpA mutants (OmpA ΔL1, ΔL2, ΔL3 and ΔL4) and showed that loop 1 is essential to confer invasiveness to E. coli ectopically expressing the OmpA mutants. Accordingly, a specific antibody against loop 1 effectively blocks OmpA function. Bioinformatics analysis indicated the presence of 2 additional OmpA-like proteins in the Coxiella genome, CBU_0307 and CBU_0936, sharing with OmpA a good degree of homology at the level of the transmembrane OmpA-like domain but no significant homology in the 4 extracellular loops. In line with these observations, CBU_0307 and CBU_0936 failed to produce internalization phenotypes when mutated by transposon insertions. Finally, experiments aiming at blocking potential OmpA interactions with a cognate receptor, effectively blocked Coxiella internalization, indicating the presence of an interacting partner at the surface of host cells, which remains to be identified. Using Galleria mellonella larvae as a surrogate in vivo model system we could reproduce the OmpA mutant phenotypes observed in cultured cells. Indeed, only wt Coxiella and the control mutant Tn1832 were able to induce the formation of nodules that were abundantly colonized by Coxiella and disrupt the organization of Galleria peripheral organs. The OmpA mutant Tn208 induced a milder formation of nodules that presented few, isolated bacteria. Accordingly, larvae infected with the OmpA mutant survived Coxiella infections longer than those infected with the control mutants.
In summary, multi-phenotypic screening of host/pathogen interactions is an efficient method for the study of infectious diseases. Here we have applied this method to Coxiella infections and identified a bacterial protein that is essential for Coxiella internalization within non-phagocytic cells. Understanding how intracellular bacteria adhere to and invade their host is essential to 1) understand the cell biology of infection and identify the candidate targets of anti-infectious molecules and 2) to develop targeted vaccines. Of note, bacterial OmpA proteins are considered as new pathogen-associated molecular patterns (PAMPs) and are among the most immuno-dominant antigens in the outer membrane of Gram-negative bacteria [50], [76]. Our laboratory currently investigates the possibility of using OmpA to develop a synthetic vaccine against Q fever.
Strains used in this study are listed in Fig. S9. Escherichia coli strains were grown in Luria-Bertani (LB) medium supplemented with ampicillin (100 µg/ml), kanamycin (50 µg/ml) or chloramphenicol (30 µg/ml) as appropriate. Coxiella burnetii NMII and transposon mutants were grown in ACCM-2 [77] supplemented with kanamycin (340 µg/ml) or chloramphenicol (3 µg/ml) as appropriate in a humidified atmosphere of 5% CO2 and 2.5% O2 at 37°C. Cells were routinely maintained in RPMI (Vero, THP-1, J774 and RAW 264.7) or DMEM (A431 and HeLa), containing 10% fetal calf serum (FCS) in a humidified atmosphere of 5% CO2 at 37°C. For experiments, THP-1 were allowed to differentiate into macrophages for 2 days in the presence of 200 nM phorbol myristate acetate (PMA, Sigma).
Hoechst 33258, rabbit anti poly-His, anti mouse and anti-rabbit HRP-conjugated antibodies and Atto-647N phalloidin were purchased from Sigma. Rabbit anti Coxiella NMII antibodies were kindly provided by Robert Heinzen. Synthesis and production of the peptide KKSGTSKVNFTGVTL, as well as the generation of the cognate antibody in rabbit (named anti-OmpA in this study) were performed by Eurogentec, Belgium. Mouse and rabbit IgG conjugated to Alexa Fluor 488 and 555 as well as Prolong Gold antifade mounting reagent were purchased from Invitrogen. Paraformaldehyde was provided by Electron Microscopy Sciences, PA.
Plasmids and primers used in this study are listed in Fig. S9. DNA sequences were amplified by PCR using Phusion polymerase (New England Biolabs) and gene-specific primers (Sigma). To create the plasmid pITR-CAT-ColE1-P311-GFP, the promoter of CBU_0311 (P311) was amplified from Coxiella RSA439 NMII genomic DNA using primers P311-XhoI-Fw and P311-Rv, GFP was amplified from pEGFP-N1 (Clontech) using primers EGFP-Fw and GFP-PITR-Rv, and P1169-CAT-ColE1 was amplified from plasmid pITR-CAT-ColE1 using primers GFP-PITR-Fw and XhoI-PITR-Rv. PCR fragments P311, GFP and P1169-CAT-ColE1 were fused by overlapping PCR. The resulting PCR product was digested with XhoI and ligated to obtain circular pITR-CAT-ColE1-P311-GFP. OmpA32-248 was amplified from Coxiella RSA439 NMII genomic DNA using primers OmpA32-248-BamHI-Fw and OmpA-EcoRI-Rv and cloned into pET28a to obtain pET28a-OmpA32-248. OmpA was amplified from Coxiella RSA439 NMII genomic DNA using primers OmpA-BamHI-shift-Fw and OmpA-EcoRI-Rv and cloned into pET27b to obtain pET27b-OmpA. Plasmids pET27b-OmpAΔL1, pET27b-OmpAΔL2, pET27b-OmpAΔL3 and pET27b-OmpAΔL4 were generated by PCR using pET27b-OmpA as template and primer pairs loop1-myc-HindIII-Fw/loop1-myc-HindIII-Rv, loop2-myc-HindIII-Fw/loop2-myc-HindIII-Rv, loop3-myc-HindIII-Fw/loop3-myc-HindIII-Rv, loop4-myc-HindIII-Fw/loop4-myc-HindIII-Rv. The PCR products were digested with HindIII and ligated to obtain the corresponding plasmids.
C. burnetii RSA439 NMII organisms were electroporated with the pITR-CAT-ColE1-P311-GFP and pUC19::Himar1C9 plasmids [34] using the following setup: 18 kV, 400 Ω, 25 µF. Bacteria were then grown overnight in ACCM-2 supplemented with 1% FBS and the following day 3 µg/ml chloramphenicol were added to bacterial cultures. Bacteria were then amplified for 4 days and plated on solid ACCM-2 for clone isolation. Seven days post-inoculation, colonies were isolated and amplified for 6 days in liquid ACCM-2 supplemented with 3 µg/ml chloramphenicol. The concentration of each isolated mutant was quantified using the PicoGreen (Invitrogen) assay according to manufacturer's instructions. To map transposon insertions, single primer colony PCR was performed on 1 µl of C. burnetii transposon mutant in stationary phase in ACCM-2. The PCR mix contained 1× HF buffer (New England Biolabs), 200 µM dNTPs, 1 µM primer SP3 and 1 U of Phusion polymerase (New England Biolabs). The PCR cycle consisted in initial denaturation (98°C, 1 min), 20 high stringency cycles (98°C, 10 sec; 50°C, 30 sec; 72°C, 90 sec), 30 low stringency cycles (98°C, 10 sec; 30°C, 30 sec; 72°C, 90 sec) and 30 high stringency cycles (98°C, 10 sec; 50°C, 30 sec; 72°C, 90 sec) followed by a final extension at 72°C for 7 min. PCR products were then sequenced at Beckman Coulter Genomics (Stansted, UK) using primer P3. Insertion sites were mapped on the annotated C. burnetii RSA493 NMI genome using MacVector (MacVector Inc.) and recorded in a relational database (FileMaker).
106 GE/ml of bacteria were inoculated in 4 ml ACCM-2 and allowed to grow for 8 days in a humidified atmosphere of 5% CO2 and 2.5% O2 at 37°C. Where needed, 3 µg/ml chloramphenicol were added to bacterial cultures. At the indicated time points bacterial concentrations were evaluated from 100 µl of cultures, using the PicoGreen (Invitrogen) assay according to manufacturer's instructions.
Vero cells were seeded into triplicate 96-wells plates (Greiner Bio one) 2 days prior to infection. Cells were then challenged with C. burnetii RSA439 NMII transposon mutants at an MOI of 100. For each plate, cells in well A1 were left uninfected and cells in wells A2 and A3 were incubated with GFP-C. burnetii RSA439 NMII at multiplicities of infection of 100 and 200. Bacterial contact with cells was promoted by centrifugation (10 min, 400 g, RT) and cells were incubated in a humidified atmosphere of 5% CO2 at 37°C. Unbound bacteria were removed after 1 h of incubation and cells were further incubated in fresh culture medium for 7 days. Plates were analyzed at a 24-hours interval using a TECAN Infinite 200 Pro operated by the Magellan software (TECAN) to monitor the variations of GFP fluorescence associated with the intracellular growth of Coxiella. Raw data were analyzed for background subtraction, normalization and quality control among triplicates using in-house developed methods. Seven days after infection, plates were fixed in 3% paraformaldehyde in PBS at room temperature for 30 minutes, rinsed in PBS and incubated in blocking solution (0.5% BSA, 50 mM NH4Cl in PBS, pH 7.4). Cells were then incubated in Hoechst 33258 diluted 1∶200 in blocking solution for 30 minutes at room temperature, rinsed and incubated in PBS. Images were acquired with an Arrayscan VTI Live epifluorescence automated microscope (Cellomics) equipped with an ORCA ER CCD camera. 6 fields/well of triplicate 96-wells plates were imaged with a 20× objective in the GFP, DAPI and Bright-field channels. Images were then processed and analyzed using CellProfiler. Briefly, the GFP channel was subtracted from the corresponding DAPI channel to avoid false identification of large Coxiella colonies as host cell nuclei, images were thresholded using the Otsu global method and host cell nuclei as well as Coxiella colonies were identified and segmented. The number, form factor and fragmentation of host cell nuclei and the number, form factor, area, perimeter, GFP intensity and compactness of Coxiella colonies were then calculated per object and per image. An average of 14000 cells per condition (infection with a given Coxiella mutant) were thus analyzed. Raw data were processed for background subtraction, normalization and quality control among the 6 fields per well and plate triplicates using in-house developed methods. Data were recorded in a relational database (FileMaker) that allowed clustering of phenotypes according to the annotated transposon insertions.
Cells were fixed in 3% paraformaldehyde in PBS at room temperature for 30 minutes, rinsed in PBS and incubated in blocking solution (0.5% BSA, 50 mM NH4Cl in PBS, pH 7.4). When appropriate, 0.05% saponin was added to the blocking solution for cell permeabilization. Cells were then incubated with the primary antibodies diluted in blocking solution for 1 h at room temperature, rinsed five times in PBS and further incubated for 45 min with the secondary antibodies diluted in the blocking solution. Fluorescent phalloidin was added to the secondary antibodies to label actin, where needed. After labeling, coverslips were mounted using Prolong Gold antifade mounting medium (Invitrogen) supplemented with Hoechst 33258 for DNA staining. For differential labeling, extracellular bacteria or beads were stained using specific antibodies without permeabilizing the cells. Intracellular bacteria or beads were visualized by green and red fluorescence, respectively. Alternatively, a second staining was performed after cellular permeabilization. Secondary antibody labeling using two different fluorochromes (before and after permeabilization) allowed discrimination between adherent extracellular bacteria/beads and those that have been internalized. Samples were analyzed with a Zeiss Axioimager Z1 epifluorescence microscope (Carl Zeiss) connected to a Coolsnap HQ2 CCD camera. Images were acquired alternatively with 63× or 40× oil immersion objectives and processed with MetaMorph (Universal Imaging Corp.). Image J and CellProfiler software were used for image analysis and quantifications. 3D reconstruction and surface rendering were performed using the Osirix software.
Transposon insertions in Dot/Icm core genes were complemented in trans as previously described [19]. Briefly, Vero cells grown on 96-wells plates were either challenged with the transposon mutants alone or in combination with wt Coxiella at a 1∶1 ratio for a total MOI of 100. Bacterial contact with cells was promoted by centrifugation (10 min, 400 g, RT) and cells were incubated in a humidified atmosphere of 5% CO2 at 37°C. Unbound bacteria were removed after 1 h of incubation and cells were further incubated in fresh culture medium for 7 days. Plates were then fixed in 3% paraformaldehyde in PBS at room temperature for 30 minutes, rinsed in PBS and incubated in blocking solution (0.5% BSA, 50 mM NH4Cl in PBS, pH 7.4). Cells were then labeled with the anti-NMII antibody to detect wt Coxiella and with Hoechst 33258 as described above for DNA labeling. 96-well plates were imaged and analyzed essentially as described above for the mutant library screening protocol with the additional acquisition of the TRITC channel to detect and segment wt Coxiella colonies. Object masking was applied using CellProfiler to specifically calculate the area of mutant Coxiella colonies growing within wt Coxiella-occupied PVs.
The terminal deoxyribonucleotidyl transferase-mediated triphosphate (dUTP)-biotin nick end labeling (TUNEL) method was used for detection of DNA fragmentation of nuclei using the In Situ Cell Death Detection Kit TMR (Roche) according to the manufacturer's instructions. Briefly, HeLa cells grown on 96-well plates in triplicate were either left untreated or challenged with the indicated Coxiella strains at an MOI of 100 and incubated at 37°C for 3 days. Cells were then either fixed and permeabilized as described above or incubated with 1 µM staurosporine for 4 hours prior to fixation. Samples were then incubated 1 h at 37°C in the dark with TUNEL reaction mixture. Samples were then washed three times with PBS, and incubated with Hoechst 33258 for DNA staining and with an anti-NMII antibody to detect bacteria in the samples infected with wt Coxiella. 96-wells plates were analyzed with an Arrayscan VTI Live epifluorescence automated microscope (Cellomics) equipped with an ORCA ER CCD camera. 10 fields/well were imaged with a 20× objective in the GFP (bacteria), TRITC (TUNEL), DAPI (nuclei) and Bright-field (cells) channels. Images were then processed and analyzed using CellProfiler. Briefly, the GFP channel was subtracted from the corresponding DAPI channel to avoid false identification of large Coxiella colonies as host cell nuclei, images were thresholded using the Otsu global method and host cell nuclei, Coxiella colonies and fragmented nuclei were identified and segmented. The percentage of fragmented nuclei over the total number of nuclei was then calculated on an average of 6000 cells per condition.
Cells were washed in PBS and fixed with 2.5% glutaraldehyde in Sorensen buffer, pH 7.2 for an hour at room temperature, followed by washing in Sorensen buffer. Fixed samples were dehydrated using a graded ethanol series (30–100%), followed by 10 minutes in graded Ethanol-Hexamethyldisilazane and finally Hexamethyldisilazane alone. Subsequently, the samples were sputter coated with an approximative 10 nm thick gold film and then examined under a scanning electron microscope (Hitachi S4000, at CRIC, Montpellier France) using a lens detector with an acceleration voltage of 20 kV at calibrated magnifications.
Galleria mellonella larvae were fixed overnight in paraformaldehyde 4% in PBS (pH 7.4). Larvae were then rinsed 3 times in PBS and cryo protected by successive incubations in PBS containing increasing concentrations of sucrose (10%, 20%, 30%). Samples were then frozen in isopentane at −80°C using a SnapFrost machine (Excilone). Consecutive 20 µm sections were then obtained from each sample using a Leica CM3050S cryostat. For indirect immuno-fluorescence, sections were permeabilized and blocked in PBS, 10% goat serum, 0.3% Triton X-100 for 1 h at room temperature. Samples were then incubated 48 hours at 4°C with the anti GFP antibody, then rinsed in PBS. Samples were then incubated with the appropriate secondary antibodies, Atto-647N phalloidin and Hoechst 33258 for 24 hours at 4°C. Samples were then washed in PBS and mounted on glass slides for microscopy analysis. Samples were analyzed either with an EVOS microscope (AMG) or with an ApoTome-equipped Zeiss Axioimager Z1 epifluorescence microscope (Carl Zeiss) connected to a Coolsnap HQ2 CCD camera. Images were acquired alternatively with a 10× objective (EVOS) or with a 63× oil immersion objective (Axioimager Z1) and processed with Image J and AxioVision (Carl Zeiss).
Genes cloned into pET27b or pET28a vectors were expressed in E. coli BL21-DE3 star pLysS (Invitrogen). Bacterial cultures were grown at 25°C to mid-exponential phase (OD600 nm = 0.5) and were induced overnight with 400 µM isopropyl-β-D- thiogalactopyranoside (IPTG). For GST expression, E. coli XL1-blue were transformed with pGEX-4T1, grown at 37°C to mid-exponential phase (OD600 nm = 0.5) and induced for 4 h with 1 mM IPTG. Bacteria were harvested by centrifugation, resuspended in lysis buffer (20 mM Tris pH 8, 300 mM NaCl, 5% glycerol, complete anti-protease (Roche)) and lysed with BugBuster (Novagen) following the manufacturer's recommendations. Lysates were then cleared by centrifugation (11 000 g, 20 min, 4°C). Proteins were purified by gravity flow using Ni2+ agarose His-select resin column (Sigma) for His-tagged proteins or glutathione-sepharose (Sigma) for GST. His-tagged and GST proteins were eluted with lysis buffer supplemented with 250 mM imidazole or 25 mM reduced glutathione, respectively.
For Coxiella membrane fractionation, 100 ml of wt Coxiella or Tn208 mutant grown in ACCM-2 for 7 days were pelleted and resuspended in 200 µl 20 mM Tris pH 8 containing 1× Complete protease inhibitor (Roche). For E. coli membrane fractionation, 30 ml of IPTG-induced or non-induced E. coli BL21-DE3 star pLysS pET27b-OmpA, pET27b-OmpAΔL1, pET27b-OmpAΔL2, pET27b-OmpAΔL3 or pET27b-OmpAΔL4 were pelleted and resuspended in 5 ml 20 mM Tris pH 8 containing 1× Complete protease inhibitor (Roche). Bacteria were sonicated using a Branson Sonifier S-450 (6 pulses of 20 s at 40% intensity) and cleared by centrifugation at 10000 g for 5 min at 4°C. Inner membrane proteins were extracted by incubation with sarkosyl (0.5% final concentration) at RT for 15 min. Outer membrane proteins were pelleted by ultracentrifugation (TLA-100, 32000 r.p.m., 30 min, 4°C) and resuspended in 2× Laemmli sample buffer. Insoluble, soluble/sarkosyl-solubilized and outer membrane fractions were resolved by SDS-PAGE and analyzed by Coomassie staining (Sigma) or immunoblotting with anti-OmpA and anti NMII antibodies.
0.5 µm fluorescent red sulfate-modified polystyrene beads (Sigma) were washed three times with 25 mM MES pH 6.1 (MES buffer). The sulfate-modified beads (7.2×109) were then mixed with either 100 µg/ml purified GST or His-OmpA32-248 and incubated at room temperature (RT) for 4 h. The GST- or His-OmpA32-248-coated beads were then washed three times with MES buffer and resuspended in MES buffer containing 1% BSA.
For fluorescent beads internalization assay, 7×107 GST- or His-OmpA32-248-coated beads in DMEM were applied to 1×105 A431 cells seeded onto glass coverslips in 24-well plates and contact was promoted by centrifugation (10 min, 400 g, RT). Cells were incubated in a humidified atmosphere of 5% CO2 at 37°C. Cells were washed three times with PBS and fixed in 4% paraformaldehyde before being processed for immunofluorescence staining. Bacteria internalization assays were performed as follow: 6×106 C. burnetii NMII GE were applied to cells (MOI 100) and contact was promoted by centrifugation (10 min, 400 g, RT). Cells were then fixed in 4% paraformaldehyde before being processed for immunofluorescence staining. For protein blocking experiments, A431 cells were pre-incubated for 1 h at 4°C with either 100 µg/ml of GST or 100 µg/ml of His-OmpA32-248 prior to the internalization assay described above. For antibody inhibition experiments, 6×106 C. burnetii RSA439 NMII GE were incubated at 4°C with increasing concentrations of either naïve rabbit serum or anti-OmpA antibodies (0.1 to 5 µg/ml) prior to the internalization assay described above. Gentamicin protection assays were performed as follow: IPTG-induced and non-induced BL21-DE3 star pLysS pET27b-OmpA, pET27b-OmpAΔL1, pET27b-OmpAΔL2, pET27b-OmpAΔL3 or pET27b-OmpAΔL4 were diluted to OD600 nm = 0.5 in DMEM and applied to cells at an MOI of 10 and contact was promoted by centrifugation (10 min, 400 g, RT). Following 1 h incubation at 37°C/5% CO2, cells were washed 5 times with PBS and incubated for 2 h with DMEM supplemented with 100 µg/ml gentamicin. Cells were then washed 5 times with PBS, lysed with 1 ml 0.1% Triton X-100 in ddH2O and serial dilutions were plated onto LB agar plates supplemented with the appropriate antibiotic for colony-forming units (CFU) assessment. Internalization frequency was determined as the number of CFU surviving the gentamicin challenge out of the total bacterial input. The results are representative of at least three independent experiments.
|
10.1371/journal.pgen.1006986 | Interplay between microtubule bundling and sorting factors ensures acentriolar spindle stability during C. elegans oocyte meiosis | In many species, oocyte meiosis is carried out in the absence of centrioles. As a result, microtubule organization, spindle assembly, and chromosome segregation proceed by unique mechanisms. Here, we report insights into the principles underlying this specialized form of cell division, through studies of C. elegans KLP-15 and KLP-16, two highly homologous members of the kinesin-14 family of minus-end-directed kinesins. These proteins localize to the acentriolar oocyte spindle and promote microtubule bundling during spindle assembly; following KLP-15/16 depletion, microtubule bundles form but then collapse into a disorganized array. Surprisingly, despite this defect we found that during anaphase, microtubules are able to reorganize into a bundled array that facilitates chromosome segregation. This phenotype therefore enabled us to identify factors promoting microtubule organization during anaphase, whose contributions are normally undetectable in wild-type worms; we found that SPD-1 (PRC1) bundles microtubules and KLP-18 (kinesin-12) likely sorts those bundles into a functional orientation capable of mediating chromosome segregation. Therefore, our studies have revealed an interplay between distinct mechanisms that together promote spindle formation and chromosome segregation in the absence of structural cues such as centrioles.
| When cells divide, they must assemble a microtubule-based structure called a spindle on which the chromosomes are segregated. While in most cell types the microtubules that comprise the spindle are nucleated and organized by centriole-containing centrosomes, female reproductive cells (oocytes) of many species lack centrioles and therefore spindles in these cells assemble using unique mechanisms. Using C. elegans as a model system, we set out to identify factors required for acentriolar spindle assembly in oocytes and found two microtubule motor proteins necessary for this process. When these motors are depleted in oocytes, microtubules fail to form stable bundles during spindle assembly, resulting in severely aberrant spindles. However, we were surprised to find that these disorganized microtubules were then able to reorganize into a spindle capable of segregating chromosomes during anaphase, revealing a second mechanism that can act to bundle and organize spindle microtubules. Studies of this unique anaphase reorganization phenotype then enabled us to uncover new proteins contributing to spindle organization, and to gain insights into the mechanisms driving chromosome segregation in this important cell type. The work presented here therefore deepens our understanding of the molecular mechanisms of acentriolar spindle assembly and chromosome segregation in oocytes.
| During mitosis, centriole-containing centrosomes duplicate and then move to opposite ends of the cell where they nucleate microtubules and form the spindle poles. However, oocytes of many species lack centrioles, and as a result, spindles in these cells assemble using a different pathway [1]. We are interested in understanding the molecular mechanisms underlying this unique, acentriolar pathway of spindle assembly.
Using C. elegans oocyte meiosis as a model, we recently found that acentriolar spindle assembly in this system proceeds by: 1) formation of a cage-like structure comprised of prominent bundles of microtubules that are constrained by the disassembling nuclear envelope, 2) reorganization of this structure such that the microtubule minus-ends are sorted away from the chromosomes to the periphery of the array where they are focused into multiple nascent poles, and then 3) coalescence of these poles until bipolarity is achieved [2]. During this process, the microtubule bundles project into the space near the homologous chromosome pairs (bivalents) and then begin to form lateral associations with them, an interaction that is maintained through anaphase. These lateral associations contribute to the alignment of bivalents at metaphase [3]. Subsequently, during anaphase, spindle morphology changes: the spindle shrinks and rotates 90 degrees such that it is perpendicular to the cell cortex, the spindle poles broaden, and the microtubule bundles reorganize into a parallel array, creating open channels [4–6]. Anaphase then proceeds through two phases, with chromosome-to-pole movement through the open channels in Anaphase A, and spindle elongation driving chromosomes further apart in Anaphase B [7]. This unique mode of chromosome segregation is kinetochore-independent [8], and instead relies on a complex of proteins containing AIR-2 (Aurora B kinase) that concentrates at the center of each bivalent [9, 10], forming a ring-like structure (the “midbivalent ring”). These rings localize to chromosomes during spindle formation [3] and then are removed from chromosomes in anaphase, remaining in the channels in the center of the spindle [8].
In this system, numerous factors have been shown to contribute to different aspects of acentriolar spindle assembly (e.g., microtubule length regulation and spindle pole formation), including MEI-1/2 (katanin), KLP-7 (MCAK), ASPM-1, dynein, and others (reviewed in [11]). Moreover, the kinesin-12 family member KLP-18 promotes spindle bipolarity [3, 12, 13], by sorting microtubule bundles and forcing the minus-ends outward where they can be organized into the spindle poles [2]. However, the factors that are required for bundling microtubules and stabilizing these bundles in the absence of centrioles are unknown. Furthermore, little is known about how the acentriolar anaphase spindle is organized and stabilized. During mitosis in C. elegans, the centralspindlin complex of CYK-4 (RhoGAP) and the kinesin-6 family member ZEN-4 (MKLP1) binds to and bundles antiparallel microtubules in the midzone of the anaphase spindle, providing stability to the structure [14]. The centralspindlin complex also localizes to the meiotic anaphase spindle in oocytes, and although this complex is required for the completion of cytokinesis and polar body formation [15], depletion has no effect on anaphase spindle morphology [8]. Another component important for anaphase spindle organization during C. elegans mitosis is the microtubule bundling protein SPD-1 (PRC1), which is required for proper central spindle structure [16–18], and localizes to the midzone in mitosis [16] and meiosis [19, 20]. However, depletion of SPD-1 from C. elegans oocytes does not produce an obvious phenotype [8], making it unclear if this protein functions during oocyte meiosis.
Now, we have identified KLP-15 and KLP-16, members of the conserved kinesin-14 family of minus-end-directed kinesins [21], as factors required for microtubule bundling and organization during acentriolar spindle assembly in C. elegans oocytes; in the absence of these proteins, spindles are unable to maintain stable microtubule bundles and as a result are severely aberrant at metaphase and early anaphase. However, despite these defects, microtubules are then able to reorganize into a spindle capable of mediating chromosome segregation during anaphase. Importantly, this unexpected spindle reorganization phenotype enabled us to gain new insights into the mechanisms underlying anaphase spindle organization and chromosome segregation during acentriolar meiosis, uncovering previously unidentified roles for SPD-1 and KLP-18 in anaphase. These studies have therefore revealed a role for minus-end kinesins in acentriolar spindle assembly in C. elegans oocytes and highlight how the interplay of multiple mechanisms functions to promote the formation of a bipolar spindle that is capable of faithfully segregating chromosomes in this specialized type of cell division.
To identify proteins required for acentriolar spindle assembly in C. elegans oocytes, we performed a targeted RNAi screen of genes previously reported to be embryonic lethal, visually screening for spindle defects in a strain expressing GFP::tubulin and GFP::histone [3]. This screen identified KLP-15 and KLP-16, two highly homologous minus-end-directed kinesins (91.1% identical in amino acid sequence and 93% identical in mRNA sequence) of the kinesin-14 family [22]. We observed the same spindle defects when we used the RNAi library clone annotated as targeting klp-15 as we did when we used the klp-16 clone, likely because both RNAi constructs target both transcripts due to the high sequence similarity between them (Figs 1A and S1). Consistent with this interpretation, our RNAi conditions caused high embryonic lethality (95.6%; S2A Fig), whereas single deletion mutants of either motor were largely viable; klp-15(ok1958) had 14% embryonic lethality and a new deletion mutant we generated, klp-16(wig1), had 2.6% embryonic lethality. Moreover, treatment of klp-15(ok1958) with the RNAi clone annotated as targeting klp-15 and treatment of klp-16(wig1) with the clone annotated as targeting klp-16 caused high embryonic lethality (89.8% and 94.6%, respectively; S2A Fig), consistent with the interpretation that both proteins are expressed, and that each RNAi library clone can target both proteins. These results suggest that KLP-15 and KLP-16 are redundant, and therefore, we refer to these proteins collectively as KLP-15/16 (in describing assays and results where we cannot distinguish between them).
Previous work from other groups had suggested a role for KLP-15/16 in the segregation of meiotic chromosomes, because in addition to embryonic lethality, inhibition of these proteins resulted in phenotypes such as polar body defects, a high incidence of male progeny (which results from non-disjunction of the X chromosome in the oocyte) and multiple female pronuclei in the one-cell stage embryo [21, 23–28]. However, a careful analysis to determine the causes of these segregation errors had not been reported. Therefore, we performed detailed live and fixed imaging of oocytes following klp-15/16(RNAi). After nuclear envelope breakdown (NEBD) is initiated in control oocytes, microtubules form prominent bundles that organize into a cage-like structure (Fig 1B, Fig 1C, arrows; S1 Movie). The microtubules are then sorted such that the minus-ends of the microtubule bundles, visualized by ASPM-1 [29–31], are on the periphery of the array, where they are organized into multiple nascent poles that coalesce until bipolarity is achieved (Figs 1B, 1C, 1D and S2C; S1 and S3 Movies [2]). In klp-15/16(RNAi) oocytes, although the microtubule cage appears to initially form normally (Fig 1C bottom; arrows), the microtubule bundles are not maintained and begin to fall apart, resulting in a disorganized array that lacks focused nascent poles. Then, the array collapses into a “microtubule ball” comprised of short microtubules surrounding the chromosomes (Figs 1B, 1C and S2C; S1 and S2 Movies). ASPM-1 localization appears largely diffuse at both the microtubule “array” and “ball” stages (Figs 1C, 1D and S3A; S4 Movie), suggesting that microtubule minus-ends are distributed throughout these structures. However, since there are examples of spindles where ASPM-1 does have areas of slight concentration within the microtubule ball structures (S3A Fig, arrowheads), it is possible that these spindles have some small degree of microtubule organization.
Our analysis therefore demonstrates that depletion of KLP-15/16 affects the early stages of spindle assembly, resulting in structures that lack prominent microtubule bundles past the cage stage. These two kinesins likely function redundantly at this stage, since spindles appeared normal in the klp-15(ok1958) and the klp-16(wig1) single mutants (Figs 2C and S3C). Interestingly, despite the severe spindle defects in oocytes following klp-15/16(RNAi), we observed that mitotic spindles in the one-cell stage embryo formed normally (Fig 1E; S5 Movie), suggesting that these proteins are not essential when centriole-containing centrosomes are present.
Given the strong phenotype observed in oocytes upon KLP-15/16 depletion, we next assessed the localization of these proteins. To this end, we generated a peptide antibody against the N-terminal 20 amino acids of KLP-16; because there is only one amino acid different between KLP-15 and KLP-16 in this region, this antibody likely recognizes both proteins (S1 Fig). Indeed, this antibody recognizes a band corresponding to the size of both proteins in a Western blot of control worms, and this band was greatly reduced when RNAi was performed using a clone from the RNAi library that had been annotated as targeting klp-16 (Fig 2A). Furthermore, this band was also reduced when klp-15(ok1958) worms were treated with the clone annotated as targeting klp-15, and when klp-16(wig1) worms were treated with the klp-16 clone (Fig 2A), further demonstrating the specificity of the antibody and confirming that RNAi treatment using either of the RNAi library clones for klp-15 or klp-16 targets both proteins.
Using this antibody, we found that KLP-15/16 localize in the cytoplasm prior to NEBD and then begin to accumulate on microtubules during the multipolar stage, becoming uniform on the spindle throughout metaphase and anaphase (Fig 2B). This localization is specific and likely represents both proteins, as it is abolished following RNAi depletion of KLP-15/16, but it is the same as wild-type in the klp-15(ok1958) mutant and the klp-16(wig1) mutant (Fig 2C). We observed a similar localization pattern in a worm strain expressing KLP-16::GFP from the endogenous locus, but this strain also revealed clear localization of KLP-16 to microtubule bundles at the cage stage (Figs 2D and S4B) and to the centrosomes and mitotic spindle microtubules in one-cell stage embryos (Figs 2E and S4C), patterns that were not apparent with the KLP-15/16 antibody (quantification in Materials and Methods). This discrepancy is likely due to variability with the fixed imaging, because even for the stages where we could observe robust staining of spindle microtubules using the KLP-15/16 antibody, not all spindles were stained. Furthermore, when we stained oocytes from the KLP-16::GFP strain with a GFP antibody, we saw similar variability (quantification in Materials and Methods) (S4D and S4E Fig), despite the fact that when this strain was viewed live, every oocyte/embryo imaged had bright KLP-16 fluorescence that appeared to mark spindle structures (S4A Fig). Therefore, we conclude that KLP-15/16 localize to spindle structures through all stages of oocyte spindle assembly, and also to microtubules in mitotic one-cell stage embryos.
Although KLP-15/16 localize to microtubule bundles at the cage stage (Figs 2D and S4B), these motors are not necessary for the formation of these bundles (Fig 1C), suggesting that they act redundantly with other microtubule associated factors at this initial stage of spindle assembly. Similarly, KLP-15/16 localize to spindle microtubules in mitotic embryos (Figs 2E, S4C and S4E), but they are not necessary for the assembly of these spindles (Fig 1E), potentially because centrosomes provide the primary source of microtubule organization in these cells. Taken together, the phenotype of klp-15/16(RNAi) and the localization pattern of these proteins support a role for KLP-15/16 in acentriolar meiotic spindle assembly where they likely stabilize the microtubule bundles formed during the cage stage. These stabilized microtubule bundles can then be sorted by other molecular motors such as KLP-18 to achieve bipolarity [2].
While filming klp-15/16(RNAi) oocytes, we made the surprising observation that although spindle assembly was severely aberrant, microtubules were often able to reorganize into a bundled structure capable of segregating chromosomes, suggesting the presence of a second, KLP-15/16-independent mechanism for bundling microtubules that operates during anaphase (S6 Movie). Therefore, we used markers of anaphase progression to carefully examine anaphase in klp-15/16(RNAi) oocytes, to better understand this mechanism.
During meiosis in wild-type oocytes, separase (SEP-1) relocates from the kinetochore to the midbivalent ring complex at anaphase onset and then disappears from the rings by late anaphase (Fig 3A) [4]. Aurora B (AIR-2), a component of the midbivalent ring complex, is removed from the rings at anaphase onset and relocalizes to the microtubules by mid anaphase (Fig 3A) [32, 33]. Therefore, we used these markers to stage oocytes following klp-15/16(RNAi), allowing us to distinguish pre-anaphase (AIR-2 on the ring structures, SEP-1 on kinetochore), early anaphase (both proteins in rings), and mid/late anaphase (AIR-2 on microtubules, SEP-1 gone).
Using these markers to stage klp-15/16(RNAi) spindles, we found that the microtubule ball configuration observed in our imaging (Fig 1B, 1C and 1D) represents a mixture of metaphase and early anaphase (Fig 3A), although the structures in early anaphase tended to be smaller (S3B Fig). (Note that the spindles that we used for our linescan analysis in Fig 1D all were within the range of volumes observed for metaphase spindles (S3B Fig)). This analysis suggests that the metaphase disorganized microtubule array begins to shrink in preparation for anaphase, similar to what happens in wild-type spindles [6]. Following this stage, when AIR-2 has relocalized to the microtubules and SEP-1 is gone in mid/late anaphase, microtubules reform into a bundled structure and chromosomes are able to segregate into distinct masses (Fig 3A). Despite this anaphase spindle reorganization, we observed segregation errors such as lagging chromosomes and segregation of chromosomes along different axes (Fig 3B and 3C; S6 Movie) that resulted in high levels of aneuploidy in MII oocytes (Fig 3D), likely due to the severely aberrant metaphase spindles that were unable to align chromosomes (Fig 1B and 1C). However, the high rates of aneuploidy also raised the possibility that the microtubules in the klp-15/16(RNAi) anaphase spindles may not be organized like in wild-type spindles, where a high concentration of microtubule minus-ends are found at the spindle poles. To test this hypothesis, we assessed the localization of ASPM-1 and KLP-18 (a kinesin that is enriched at the poles of wild-type oocyte spindles [12]), and found that the microtubules of anaphase spindles in klp-15/16(RNAi) oocytes, although bundled, are likely not organized properly, since ASPM-1 and KLP-18 are distributed throughout the entire spindle instead of being enriched at the poles (Fig 3B and 3C). Although it is possible that microtubules within the bundles are properly organized and that the signals to localize ASPM-1 and KLP-18 are defective, we favor the interpretation that the secondary mechanism we identified bundles microtubules without first sorting them, resulting in bundles comprised of microtubules of mixed polarity.
Next, we wanted to uncover factors that are responsible for bundling anaphase microtubules in klp-15/16(RNAi) oocytes. Two possible candidates are the centralspindlin complex (comprised of ZEN-4 and CYK-4) and SPD-1, since these proteins have clearly defined roles during anaphase in C. elegans mitosis [14, 16, 18, 34] and have been shown to concentrate at the midzone of the anaphase spindle in C. elegans oocytes [8, 19, 20]. Therefore, we assessed the localization of SPD-1 and ZEN-4 at high resolution on C. elegans oocyte spindles. As expected from previous studies, we found that neither centralspindlin (ZEN-4) nor SPD-1 localize to metaphase spindles in control oocytes (Fig 4A). However, during anaphase, ZEN-4 and SPD-1 both become enriched in a short region at the center of the spindle (Fig 4A), with similar though non-identical localization (Fig 4C). Following klp-15/16(RNAi), we observed a similar pattern, with neither ZEN-4 nor SPD-1 present on the disorganized spindle structures prior to anaphase, but then prominent localization on the bundled microtubules between the sets of segregating chromosomes during anaphase (Fig 4B). This localization was clear even in spindles with multiple sets of segregating chromosomes, where the bundles were not all oriented along the same axis (Fig 4B, bottom zoom). Therefore, because centralspindlin and SPD-1 both localize to microtubule bundles following klp-15/16(RNAi), these factors are in a location where they could potentially contribute to the anaphase-bundling mechanism we identified.
To test this hypothesis, we assessed a potential functional role for these proteins in anaphase microtubule bundling. In previous work, co-depletion of ZEN-4 and SPD-1 did not affect anaphase spindle structure [8], suggesting that these proteins may not play a role in oocytes. However, our studies have revealed a mechanism that operates in parallel with KLP-15/16 (since KLP-15/16 are normally present on anaphase microtubules, Fig 2B and 2D). Thus, we expect that single depletion of this putative anaphase bundling factor may have only a mild (or no) anaphase phenotype, but that depletion of KLP-15/16 in combination with the secondary factor would completely abolish anaphase bundling. Therefore, we tested each candidate by single depletion and also by co-depletion/inhibition with KLP-15/16 (Fig 5A), and then scored microtubule bundling in mid/late anaphase (using SEP-1 and AIR-2 as markers to stage anaphase, as before; Fig 5B). In addition to microtubule bundling, we also assessed chromosome segregation as a functional readout for anaphase spindle organization, by scoring whether chromosomes were able to segregate into distinct masses (Fig 5B).
Using these assays, we found that both single and double inhibition/depletion of ZEN-4 and SPD-1 had little effect on anaphase microtubule bundling and chromosome segregation (Fig 5A and 5B), consistent with a previous study [8]. Moreover, oocytes where both ZEN-4 and KLP-15/16 were inhibited/depleted appeared similar to klp-15/16(RNAi) alone, with most spindles containing bundled microtubules that were able to segregate chromosomes. However, we found that co-depletion of KLP-15/16 and SPD-1 largely abolished anaphase microtubule bundling and chromosome segregation (Fig 5A and 5B) and resulted in spindles with shorter microtubule lengths (Fig 5C). Furthermore, we observed an increase in the percentage of embryos with a single large polar body and no maternal pronucleus under these conditions, suggesting that the meiotic divisions lacked a functional spindle on which DNA could segregate (Fig 6A, 6B and 6C). Interestingly, our estimations of spindle microtubule lengths revealed that the microtubules in the spd-1(RNAi) condition were somewhat longer than microtubules in the control (Fig 5C), suggesting that SPD-1 may perform a subtle role in regulating spindle length in anaphase. This observation is reminiscent of studies of mitotic anaphase in C. elegans, where an SPD-1 mutant displays larger distances between segregating chromosomes than wild-type embryos, suggesting that the microtubule crosslinking activity of SPD-1 can act to slow the rate of spindle midzone elongation [35]. Taken together, these data highlight a previously unknown role for SPD-1 on acentriolar anaphase spindles.
Given this finding, we more carefully assessed the dynamics of SPD-1 loading onto the spindle during anaphase. Live imaging of control oocytes expressing SPD-1::GFP and mCherry::histone revealed that SPD-1 begins to load onto the spindle between segregating chromosomes shortly after spindle rotation and then continues to accumulate as anaphase progresses (Fig 4D; S7 Movie), consistent with previous studies [19, 20]. Similar to control oocytes, following depletion of KLP-15/16, SPD-1 loads onto microtubules in early anaphase, at the microtubule ball stage (Fig 4D; S7 Movie). Subsequently, as SPD-1 accumulates on the spindle, prominent bundles begin to form (Fig 4D; S7 Movie). This localization pattern, in combination with our functional analysis, is consistent with the interpretation that loading of SPD-1 in early anaphase provides a secondary bundling activity that provides spindle stability and allows for chromosome segregation.
Our SPD-1::GFP imaging also revealed that when SPD-1 loads onto microtubules in klp-15/16(RNAi) oocytes, the forming bundles start out randomly oriented but are then restructured into a largely parallel array (Fig 4D; S7 Movie). Therefore, in addition to SPD-1 bundling microtubules, there is another mechanism working to reorganize these newly formed microtubule bundles into a functional orientation along which chromosomes are able to segregate. One candidate factor that could provide this function is KLP-18, since this motor sorts microtubule bundles during spindle assembly [2], and is present on anaphase spindles following klp-15/16 depletion (Fig 3C). It is currently unknown whether KLP-18 also functions during anaphase, because the requirement for this protein earlier during spindle assembly has made it difficult to assess an anaphase-specific role; in klp-18 mutants or RNAi, chromosomes do not segregate into distinct groups in anaphase (Fig 5D) because the spindles are monopolar prior to anaphase onset [4, 12]. However, the KLP-15/16 depletion phenotype offers a unique opportunity to address this question, since this condition has revealed a sorting activity that operates specifically during anaphase to generate parallel arrays of microtubule bundles. Notably, this activity does not require that the microtubules have been sorted previously; following klp-15/16(RNAi), the microtubule bundles start out randomly oriented (Fig 4D) yet they can still be organized into a parallel array. Therefore, this feature allowed us to explore a potential role for KLP-18 during anaphase by co-depleting it with KLP-15/16.
To determine if KLP-18 could be required for this anaphase reorganization activity, we depleted KLP-15/16 in a KLP-18 mutant, klp-18(tm2841), which results in a predicted early stop that is thought to eliminate KLP-18 function [2]. We then stained the spindles for SEP-1 and AIR-2 to stage them as before, to determine if microtubules were able to reorganize into spindles capable of mediating chromosome segregation in late anaphase, as they do following klp-15/16 depletion in the wild type strain (Figs 3A and 5A). Notably, we found that depletion of KLP-15/16 in the klp-18(tm2841) mutant results in a complete failure of microtubule reorganization and chromosome segregation in late anaphase (Fig 5D), despite the fact that SPD-1 was still able to target to the microtubules (Fig 5E) and that the early anaphase configuration appeared similar to KLP-15/16 depletion in the wild-type strain (Figs 3A and S5). These results therefore suggest that KLP-18 could provide the anaphase spindle reorganization activity that we observe in the klp-15/16(RNAi) condition. Although we cannot completely rule out the possibility that the metaphase defect in klp-18 mutant oocytes prevents the microtubule reorganization that normally occurs following KLP-15/16 depletion, we think that our data are at least suggestive that KLP-18 provides this activity during anaphase and that it may therefore have an anaphase role in wild-type spindles. Taken together, we therefore propose that two complementary activities facilitate the reorganization of anaphase spindle microtubules following KLP-15/16 depletion: 1) SPD-1 loads in early anaphase to generate prominent microtubule bundles of mixed polarity, and 2) KLP-18 acts to orient these bundles into a parallel array that is capable of segregating chromosomes.
Finally, we wanted to further investigate the mechanism of chromosome segregation during KLP-15/16-independent anaphase. During wild-type meiosis, microtubule bundles run along the sides of chromosomes prior to anaphase onset. These lateral associations remain in place during anaphase, creating channels that the chromosomes reside in as they move towards spindle poles [4] and then spindle elongation drives chromosomes further apart [7, 8]. Given that microtubules are completely disorganized prior to anaphase onset following KLP-15/16 depletion and, unlike wild-type spindles, have no discernable lateral associations with chromosomes, we wanted to determine what types of microtubule-chromosome contacts were established during anaphase to facilitate segregation.
First, we asked if anaphase spindles in klp-15/16(RNAi) oocytes are able to form any channels that are analogous to those observed in wild-type oocytes. To this end, we stained spindles for SUMO, to mark the ring structures [36], and SPD-1, to mark anaphase microtubule bundles. In control spindles, each channel is comprised of a pair of separating chromosomes with a ring in between, and SPD-1 marks the microtubule bundles adjacent to the ring. Therefore, line scans of these components in control spindles show an alternating pattern of SUMO/SPD-1 and SUMO/microtubules across the channels (Fig 7A). In klp-15/16(RNAi) oocytes, we found similar alternating patterns of these markers in a significant number of spindles (12/18 klp-15/16(RNAi) spindles examined; Fig 7A); showing that the spindles are capable of forming microtubule channels during anaphase. Importantly, we also observed microtubules associating laterally with the segregating chromosomes (Fig 7B arrows and Fig 7C; these associations were seen in 22/31 klp-15/16(RNAi) spindles) suggesting that this type of association can be established in anaphase, even if these associations are not in place at anaphase onset.
Despite the fact that channels can form during anaphase in klp-15/16(RNAi) oocytes, we also found that some rings appeared to be on the periphery of the spindle, demonstrating that not all separating chromosomes end up in a channel with a ring in the center (Fig 7A, arrowheads). To gain insight into this variability, we looked earlier in anaphase before the microtubules were reorganized into bundles. In early anaphase spindles, at the “microtubule ball” stage when homologous chromosomes first begin to come apart, we observed some rings embedded in the microtubule ball close to the separating chromosomes, but also some rings towards the periphery of the structure (Fig 7B). Subsequently, when the microtubules are bundled and aligned into parallel arrays, rings can be seen both in channels between segregating chromosomes and also completely outside of the reorganized spindle (Fig 7B). This is likely due to the fact that microtubule bundling and reorganization are occurring as chromosomes begin to come apart. This results in some chromosomes and rings becoming organized within channels while others are not. This behavior may also contribute to the presence of lagging chromosomes in these spindles (Figs 3B, 5A and 7A). Therefore, while it is unlikely that the complete formation of a ring-containing channel is essential for chromosome segregation, the fact that lateral associations are established suggests that they could contribute to segregation in this context.
Taken together, our data have revealed two distinct mechanisms that act to bundle microtubules in acentriolar spindles. Prior to spindle assembly, KLP-15/16 localize diffusely throughout the cytoplasm, which is in contrast to kinesin-14s from other organisms that have been shown to be sequestered in the nucleus [37, 38]. Then, as acentriolar spindles begin to form, KLP-15/16 load onto microtubule bundles during the cage stage, stabilizing them to facilitate spindle assembly (Fig 8). This proposed function is consistent with previous studies of kinesin-14s in other organisms, which demonstrated that this family of kinesins is required for acentriolar spindle formation and localize uniformly to acentriolar spindle microtubules [30, 39–41]. However, while depletion of kinesin-14s in other organisms predominantly results in defects such as unfocused poles and splayed microtubules, depletion of KLP-15/16 in C. elegans oocytes completely prevents bipolar spindle formation and abolishes microtubule bundling prior to anaphase, implicating these proteins in the stabilization of microtubule bundles comprising the acentriolar meiotic spindle. Furthermore, unlike most other organisms where kinesin-14s perform essential roles in mitosis [40, 42–44], KLP-15/16 are not required for mitotic spindle formation in C. elegans, suggesting a unique role for KLP-15/16 that is specific to acentriolar spindle assembly; this finding is also reminiscent of studies in Drosophila, where inhibition of the kinesin-14 Ncd has a much more severe phenotype in oocytes than it does in mitosis [38, 42, 45]. The presence of the “microtubule ball” comprised of short microtubules that ultimately forms prior to anaphase following klp-15/16(RNAi) raises the intriguing possibility that the function of KLP-15/16 could be to stitch together short microtubules into longer bundles that can then be sorted and organized into a bipolar spindle. Since kinesin-14s contain a motor domain in the C-terminus and a microtubule binding domain in the N-terminus [37, 43, 46], and it has been reported that this class of kinesins can stabilize and cross-link microtubules in a parallel configuration [47], it is possible that these motors could contribute to this stitching activity. This interpretation is consistent with a previous electron microscopy study, which reported the presence of many short microtubules in a partial reconstruction of a C. elegans oocyte spindle [48], and also with a study in Xenopus egg extracts that demonstrated that meiotic spindles are comprised of tiled arrays of short microtubules [49]; therefore a microtubule stitching activity might be something that is especially important in the context of acentriolar meiosis. Our studies suggest that KLP-15/16 could be factors that organize these short microtubules into longer bundles capable of mediating chromosome congression and segregation.
In addition to revealing an important function for KLP-15/16, our studies have yielded insights into previously unknown mechanisms promoting accurate chromosome segregation during acentriolar meiosis. First, we found that the PRC1-family protein SPD-1 provides a secondary activity that stabilizes microtubule bundles during anaphase (Fig 5A, 5B and 5C). This activity was previously unidentified, as prior depletion of SPD-1 [8], confirmed by our own studies (Fig 5A), failed to reveal an obvious anaphase defect. However, this is likely because KLP-15/16 are present on the anaphase spindle stabilizing the microtubule bundles (Fig 2B), making SPD-1 non-essential until these proteins are depleted. This discovery is similar to findings in fission yeast, where the SPD-1 homolog Ase1 is not essential on its own but provides a backup mechanism for bipolar spindle assembly under conditions where the kinesin-5 motor Cut7 and the kinesin-14 motor Pkl1 are deleted [50].
During mitosis in other organisms, homologs of SPD-1 are known to crosslink antiparallel microtubules [51–53], and our data are consistent with SPD-1 performing a similar function in C. elegans oocytes. During wild-type meiotic anaphase, this protein loads onto the central region of the spindle (Fig 4D), where this putative crosslinking activity could reinforce anaphase spindle structure. Under KLP-15/16 depletion conditions, SPD-1 loads at the microtubule ball stage (Fig 4D), which contains many short microtubules that are likely randomly oriented (Fig 1C); given this configuration, the ability to crosslink antiparallel microtubules would enable SPD-1 to bundle microtubules. These SPD-1-stabilized microtubule bundles could then be sorted and aligned into a parallel array by the action of KLP-18. Therefore, our studies have uncovered a new function for SPD-1 on C. elegans acentriolar spindles, and also represent the first demonstration that PRC1-family proteins play a role during oocyte meiosis. Furthermore, our work also suggests that KLP-18 may be functional during anaphase in these cells, since it appears to organize the microtubule bundles generated by SPD-1 (Fig 5D and 5E), suggesting that this motor not only plays roles during bipolar spindle formation, but may also be required to maintain spindle organization as chromosomes segregate.
Finally, our work has also shed light on the mechanisms driving chromosome segregation in this unique form of anaphase. We found that anaphase spindles in KLP-15/16-depleted oocytes are sometimes able to form channels with lateral microtubule-chromosome associations (Fig 7), despite the lack of microtubule bundles prior to anaphase onset. These data suggest that this form of microtubule-chromosome contact is preferred during anaphase and points to a role for the chromosomes providing significant structural cues for spindle organization as these associations can be established de novo following KLP-15/16 depletion. However, given that these laterally-associated bundles may be comprised of microtubules of mixed polarity (Fig 3B and 3C), it is improbable that they would be able to efficiently facilitate directional chromosome movement (a mechanism we proposed to drive normal Anaphase A [4]). Therefore, we suggest that the primary force driving segregation in the absence of KLP-15/16 is the elongation of these lateral bundles in an Anaphase-B type mechanism.
The discovery that lateral associations are established during anaphase is also interesting since two other types of chromosome-spindle contacts have been proposed to facilitate segregation during wild-type anaphase: 1) elongating microtubules contacting the inside surfaces of separating chromosomes to provide a pushing force [8] and 2) chromosomes contacting the spindle poles, so that outward pole separation can drive segregation in Anaphase B [7]. It is possible that the first type of association contributes to segregation during KLP-15/16-independent anaphase; since not every chromosome ends up in a normal microtubule channel (Fig 7), some microtubules might randomly make contacts with chromosome surfaces and provide a pushing force, contributing to segregation alongside the bundles that are laterally-associated. Indeed, our data are consistent with this idea since we observe non-laterally-associated microtubules in the reorganized klp-15/16(RNAi) anaphase spindles (Fig 7). In contrast, the second model proposed that spindle shrinkage enables the chromosomes to establish a physical tether to a cross-linked network of microtubules and pole proteins; outward sliding of interpolar microtubules would then drive the poles and the tethered chromosomes apart [7]. Our observation that pole proteins KLP-18 and ASPM-1 are distributed throughout klp-15/16(RNAi) spindles both prior to and during anaphase (Figs 1C, 1D, 3B and 3C) makes it difficult to imagine how such a tether would efficiently form in this context, and we therefore speculate that pre-established spindle poles may not be absolutely required to segregate chromosomes in C. elegans oocytes (although we cannot rule out the possibility that other pole proteins exhibit a higher level of organization in these mutant spindles). Moreover, our data also raise the possibility that spindle elongation could be capable of driving segregation even when the polarity of microtubules within the spindle is disrupted, potentially revealing an unusual mode of chromosome segregation that operates in this mutant context.
In summary, our studies have uncovered a crucial role for KLP-15 and KLP-16 in C. elegans acentriolar spindle assembly, revealed a second, anaphase-specific mechanism dependent on SPD-1 operating in parallel to these kinesins, and provided new insights into anaphase spindle organization and chromosome segregation mechanisms during acentriolar meiosis.
In this study, ‘wild-type’ refers to N2 (Bristol) or EU1067 worms grown on NGM/OP50 plates, and ‘control’ refers to the RNAi vector control (L4440).
N2 (Bristol)
ANA065: adeIs1[pMD191, mex-5::spd-1::GFP] II (gift from Marie Delattre)
ANA072: adeIs1[pMD191, mex-5::spd-1::GFP] II; ltIs37[pAA64; pie-1::mCherry::his-58; unc-119(+)] IV (gift from Marie Delattre)
EU716: zen-4(or153) IV (from the CGC). For experiments using zen-4(or153), plates were shifted to 25°C 16–18 hours before dissection and fixation.
EU1067: unc-119(ed3) ruIs32[unc-119(+) pie-1::GFP::H2B] III; ruIs57[unc-119(+) pie-1::GFP::tubulin] (gift from Bruce Bowerman)
OD57: unc-119(ed3) III; ltIs37[pAA64; pie-1::mCherry::his-58; unc-119(+)] IV; ltIs25[pAZ132; pie-1::GFP::tba-2; unc-119 (+)] IV (gift from Arshad Desai)
RB1593: klp-15(ok1958) I. ok1958 is a deletion allele of the last 391 amino acids of KLP-15 (from the CGC)
SMW13: klp-18(tm2841)IV/nT1[qIs51]; unc-119(ed3) ruIs32[unc-119(+) pie-1::GFP::H2B] III; ruIs57[unc-119(+) pie-1::GFP::tubulin] (Wolff et. al., 2016)
SMW15: klp-16(wig1) I. This strain was generated using a CRISPR approach detailed below.
SMW16: Pklp-16::klp-16::GFP (C1971>A–PAM site mutation) I. This strain was generated using a CRISPR approach detailed below.
SMW18: (SMW16 x OD56) Pklp-16::klp-16::GFP (C1971>A–PAM site mutation) I; ltIs37 [(pAA64) pie-1::mCherry::his-58 + unc-119(+)] IV
A CRISPR-based approach [54, 55] was used to generate an endogenously tagged KLP-16::GFP strain (SMW16). Briefly, 27μM recombinant Cas9 protein (IDT) was co-injected with 13.6μM tracrRNA (IDT), 4μM dpy-10 crRNA (5’—GCUACCAUAGGCACCACGAG- 3’) (IDT), 1.34μM dpy-10 repair oligo (Ultramer from IDT; 5’ -CACTTGAACTTCAATACGGCAAGATGAGAATGACTGGAAACCGTACCGCATGCGGTGCCTATGGTAGCGGAGCTTCACATGGCTTCAGACCAACAGCCTAT- 3’), 9.6μM klp-16 crRNA (5’—UGUCUAGUUCAUAGACAUCU- 3’) (IDT); and 136ng/μL ssDNA klp-16 repair template into N2 worms, that were then allowed to produce progeny. Worms from plates containing rollers and dumpys were screened for GFP expression, and homozygous KLP-16::GFP worms were identified by PCR screening. To make the klp-16 repair template, we generated a C-terminal LAP tag using a GBlock (IDT) and Gibson Assembly to create an S-TEV-GFP construct. The tag was then amplified using PCR with primers that contained homology to the klp-16 gene with the final product containing 68 bp of homology upstream of the klp-16 stop codon and 100 bp of homology downstream of the stop codon. ssDNA was generated by asymmetric PCR. SMW16 (KLP-16::GFP) was also crossed with OD56 (mCherry::histone) to generate SMW18: Pklp-16::klp-16::GFP (C1971>A—PAM site mutation) I; ltIs37 [(pAA64) pie-1::mCherry::his-58 + unc-119(+)] IV.
A CRISPR-based approach similar to the one above was used to generate a worm strain with a ~600 bp deletion in the klp-16 locus beginning ~100 bp upstream of the start codon. Essentially the same approach was used as above; the differences being two crRNAs (4.8μM) (5’- AGGCGGAGUUUAAGUUUGAG-3’ and 5’- CUCCUCAAGAAGCGUCACUU-3’) (IDT) (one upstream and one downstream of the klp-16 start codon, respectively), and a ssDNA oligo (4μM) (Ultramer from IDT; 5’-CAGCCATCTCACGCTCCAATTGCGCATTTCTCTCCTCAAGAAGCGTCACTTCTCAAACTTAAACTCCGCCTCTGAAAATTCCCGCCAAATCGGATGGATTAC-3’) were used in the injection mix. The repair Ultramer sequence is homologous to the sequence just upstream and downstream to the two CRISPR cut sites thereby deleting the ~600 base pairs. Worms from plates containing rollers and dumpys were screened by PCR and homozygous mutants were isolated.
Protein domains were determined using PsiPred [56] and Paircoil2 [57]. Protein sequences were analyzed using Clustal Omega [58]. Proline-rich regions of proteins have been shown to bind microtubules [46]. The proline content of amino acids 1–149 is 14% for KLP-15 and 13% for KLP-16.
From a feeding library [26, 59], individual RNAi clones were picked and grown overnight at 37°C in LB with 100μg/ml ampicillin. Overnight cultures were spun down and plated on NGM (nematode growth media) plates containing 100μg/ml ampicillin and 1mM IPTG. Plates were dried overnight. Worm strains were synchronized by bleaching gravid adults and letting the eggs hatch overnight without food. L1s were then plated on RNAi plates and grown to adulthood at 15° for 5–6 days.
Young adult worms grown on control plates or plates containing RNAi-expressing bacteria were transferred to new plates containing either control or RNAi-expressing bacteria and allowed to lay eggs for 24 hours at 15°C before being moved to another fresh plate of either control or RNAi-expressing bacteria. The eggs were allowed to hatch for 24 hours and then the progeny (eggs and hatched worms) were counted. For each parent worm this process was repeated twice, resulting in three days of progeny being counted. For each condition, the progeny of at least 15 worms were scored.
Immunofluorescence was performed by freeze cracking embryos and plunging into -20°C methanol as described [60]. Embryos were fixed for 35–45 minutes, rehydrated in PBS, and blocked in AbDil (PBS plus 4% BSA, 0.1% Triton X-100, 0.02% Na-Azide) for 30 minutes. Primary antibodies were incubated overnight at 4°C. The next day, embryos were washed 3x with PBST (PBS plus 0.1% Triton X-100), incubated in secondary antibody for 1 hour and 15 minutes, washed again as before, incubated in mouse anti-α-tubulin-FITC for 1.5 hours, washed again, and incubated in Hoechst (1:1000 in PBST) for 15 minutes. Embryos were then washed 2x with PBST, mounted in 0.5% p-phenylenediamine, 20mM Tris-Cl, pH 8.8, 90% glycerol or ProLong Gold antifade Mountant (Molecular Probes), and sealed with nail polish; except for the overnight primary, the entire procedure was performed at room temperature. For experiments using the rabbit anti-KLP-16 antibody and staining of SPD-1::GFP with mouse anti-GFP, embryos were blocked in AbDil overnight at 4°C and incubated in primary antibody for 2 hours at room temperature. Primary antibodies used in this study: rabbit anti-ASPM-1 (1:5000, gift from Arshad Desai), rabbit anti-SEP-1 (1:400; gift from Andy Golden), rabbit anti-KLP-18 (1:10,000, gift from O. Bossinger), rabbit anti-ZEN-4 (1:500; gift from Michael Glotzer), mouse anti-SUMO (1:500; gift from Federico Pelisch), mouse anti-GFP (1:200; Invitrogen). Rat anti-AIR-2 was generated by Covance using the C-terminal peptide sequence KIRAEKQQKIEKEASLRNH (synthesized by the Peptide Synthesis Core Facility at Northwestern University), then affinity purified and used at 1:1000. Rabbit anti-KLP-16 was generated by Covance using the N-terminal peptide sequence CMNVARRRSGLFRSTIGAPPK (synthesized by the Peptide Synthesis Core Facility at Northwestern University), then affinity purified and used at 1:2000. Rabbit anti-SPD-1 was generated by Proteintech using the C-terminal peptide sequence CIASSTPSSAKKVLTRRNQFL, then affinity purified and used at 1:1000. Directly conjugated mouse anti-α-tubulin-FITC (DM1α, Sigma) and Alexa-fluor directly conjugated secondary antibodies (Invitrogen) were used at 1:500. All antibodies were diluted in AbDil.
All fixed imaging and high resolution imaging of KLP-16::GFP and KLP-16::GFP;mCherry::histone was performed on a DeltaVision Core deconvolution microscope with a 100x objective (NA = 1.4) (Applied Precision). This microscope is housed in the Northwestern University Biological Imaging Facility supported by the NU Office for Research. Image stacks were obtained at 0.2μm z-steps and deconvolved using SoftWoRx (Applied Precision). All images in this study were deconvolved and displayed as full maximum intensity projections of data stacks encompassing the entire spindle structure, unless stated otherwise. For KLP-16::GFP and KLP-16::GFP;mCherry::histone imaging, live worms were mounted in anesthetic (0.2% tricaine, 0.02% levamisole in M9).
Two-color live imaging was performed using a spinning disk confocal microscope with a 63x HC PL APO 1.40 NA objective lens. A spinning disk confocal unit (CSU-X1; Yokogawa Electric Corporation) attached to an inverted microscope (Leica DMI6000 SD) and a Spectral Applied Imaging laser merge ILE3030 and a back-thinned electron-multiplying charge-coupled device (EMCCD) camera (Photometrics Evolve 521 Delta) were used for image acquisition. The microscope and attached devices were controlled using Metamorph Image Series Environment software (Molecular Devices). Typically, ten to twelve z-stacks at 1μm increments were taken every 20–45 seconds at room temperature. Image deconvolution was done using AutoQuant X3 (Media Cybernetics Inc.). Images were processed using ImageJ. Images are shown as maximum intensity projections of entire spindle structure. Live, intact worms were mounted on 5% agarose, M9 pads in 50% live imaging solution (modified S-basal [50mM KH2PO4,10mM K-citrate, 0.1M NaCl, 0.025mg/ml cholesterol, 3mM MgSO4, 3mM CaCl2, 40mM serotonin creatinine sulfate monohydrate]), 50% 0.1 micron polystyrene Microspheres (Polysciences Inc.), and covered with a coverslip. The spinning disk microscope is housed in the Northwestern University Biological Imaging Facility supported by the NU Office for Research.
For S3 Movie, EU1067 worms were picked into a solution of tricaine (2%) and tetramisole (0.4%), and incubated for ~30 min. Worms were then pipetted onto a 3% agarose pad, covered with a coverslip, and imaged immediately on a DeltaVision Core deconvolution microscope (same as above). Image stacks were obtained at 1μm z-steps at 10 second intervals using 2 × 2 binning, and then deconvolved. Video images are full projections of data stacks.
Fig 1D: Slides made on the same day were imaged within an 8 hour window on a DeltaVision Core deconvolution microscope (see Microscopy section) using the same exposure conditions and times for all slides. In ImageJ, linescans of 154 x 75 pixels (L x W) were performed on 6 z-slice sum projections of representative spindles from control (n = 8) and klp-15/16(RNAi) (n = 9) embryos. In control spindles, the linescans were done along the pole-to-pole axis. In spindles from klp-15/16(RNAi) embryos, linescans were done straight along the x-axis of the image, since these spindles lack a discernable orientation. The average fluorescence intensity for each channel was graphed (solid line) along with the SEM (standard error of the mean) (shaded area) using the ggplot package in R Studio. The y-axes of the graphs are the same between control and experiment for a given channel.
Fig 2B: α-KLP-16 staining for each stage of spindle assembly in wild-type oocytes/embryos. Oocytes in prophase with the nuclear envelope intact were scored as “localized” if the KLP-15/16 signal was primarily cytoplasmic. During all other stages, oocytes were scored as “localized” if the KLP-15/16 signal was colocalized with spindle microtubules. The quantification is as follows: diakinesis 81.8% (n = 22), cage 0% (n = 13), multipolar 31.3% (n = 67), bipolar 64% (n = 114), anaphase 51.9% (n = 79), mitotic spindles 3.7% (n = 27). Although not every spindle is stained, we think that this represents variability with the immunofluorescence procedure and with the antibody (since we see 100% localization of KLP-16::GFP to oocyte spindle microtubules and to mitotic spindles; see S4A Fig).
Fig 2C: α-KLP-16 staining was scored in klp-15/16(RNAi) oocytes. The number of oocytes in which we could discern any spindle staining is as follows: microtubule ball stage 8% (n = 75), anaphase 2.9% (n = 35). α-KLP-16 staining was scored in klp-15(ok1958) oocytes. Oocytes in prophase with the nuclear envelope intact were scored as “localized” if the KLP-15/16 signal was primarily cytoplasmic. During all other stages, oocytes were scored as “localized” if the KLP-15/16 signal was colocalized with spindle microtubules. The quantification is as follows: diakinesis 100% (n = 3), cage 14% (n = 7), multipolar 44% (n = 25), bipolar 72.2% (n = 18), anaphase 38% (n = 21). α-KLP-16 staining was scored in klp-16(wig1) oocytes as above. The quantification is as follows: diakinesis 100% (n = 10), cage 42.9% (n = 7), multipolar 89% (n = 19), bipolar 81.8% (n = 11), anaphase 79.2% (n = 24). As with the control strain (see Fig 2B quantification above), we think that the incomplete staining we observe is due to variability with the procedure and antibody.
Fig 3B: Linescans of control anaphase spindles and anaphase spindles from klp-15/16(RNAi) oocytes stained for ASPM-1 were performed using the arbitrary profile tool in SoftWorx (Applied Precision). A spindle was scored as having staining at poles if the ASPM-1 signal was enriched at two ends of the spindle near the segregating chromosomes. ASPM-1 was enriched at the poles of 21/24 control spindles, but was largely diffuse along spindle microtubules in anaphase of klp-15/16(RNAi) oocytes (only 9/26 spindles could be classified as having any type of ASPM-1 enrichment, and this enrichment was not as strong as in the control spindles).
Fig 3C: Linescans of control anaphase spindles and anaphase spindles from klp-15/16(RNAi) oocytes stained for KLP-18 were performed using the arbitrary profile tool in SoftWorx (Applied Precision). A spindle was scored as having staining at poles if the KLP-18 signal was enriched at two ends of the spindle near the segregating chromosomes. KLP-18 was enriched at poles in 11/12 control spindles but was diffuse in klp-15/16(RNAi) spindles (0/6 had KLP-18 concentrated into poles).
Fig 3D: Aneuploidy in MII embryos was quantified by counting the number of chromosomes in MII in immunofluorescence images of control and klp-15/16(RNAi) embryos. An embryo was scored as ‘aneuploid’ if the number of chromosomes was not 6.
Fig 4A and 4B: ZEN-4 and SPD-1 staining was scored in control spindles and spindles from klp-15/16(RNAi) oocytes. Staining of metaphase and anaphase spindles (staged by AIR-2 localization) was scored for each condition. ZEN-4: Control metaphase 1.6% (n = 63), Control anaphase 90% (n = 20); klp-15/16(RNAi) metaphase 2.9% (n = 134), klp-15/16(RNAi) anaphase 80.3% (n = 66). SPD-1: Control metaphase 9.5% (n = 42), Control anaphase 97% (n = 66); klp-15/16(RNAi) metaphase 1% (n = 100), klp-15/16(RNAi) anaphase 88.2% (n = 85).
Fig 5A and 5B: Quantification of microtubule bundling and chromosome segregation was done using immunofluorescence images of anaphase spindles with SEP-1 gone and AIR-2 relocalized to the microtubules (mid/late anaphase) for the conditions shown. We scored microtubule bundling by eye, looking through the entire z-stack in SoftWorx (Applied Precision). An anaphase spindle was scored as “bundled” if one or more microtubule bundles were discernable. Chromosomes were scored as “segregated” if two or more distinct masses of chromosomes were observed. The simple matching coefficient (SMC) for microtubule bundling and chromosome segregation = 0.82 (n = 251); in other words, 82% of the spindles were scored as microtubules bundled and chromosomes segregated or as no microtubule bundles and no chromosome segregation.
Fig 5C: To approximate anaphase microtubule lengths, we used the measure distances tool in SoftWorx (Applied Precision). Using this tool, a line was manually drawn (point by point) along the most prominent spindle microtubule bundle through the 3D stack of an image to measure its full length.
Fig 6A and 6B: Polar body number and maternal pronuclei number were quantified by scoring live EU1067 worms mounted in anesthetic (0.2% tricaine, 0.02% levamisole in M9) on a Leica DM5500B fluorescent microscope.
Fig 7A: Linescans of control anaphase spindles and anaphase spindles from klp-15/16(RNAi) oocytes stained for tubulin, SUMO, and SPD-1 were performed using the arbitrary profile tool in SoftWorx (Applied Precision). A spindle was scored as having oscillations if one or more instances of alternating MTs/SUMO and SUMO/SPD-1 signal was observed. This analysis was done by examining both single z-slices and max projections of spindles. Oscillations were observed in 9/11 control anaphase spindles and 12/18 klp-15/16(RNAi) anaphase spindles.
Fig 7C: Lateral microtubule associations to chromosomes were scored in control anaphase spindles and anaphase spindles from klp-15/16(RNAi) oocytes. A spindle was scored as having lateral microtubule/chromosome associations if a microtubule appeared to contact and run along the side of a chromosome. This analysis was done by examining both single z-slices and max projections of spindles. We observed clear lateral associations in 31/35 control anaphase spindles and in 22/31 klp-15/16(RNAi) anaphase spindles.
S2C Fig: Live, intact worms expressing GFP::tubulin, GFP::histone (EU1067) fed either control or klp-16(RNAi)-expressing bacteria were anesthetized in 0.2% tricaine, 0.02% levamisole in M9 and viewed on a Leica DM5500B widefield fluorescence microscope. Spindles in embryos in the -1, spermatheca, and +1 positions within the gonad were scored for microtubule organization by eye. A spindle was scored as “multipolar” if it had prominent microtubule bundles that formed more than two organized poles. A spindle was scored as an “array” if the microtubule structure lacked prominent bundles and organized poles. A spindle was scored as “MT ball” if the microtubule structure had collapsed around the chromosomes and lacked prominent bundles and organized poles.
S3B Fig: Spindle volumes were measured using the surfaces tool in Imaris (Bitplane). Using the full 3D image stack, this tool renders a 3D surface based on fluorescence signal (for our analysis, we used the tubulin signal). The volume of this 3D surface is then measured. The volumes of metaphase and early anaphase spindles (staged by SEP-1/AIR-2 localization) from klp-15/16(RNAi) oocytes were measured and compared to the volumes of the spindles used for the linescan measurements in Fig 1D. This analysis allowed us to conclude that the spindles used in our linescans for Fig 1D are within the range of metaphase spindles based on spindle volume.
S4A Fig: Live, intact worms expressing KLP-16::GFP, mCherry::histone (SMW18) were anesthetized in 0.2% tricaine, 0.02% levamisole in M9 and viewed on a Leica DM5500B widefield fluorescence microscope. The localization of KLP-16::GFP was scored in oocytes/embryos in the -1, spermatheca, and +1 positions within the gonad. KLP-16::GFP signal was scored as cytoplasmic if it was absent/reduced in the nucleus. Because the localization of KLP-16::GFP on the spindle looks identical to the organization of GFP::tubulin, we scored localization for the following categories: cage, multipolar, bipolar, and anaphase. The organization of the chromosomes visualized by mCherry::histone was used to identify and better stage the spindles.
S4C Fig: α-GFP staining for each stage of spindle assembly in SMW16 (KLP-16::GFP). Oocytes in prophase with the nuclear envelope intact were scored as “localized” if the α-GFP signal was primarily cytoplasmic. During all other stages, oocytes were scored as “localized” if the α-GFP signal was colocalized with spindle microtubules. The quantification is as follows: diakinesis 100% (n = 2), cage 0% (n = 4), multipolar 76.5% (n = 17), bipolar 96.2% (n = 26), anaphase 83.3% (n = 24), mitotic spindles 100% (n = 5). As with the KLP-15/16 antibody, we think that the lack of staining in all embryos represents variability with the immunofluorescence procedure, since we see robust localization of KLP-16::GFP to microtubules when we visualize this strain live.
Seventy-five EU1067, RB1593, or SMW15 worms were picked off of control, klp-16(RNAi) (EU1067 and SMW15), or klp-15(RNAi) (RB1593) plates onto new, empty (no bacteria) plates. The worms were washed off the plates with cold M9 and transferred to a 1.5ml microcentrifuge tube. Worms were pelleted by spinning at 2000 rpm for 1 minute, and the tube was put on ice for ~2 minutes to allow worms to slow down and form a tight pellet. The M9 was removed and the tube was filled with fresh, cold M9 and mixed. The worms were washed a total of 3 times. After the final wash, as much M9 was removed as possible and 2X SDS sample buffer was added to the remaining worm/M9 mixture and boiled for 10 minutes. Samples were run on a 10% SDS-PAGE gel and blotted. For western analysis, rabbit anti-KLP-16 antibody (1:10,000) and mouse anti-tubulin (1:5000) (Sigma, DM1α) were used.
|
10.1371/journal.pcbi.1003193 | Gause's Principle and the Effect of Resource Partitioning on the Dynamical Coexistence of Replicating Templates | Models of competitive template replication, although basic for replicator dynamics and primordial evolution, have not yet taken different sequences explicitly into account, neither have they analyzed the effect of resource partitioning (feeding on different resources) on coexistence. Here we show by analytical and numerical calculations that Gause's principle of competitive exclusion holds for template replicators if resources (nucleotides) affect growth linearly and coexistence is at fixed point attractors. Cases of complementary or homologous pairing between building blocks with parallel or antiparallel strands show no deviation from the rule that the nucleotide compositions of stably coexisting species must be different and there cannot be more coexisting replicator species than nucleotide types. Besides this overlooked mechanism of template coexistence we show also that interesting sequence effects prevail as parts of sequences that are copied earlier affect coexistence more strongly due to the higher concentration of the corresponding replication intermediates. Template and copy always count as one species due their constraint of strict stoichiometric coupling. Stability of fixed-point coexistence tends to decrease with the length of sequences, although this effect is unlikely to be detrimental for sequences below 100 nucleotides. In sum, resource partitioning (niche differentiation) is the default form of competitive coexistence for replicating templates feeding on a cocktail of different nucleotides, as it may have been the case in the RNA world. Our analysis of different pairing and strand orientation schemes is relevant for artificial and potentially astrobiological genetics.
| The dynamical theory of competing templates has not yet taken the effect of sequences explicitly into account. One might think that complementary sequences have very limited competition only. We show that, despite interesting sequence effects, competing template replicators yield to Gause's principle of competitive exclusion so that the number of stably coexisting template species cannot exceed the number of nucleotide species on which they grow, although one of the findings is that plus and minus strands together count as one species. Thus up to four different templates/ribozymes can constitute the first steps to an early, segmented genome: we suggest that other mechanisms build on this baseline mechanism.
| Gause (1934) in the Golden Age of theoretical ecology formulated the principle of competitive exclusion, proposing in effect what usually is being referred to as “complete competitors cannot coexist” [1]. Later investigations have confirmed that in stable steady state the number of coexisting species cannot be larger than the number of resources, provided that growth rates depend linearly on resource concentrations and that we look for coexistence at fixed densities [2]–[4]. For maximal coexistence to occur, the competitors must consume the resources in different proportions. Since the seminal experiments of Spiegelman [5] and the deep theoretical insights of Eigen [6], nucleic acid replication kinetics has been under repeated scrutiny. In the “default” model of Eigen with constant total population concentration the fastest replicator (and its associated mutant cloud) wins, consonant with “survival of the fittest”; the tacit assumption being that the competing sequences are complete competitors in the sense of Gause. More detailed investigations of RNA replication kinetics have greatly improved these models, taking into account saturation of the replicase enzyme, asymmetry of plus and minus RNA strands, and replicationally inert double-strand formation [7]–[9]; the latter phenomenon yielding coexistence due to the self-limitation of growth. Von Kiedrowski [10], [11] discovered a somewhat similar phenomenon for his artificial non-enzymatic chemical self-replicators growing parabolically, where self-limitation of growth arises from reversible double-strand formation. Szathmáry and Gladkih [12] showed that the consequential parabolic growth leads to stable dynamical coexistence. Yet none of these models included a detailed analysis of base composition and sequence effects on coexistence. In this paper we remedy this deficiency.
We explicitly take into consideration the concentration of up to four different building blocks (“nucleotides”, with the aim that the model should be general enough to deal with different number of bases and base-pairing modes [13]–[15]) and a large number of competing different sequences, in order (i) to present, at least in part, the missing theory of competing template replicators having different sequences and (ii) to answer the question whether Gause's principle holds for such replicators.
During the forthcoming analysis we deliberately introduce some simplifications. We assume that template replication rates depend on nucleotide concentration linearly (there are no cooperative effects) and that the dynamics of these abiotic resources are not periodically forced, for example. We neglect replicase enzymes and assume that template and replica separate irreversibly upon completion of elongation. The kinetic effects are simplified to the extent that the elongation rate of template polymerization depends only on the identity of the inserted nucleotide and nothing else. We know that this is a crucial simplification but already with this rule different sequences may assume very different kinetic phenotypes. In agreement with this, we neglect secondary and tertiary structures.
The raison d'être for these assumptions is that we would like to demonstrate the effect of competition for resources of competing template sequences as simply and clearly as possible. (We note that as mentioned above, irreversible or reversible pair formation can lead to coexistence, and that enzyme saturation leads to linear growth instead of exponential.) We deliberately want to see the dynamics of coexistence under irreversible exponential growth tendency, as a kind of worst case. The effect of the sequence diversity of templates on dynamical coexistence is not trivial. If there are two resources A and B, then it is trivial that sequences of and may coexist. But what about and ? Are these sufficiently similar for competitive exclusion or sufficiently different for competitive coexistence? And have and got the same features in competition, or not? This question is relevant since a recent study [16] in the ecological literature indicated that life-history traits of organisms can promote dynamical coexistence on limiting resources beyond the effect of simple resource partitioning. Thus two templates with the same nucleotide composition but substantially different sequences may be regarded as adopting two different life history strategies. As replication proceeds sequentially, templates might be regarded as consuming different resources during different stages of their life histories. What is the effect (if any) of this stage-structure on template coexistence?
Some of the effects that we show in this paper are far from trivial. Our calculations show the effect of resource partitioning on template coexistence and shed more light on early molecular evolution, which surely was affected by sequence effects of template replicators.
To understand the mechanism of coexistence of template replicators (sequences) we formulated the dynamics of polynucleotide replication. Here we only explain the necessary basics of our formalism, for the mathematical model see Models Section, for further details see Text S1. Template replicators are assumed to be single-stranded with double-stranded replication intermediates (as for RNA). As a simplification, metabolism responsible for replication is restricted to the common pool of shared monomers, which are either fed from within (protocells) or from the outside (flow reactor).
A sequence is a single polynucleotide strand of the form of length where stands for any monomer at the position. As an example, a sequence of monomers in case of RNA could be with , , etc. A sequence pair is a double-stranded polynucleotide molecule. It can be represented by only one of its strands as it defines the complement strand unambiguously (note that we do not deal with strand separation and treat the two strands as separate sequences). For example the sequence defines its complementary pair . An intermediate complex is a complete sequence and its incompletely built complementary sequence during the duplication phase. For example, one intermediate complex of the above RNA sequence pair is:
An -group of sequence pairs consists of such sequence pairs. Often members of a group are represented by one sequence of each pair for sake of simplicity. Concerning the dynamics, for each type of monomer we introduce a specific rate constant that defines the speed of elongation of the sequence; also for each monomer type and for each intermediate there is a specific degradation rate constant.
First, we investigate the more realistic but also more complex systems that can only be solved numerically and later we gradually traverse to simplified systems that can be handled fully analytically. Such systems, though simplified, provide powerful rules about the mechanisms of coexistence which still can be translated and applied to the realistic cases. Accordingly, first we numerically analyze the complementary replication of templates corresponding RNA replication. Second, we deal with the simpler homologous replication where monomers pair with identical types (non-complementary base-pairing); we also introduce parallel strand polarity as opposed to antiparallel polarity (like in case of RNA replication). The difference between complementary and non-complementary pairing and parallel and antiparallel strand polarity is given in Fig. 1. Third, as a further simplification of the previous system, we assume uniform degradation rates for replication intermediates and even identical elongation rates for the different monomer types to obtain analytical results.
The following model of template replication models RNA replication, dealing with 4 monomers, complementary pairing and antiparallel strand polarity. We have investigated the coexistence of groups of sequence pairs of length . As the sequence space is huge (), consequently, the Cartesian product of the sequence space yielding all possible combinations of sequence pairs is even more huge (). Therefore it is usually impossible to investigate the coexistence of all possible combinations of sequences. Instead, we used a reasonably large sample of the full combined space to estimate the probability and stability of coexisting sequences over four different monomers. We have investigated sequences of length , as this is the maximum length for which the space could be reasonably analyzed. Our analysis was performed using the following two methods (for parameters, see Text S1):
The results of the analysis of coexistence of complementary pairs of sequences of length 4 can be seen in Tables 1 and 2 (methods M1 and M2, respectively). Accordingly, we can state that the increasing number of complementary pairs reduces the probability and stability of coexistence and that on four different monomers a maximum of four different sequence pairs could possibly coexist. Thus Gause's principle (against first intuition) limits the number of sequence pairs instead of individual sequences because of the dynamical coupling between the template and its complement (the plus-minus ensamble behaves like a single replicator, see [6], [17]). In the intermediate case of non-complementary pairing with antiparallel polarity, it is still the number of sequence pairs that limits coexistence (and not individual sequences), though due to a better partitioning of resources coexistence is more probable on average (for details, see Text S1).
Coexistence can be visualized in case of two coexisting sequence pairs of length in two dimensions: each cell of a matrix represents a certain combination of two pairs, and is labeled by the first sequence of the first pair (rows) and the first sequence of the second pair (columns). Sequences are ordered according to standard lexicographic ordering along the horizontal and vertical axes (see Fig. 2). The higher probability and mean stability of coexistence in the bottom corner correspond to the “niche partitioning”: first sequence consumes more , while the second sequence consumes more .
Our model is applicable for any sequence length , predicting whether there is coexistence for any combination of sequences. However, for larger values of the full combined sequence space is enormous therefore it is impossible to perform exhaustive search for all coexisting cases. By brute force search we have found some illustrative examples, according to Method M2 (for details see Text S1). Our results (tested up to ) show that 8 sequences (4 sequence pairs) can stably coexist (linear asymptotic stability was explicitly tested). There is no theoretical (but computational) objection to applying this model for even longer sequences.
In this section we deal with a simplified system: we restrict our attention to homologous base pairing ignoring the polarity of strands (see Fig. 1), allowing only two monomer types (see the extended model of four monomers in the Text S1). Due to homologous pairing of monomers and the lack of different polarities of the strands, template and copy are identical, thus a pair of pairs consists of a total of two different sequences. Intermediates of the first sequence (pair) are denoted as , intermediates of the second are denoted by , their concentrations by and (as were introduced previously), monomers are denoted by and their concentrations by and , (). Note that results in this section hold for all cases when template and copy are identical: this happens not only in the case of homologous pairing with parallel orientation, but also for some exotic cases of antiparallel strands, like palindromic sequences of homologous pairing and reverse-palindromic sequences of complementary pairing (Fig. 1). We briefly note that such palindromes are not likely to coexist any more than non-palindromic sequence pairs.
In this paper we have provided the foundations of the hitherto missing theory of template replication where replication intermediates and different sequences are explicitly taken into account. Under the assumption of fixed stable steady state densities for resources and competitors Gause's principle [1] fully rules over replicator dynamics: coexistence of more replicators than the number of limiting resources (nucleotides) is not asymptotically stable. We have found, however, that template and copy (or plus and minus strands) count as one replicator, since they are stoichiometrically coupled.
We have found cases of coexistence where Gause's principle seems to be violated in that two sequences can coexist with exactly the same nucleotide composition but adequately different sequences: this is a version of the stage-structure effect on coexistence found in theoretical ecology [16]. The part of a sequence that is replicated earlier has a stronger effect than that replicated later, since replication intermediates corresponding to positions in the front are more abundant, hence they influence competitive dynamics more strongly. We have demonstrated the trend that the stability of coexistence in terms of the leading eigenvalue decreases with sequence length. This may be considered bad news; however we should not forget that a good share of ribozymes [18] and aptamers [19] is smaller than 100 nucleotides, for which one still would get acceptable local stability values. (Note that the smallest known ribozyme consists of 5 nucleotides [20].)
The relevance of our findings can be questioned on the grounds that ribozyme replicators should have been longer than considered in this paper. This objection partially loses force if one considers known ribozyme sizes and the early constraints on replication. We discuss these issues in turn. The smallest ribozymes known are: (1) the trinculeotide that catalyses metal ion-dependent cleavage of RNA [21], (2) the pentanucleotide promoting peptide bond formation [22], and (3) the 19 nucleotides long minimalized hammerhead ribozyme for RNA cleavage [23]. Note that these ribozymes are devoid of complex secondary structures that would significantly alter their potential replication kinetics. Nevertheless, Yarus notes that the stable conformation of a cavity formed by the single-stranded overhang beyond the three base pair formed between enzyme and substrates seems to be essential for catalysis [22]. Regarding the replication issue, it is generally believed that small RNA replicators preceded long ones, partly supported by the non-enzymatic replication in the von Kiedrowski experiments [10]. In fact, the production of a generalized replicase ribozyme that could replicate long RNA-s is an unsolved problem. This prompted Ellington [24] to suggest a collectively autocatalytic set consisting of a modest replicase and a ligase. In such a system only small fragments would be replicated, followed by ligation to yield the longer ribozyme structures. Noteworthy in this regard is the case of the collectively autocatalytic, ligating set of Lehman [25], in which fragments of lenghts 43, 65, 55 and 52 are used as pieces in the assembly. It remains to be seen whether these fragments would stably coexist when replicated, using the right combination of the resource and structure fitness landscapes. Of course, we might find in the future ribozymes that could be assembled from even smaller pieces; for such cases our theory would almost immediately apply. In any case, we predict that dynamical coexistence of small, functionally important RNA replicators will be demonstrated in the near future.
Mechanisms for template coexistence have been in the focus of models of primordial replicator evolution (cf. [6], [26]). Here we have shown that up to four replicator pairs (plus and minus sequences) can stably coexist in the same environment without any special coupling. Thus we argue that for any special theory showing that different template replicators can coexist one might find that in effect up to different replicators may coexist without explicit representation of the four nucleotides as resources. This calls for further investigations.
Recently there has been an upsurge in interest in exo/astrobiology. It is in this context that we have deliberately presented results for homologous pairing also, even with parallel orientation of the strands. Although such configurations are not unheard of even in our world, we wanted to see how such features would in general affect dynamical coexistence of template replicators.
We have obtained the fitness landscapes through a distribution of elongation and degradation rates. The main reason behind this is tractability: although the 2D structures as phenotypes of RNA molecules can be calculated for most cases, this does not automatically yield phenotypes in terms of replication rates. We are temporarily satisfied with the phenotype richness that our local rules provided (see Fig. S3). What is more, we predict that the main finding that Gause rules over competitive coexistence of template replicators in stable steady state would not be violated even with more complex fitness landscapes.
In each experiment, we integrated the system of sequence pairs (Eqs. (5)–(9) extended with the dynamics of the rest pairs) until convergence (when the difference of the concentrations of any two intermediates at two successive time steps is less than ) or until extinction (if the concentration of an intermediate is less than the corresponding sequences pair is treated as extinct). We are interested in how many sequence pairs can coexist maximally on four different monomers. According to Gause's principle, one would expect a maximum of two sequence pairs to coexist, as that yields four different sequences. Since members of the pairs are stoichiometrically coupled, this should affect the dynamics, allowing different mechanisms of coexistence.
Let us introduce two complementary sequences:where is the is the type of the monomer at position . Since is the complement of , the overbar denotes the complementing monomer pair ( and , thus , etc.). Replication of the sequences happens as builds up stepwise along . Using the notation above, the intermediate complexes during replication are:When the new copy is completed along the other template, the two strands separate instantaneously yielding and . The schema of the reactions is as follows. Equations (2), (3) and (4) correspond to the replication of and strands and the generation/degradation of components, respectively.(2)(3)(4)where is the elongation rate constant for the given monomer at position , and denote degradation rates for the type of monomer and for the species. The corresponding dynamics of the intermediates is as follows (, denotes concentration of and , respectively, denotes the concentration of the monomer in the sequence, overbar denotes the complementing pair ):(5)(6)(7)(8)and the dynamics of the monomer , where denotes the concentration of monomer , etc. is:(9)where if , otherwise 0 (Kronecker delta).
The extension of the dynamics for more sequence pairs (i.e., to more than one copy and template) is straightforward. The dynamics of the intermediates is independent for each pair and the dynamics of the monomers provides the coupling between the equations of different pairs of sequences. Because of the cross-coupling of equations, no analytical solution was found (some analytical results will be presented for simplified cases). For the numerical integration of the ODE system to find steady-state solutions we have used the CVODE code from the SUNDIALS project of the Lawrence Livermore National Laboratory [27].
Uniform degradation rates of sequence intermediates allow for a completely analytic approach. For the positivity test of concentrations, we introduce the following notation for the constants of the power sum of :(10)(11)The solution of the dynamics of the non-complementary pairing system with uniform degradation rates for intermediates and monomers ( and , respectively, for all ) provides the concentrations of the last intermediates of and :(12)(13)where and are the stable steady state monomer concentrations (for detailed derivation, see Text S1).
Let us assume that influx can counter degradation. In this case the condition of coexistence ():(14)To sum up, coexistence is possible if and are of different signs.
If the two elongation rate constants of the two monomers are identical () the parameter equals 1, thus the simple criteria of coexistence are the following:(15)
For example, sequences = and = are not coexisting ( and ), according to Gause's principle, as for both sequences is the limiting resource for which they compete. On the other hand, = and = are coexisting ( and ).
Sequences = and = demonstrate an example of irregular coexistence seemingly violating Gause's principle, as both have -majority. Though, according to Eq. (15), ( and ), and thus behaves as having -majority.
|
10.1371/journal.ppat.1004891 | Staphylococcus aureus Survives with a Minimal Peptidoglycan Synthesis Machine but Sacrifices Virulence and Antibiotic Resistance | Many important cellular processes are performed by molecular machines, composed of multiple proteins that physically interact to execute biological functions. An example is the bacterial peptidoglycan (PG) synthesis machine, responsible for the synthesis of the main component of the cell wall and the target of many contemporary antibiotics. One approach for the identification of essential components of a cellular machine involves the determination of its minimal protein composition. Staphylococcus aureus is a Gram-positive pathogen, renowned for its resistance to many commonly used antibiotics and prevalence in hospitals. Its genome encodes a low number of proteins with PG synthesis activity (9 proteins), when compared to other model organisms, and is therefore a good model for the study of a minimal PG synthesis machine. We deleted seven of the nine genes encoding PG synthesis enzymes from the S. aureus genome without affecting normal growth or cell morphology, generating a strain capable of PG biosynthesis catalyzed only by two penicillin-binding proteins, PBP1 and the bi-functional PBP2. However, multiple PBPs are important in clinically relevant environments, as bacteria with a minimal PG synthesis machinery became highly susceptible to cell wall-targeting antibiotics, host lytic enzymes and displayed impaired virulence in a Drosophila infection model which is dependent on the presence of specific peptidoglycan receptor proteins, namely PGRP-SA. The fact that S. aureus can grow and divide with only two active PG synthesizing enzymes shows that most of these enzymes are redundant in vitro and identifies the minimal PG synthesis machinery of S. aureus. However a complex molecular machine is important in environments other than in vitro growth as the expendable PG synthesis enzymes play an important role in the pathogenicity and antibiotic resistance of S. aureus.
| Peptidoglycan forms the stress-bearing sacculus that prevents lysis of bacteria due to turgor pressure. The integrity of peptidoglycan is therefore essential for bacterial survival and its synthesis is the target of many important antibiotics, such as penicillin. The final steps of peptidoglycan synthesis are catalyzed by penicillin-binding proteins, enzymes that are proposed to work in multi-enzyme complexes. We show that seven of the nine genes encoding peptidoglycan synthesis enzymes can be deleted from the Staphylococcus aureus genome without affecting normal growth and cell morphology in vitro, identifying the minimal peptidoglycan synthesis machinery of this organism. Identification of minimal machineries is key for synthetic biology efforts towards the design of systems with reduced complexity. However, the non-essential peptidoglycan synthetic proteins are important for survival of S. aureus in more challenging environments, such as in the presence of antibiotics that target cell wall synthesis or within the host, as shown by the inability of the mutant strain to establish a successful infection and kill Drosophila flies.
| Many cellular functions are performed by molecular machines that are composed of multiple proteins. Consequently, it is often difficult to determine the precise role of each protein within such a complex. In part this is due to functional redundancy, or to the interdependency of proteins that can result from a recruitment hierarchy or from a requirement of the physical presence of individual proteins to the stability of the entire complex. One approach for the identification of the essential components of a cellular machine consists of determining its minimal protein composition. This information is also key for synthetic biology efforts towards the design of systems with reduced complexity.
One example of a molecular machine that was proposed almost two decades ago [1] but is not yet fully characterized, is the protein complex responsible for the synthesis of peptidoglycan (PG). PG, the main constituent of the bacterial cell wall, is a macromolecule composed of long glycan chains of alternating N-acetylglucosamine and N-acetylmuramic acid units, cross-linked by flexible peptide bridges. The resulting mesh forms a stress-bearing sacculus that envelopes the bacterial cell and prevents lysis due to turgor pressure. The integrity of PG is therefore absolutely essential for bacterial survival and many important antibiotics, such as β-lactams and glycopeptides, target penicillin-binding proteins (PBPs), the enzymes involved in the final stages of PG synthesis. PBPs catalyze the two reactions—transglycosylation and transpeptidation—required to synthesize the glycan strands and to crosslink them via peptides, respectively.
PBPs have been proposed to work in multi-enzyme complexes that may also include cell wall hydrolases and other PG synthesis proteins [1,2]. These complexes would facilitate the coordinated activity of PG synthases and hydrolases, to ensure that growth of the PG mesh occurs without endangering the integrity of the PG sacculus. However, despite years of work from several groups, this hypothetical complex has not been isolated and, particularly in Gram-positive bacteria, we currently lack strong evidence for its existence.
One of the difficulties in studying PG synthesis is the large number of PBPs with apparent redundancy during growth in rich media. For example, the two best-studied bacterial species, Escherichia coli and Bacillus subtilis, have 12 and 16 PBPs, respectively [3], making it difficult to unravel the specific role of each of these proteins. Staphylococcus aureus, a Gram-positive bacterial pathogen, is a particularly good model to study cell wall synthesis, as it contains only four PBPs, PBP1-4 in methicillin-susceptible S. aureus (MSSA) strains, or five PBPs in methicillin-resistance S. aureus (MRSA) strains. The latter contain an additional PBP, PBP2A, which is the main determinant for β-lactam resistance due to its low affinity for these antibiotics [4]. The role of the other four, native, staphylococcal PBPs has been reasonably well studied. PBP1 is an essential protein with transpeptidase (TPase) activity, and is involved in cell division and separation [5,6]. PBP2 is the only bi-functional PBP in S. aureus, with both TPase and transglycosylase (TGase) activities. This protein is essential in MSSA strains, but not in MRSA strains, given that the TPase domain of PBP2 can be replaced by that of PBP2A [7]. However, PBP2 becomes essential in the presence of β-lactams, when cooperation between its TGase domain and the TPase domain of PBP2A is required for survival [8]. PBP3 and PBP4 are non-essential proteins. The function of the former remains largely unknown [9], while the latter was shown to be required for the synthesis of highly cross-linked PG [10]. PBP4 has also been described as important for β-lactam resistance in community-acquired MRSA (CA-MRSA) strains [11].
Besides PBPs, the S. aureus genome encodes four additional, non-essential, proteins with proven or hypothesized roles in PG synthesis. These are two monofunctional TGases, MGT and SgtA [12–14], and two auxiliary proteins, FmtA and FmtB, which have homology to TPase domains, and therefore are thought to have a role in cell wall synthesis [15–17].
The study of PG synthesis, the target of β-lactam antibiotics, is particularly relevant in S. aureus, as MRSA strains are currently one of the leading causes of hospital- and community-acquired infections, causing more deaths in the USA than tuberculosis and AIDS combined [18]. The high frequency of antibiotic resistance emphasizes the need for the identification of novel drug targets and of strategies to increase the efficacy of available antibiotics against this pathogen.
Here we describe the construction of a S. aureus strain encoding the minimal number of PG synthesis enzymes required for survival, which constitute the most relevant targets for the development of antibiotics capable of inhibiting the final stages of PG synthesis. We show that S. aureus cells are capable of PG synthesis catalyzed solely by the transpeptidase PBP1 and the bi-functional PBP2. However, these cells show loss of resistance to cell wall-targeting antibiotics and decreased virulence in a Drosophila infection model, highlighting the important roles of the non-essential enzymes for survival in natural environments, such as a host.
To determine the exact number of proteins likely to play a direct role in PG synthesis in S. aureus, we used profile Hidden Markov Models (pHMMs) [19], a sensitive bioinformatics approach that can identify both close and distantly related homologues. We scanned the S. aureus genome with pHMMs based on known transpeptidase and transglycosylase enzymes (SCOP-56601 and SCOP-159832). This analysis identified all nine previously known enzymes (one bi-functional) with a demonstrated or suggested role in PG synthesis (see Introduction) comprising a total of ten active domains, 7 transpeptidases (TPases) and 3 transglycosylases (TGases) (Fig 1).
Next, we extended this analysis to 1295 organisms from the bacterial kingdom to determine the absolute minimal number of PG synthesis domains present in a bacterium whose genome has been fully sequenced. Fig 1 shows the results for selected species (mainly disease causing and model organisms) and illustrates the variety in the number of enzymes predicted to be involved in PG synthesis. The number of encoded TPases and TGases varied dramatically from as many as 62, encoded by Streptosporangium roseum, to none in intracellular pathogens such as Mycoplasma pneumoniae, which are often considered to be lacking PG. The lowest number of enzymes in free-living bacteria was found in the non-pathogenic bacterium, Sphaerochaeta coccoides (1 TPase), though studies of its PG composition have not been described. Among species with known PG composition, Helicobacter pylori has the smallest set of PG synthesis enzymes, with one bi-functional enzyme (PBP1) and two monofunctional TPases (PBP2 and PBP3) [20]. In this organism PBP2 is essential for viability [21] and depletion of PBP1 or PBP3 causes significant morphological defects [22].
The non-essential nature of PBP3, PBP4 and PBP2A for survival of S. aureus has been previously described [4,9,10,23]. MGT and SgtA as well as FmtA and FmtB are also individually dispensable, which could be due to redundant functions of these proteins [12,15,17]. In order to assess the effect of the simultaneous lack of these enzymes upon growth and survival, we sequentially deleted from the chromosome of the MRSA strain COL, the genes pbp3, pbpd (encoding PBP4), mgt, sgtA, fmtA, fmtB, and mecA. In-frame, markerless deletions of each gene were successively constructed, as outlined in S1 Fig, to produce the strain COL MIN. This strain contains a minimal PG synthesis machinery, as it encodes only two of the nine known proteins with PG synthesis activity—PBP1 and PBP2. Western blotting of whole cell extracts was performed to confirm the absence of PBPs (S2 Fig) and all gene deletions and lack of gene duplications were confirmed by PCR (S2 Fig) and genome sequencing (see below).
Surprisingly, S. aureus was able to grow normally both in rich medium as well as in minimal medium in the absence of seven PG synthesis enzymes (Fig 2A and 2B), requiring only PBP1 and PBP2 for viability under laboratory conditions.
PBP1 is an essential protein [6], thus in order to verify if this was also the case in COL MIN, we constructed strain COL MIN PBP1i, in which the native copy of the gene encoding for PBP1 was placed under the control of the IPTG-inducible Pspac promoter in the COL MIN background. Growth analysis showed that upon depletion of the protein (via growth in the absence of IPTG), cells do not grow, confirming the essentiality of PBP1 in COL MIN (Fig 2C). Given that PBP1 is part of a large operon, polar effects were ruled out by introducing a replicative plasmid expressing PBP1 into COL MIN PBP1i and showing that the resulting strain (COL MIN PBP1i pSKP1) is able to grow in the absence of IPTG (S3 Fig).
PBP2 expression in the absence of either PBP2A [7] or MGT [12] has been previously described as essential. Thus we placed PBP2 expression under the control of the IPTG-inducible Pspac promoter in the background of COL MIN and confirmed that the resulting strain does not grow in the absence of the inducer (Fig 2D). Although the gene encoding PBP2 is part of an operon, the other gene in the operon, recU, is not essential [24].
To determine if COL MIN had suppressor mutations that allowed survival in the absence of seven PG synthesis enzymes, we sequenced its entire genome (Table A in S1 Text). This approach confirmed the deletion of pbp3, pbpd, mgt, sgtA, fmtA, fmtB, and mecA in COL MIN. A total of fifteen mutations were identified in COL MIN that were absent in the parental strain COL. Of these, 9 were very close to the excision sites corresponding to deletions of sgtA, mgt, fmtB and mecA and are therefore unlikely to have any effect upon other genes. Five genetic differences between COL MIN and the parental strain COL were SNPs in non-coding regions of the genome, close to genes of unknown function or unrelated to peptidoglycan synthesis. Only one SNP was in a coding region, in the gene encoding the molybdenum-binding protein ModA. In Gram-negative organisms this protein is involved in the uptake of nutrients including metals such as molybdenum [25] and therefore it is unlikely that the ModA SNP found in COL MIN promotes new PG synthesis activities or decreases PG autolysis.
Given that genome sequencing did not elucidate any mechanism of adaptation of COL to the loss of seven proteins required for peptidoglycan synthesis, we compared the over all gene expression profile of COL and COL MIN by total RNA sequencing (RNA-Seq). Expression of approximately 155 genes was significantly altered (p<0.05 and over two-fold difference, Table B in S1 Text). Of these, 59 are part of the large lysogenic prophage L54a that may be excised in a subpopulation of COL MIN cells. Analysis of the remainder showed upregulation of the vraSRF operon, which controls the so-called “cell wall stress stimulon”, known to be regulated in response to cell wall damage [26]. The vraS and vraR genes were up regulated more than 3-fold in COL MIN compared to COL. Almost half of the 46 genes shown to be regulated by the VraSR system [26] also showed increased expression of the transcripts and of these 11 showed a fold-difference in expression of greater than 2-fold. Another relevant observation was the down regulation of a number of known or putative PG hydrolases (sle1 (2.3 fold), lytD (1.7 fold), SACOL2298 (1.3 fold), sceD (3.6 fold), isaA (2.8 fold) and lytM (3.3 fold)), presumably to reestablish a required balance between PG synthesis and autolysis. Interestingly, none of the proteins involved in the synthesis of PG precursors (Mur A-G, FemABX, MraY) showed differences in gene expression in the mutant strain.
To determine if the lack of seven PG synthesizing enzymes in COL MIN resulted in cell morphology defects, cells were examined with transmission electron microscopy and with super-resolution fluorescence microscopy using the DNA dye Hoechst 33342 and a fluorescent derivative of vancomycin (Van-FL), which labels PG by binding the terminal D-Ala-D-Ala residues of the muropeptides. COL MIN showed no morphological changes compared to the parental strain COL, when observed by Structured Illumination Microscopy (SIM) and electron microscopy (Fig 3A and 3C). The fraction of cells with septa (COL 38.1%, COL MIN 37.6%) and the average cell diameter (COL 1.06 μm ± 0.08, COL MIN 1.07 μm ± 0.08) remained unchanged. These results suggest that, in COL MIN, the remaining PG synthesis enzymes were sufficient to maintain the normal morphology of the cell. In wild type S. aureus, both PBP1 and PBP2 have been shown to localize to the division septum [5,27], where cell wall synthesis takes place [28,29]. Given that PBP1 or PBP2 have been reported to interact with other PG synthesis enzymes in bacterial two hybrid assays [12,30], it was possible that these proteins required the presence of other members of a putative cell wall synthesis complex for correct localization. We therefore investigated the localization of PBP1 and PBP2 in COL MIN by immunofluorescence, in strains lacking the spa gene given that its gene product, Protein A, binds with high affinity to IgG molecules, and using FtsZ as a control for septal localization. Both PBP1 and PBP2 maintained their normal septal localization in the absence of the other PG synthesis enzymes (Fig 3D) indicating that they do not depend upon them for correct localization.
We also verified whether the presence of PBP1 and PBP2 was sufficient for synthesis of PG with a composition similar to the parental strain COL. For that purpose, PG was purified from the parental strain COL and COL MIN and digested with muramidase to isolate muropeptides. When muropeptides were analyzed by reverse-phase-high pressure liquid chromatography (HPLC), the major difference between the two strains was a reduction in highly cross-linked muropeptides, (peak V and above, S4 Fig). This phenotype was expected as PBP4, and to a smaller extent FmtA, both absent in COL MIN, were previously shown to be responsible for the synthesis of highly cross-linked PG in S. aureus [17,31]. Analysis of single deletion mutants in genes encoding PG synthesis enzymes confirmed that the decrease in crosslinking in COL MIN was due to the lack of PBP4 and FmtA (S4 Fig).
To analyze the glycan strands from COL and COL MIN, we performed a sequential digestion of the PG with lysostaphin (to cleave the pentaglycine bridge between different glycan strands) followed by LytA amidase (to remove the stem peptides from the glycans) as previously described [32]. The glycans were then analyzed by HPLC and showed only minor differences in length or composition (S5 Fig). We have previously reported that PBP2 is the major transglycosylase in S. aureus [12] and the results here confirm that removal of MGT and SgtA, the other enzymes with transglycosylase activity, has no major effect upon glycan chain length or composition. Thus our results show that S. aureus can form a functional cell wall using only PBP1 and PBP2.
Importantly, the minimal strain described above was constructed in the background of MRSA strain COL. MRSA strains are adapted to live in the presence of β-lactam antibiotics, i.e., in conditions where their native transpeptidases are inactivated. We therefore questioned if the genes pbp3, pbpd, mgt, sgtA, fmtA and fmtB could also be deleted in an MSSA strain, which is susceptible to β -lactams. S6 Fig shows that strain Newman MIN, which encodes only PBP1 and PBP2 as PG synthesis enzymes, was also viable in rich and minimal medium.
We have described above that deletion of genes encoding seven of the nine identified PG synthesizing enzymes from the genome of S. aureus has little effect upon viability, growth, morphology or PG composition in rich medium. However, in vitro growth assays are not sufficient to study the fitness of a pathogen such as S. aureus, best known for its ability to resist various antibiotics, in particular β-lactams, which target cell wall synthesis. Therefore we measured the minimum inhibitory concentration (MIC) of an array of antibiotics for COL and COL MIN, Newman and Newman MIN. In the absence of the non-essential PG synthesis proteins, S. aureus became exquisitely sensitive to antibiotics directly targeting PG synthesis enzymes (β-lactams and moenomycins), but retained normal resistance to antibiotics targeting other pathways (Table 1). For example, in COL, oxacillin resistance is only slightly reduced (2 fold) in the combined PBP3, PBP4, MGT and SgtA mutant (Table C in S1 Text). Upon additional deletion of fmtA, resistance to antibiotics that target TPases decreases dramatically, even though the strain retains PBP2A and PBP2, the two enzymes capable of sustaining PG synthesis in the presence of β-lactams [7,8]. FmtA has been previously described to have a role in β-lactam resistance [17], as interruption of the corresponding gene resulted in an 8-fold decrease in oxacillin resistance. Here, deletion of fmtA in a background already lacking PBP3, PBP4, MGT and SgtA led to a greater than 100-fold decrease in the oxacillin MIC, indicating that there is a cumulative effect on resistance upon deletion of the genes encoding these proteins (Table C in S1 Text).
Finally, we tested if COL MIN was able to establish a successful infection, using Drosophila melanogaster as a model organism. D. melanogaster has been used to show that the composition of the cell surface of S. aureus has a crucial role in the ability of these bacteria to avoid host recognition and survive inside the host [33].
To determine whether the deletion mutant strain was affected in virulence, we injected Drosophila flies with equal numbers of COL and COL MIN S. aureus cells and determined the ability of both strains to kill Drosophila. Over 90% (n≈90) of flies injected with the parental strain COL were killed within 96 hours, while only 12% of the flies infected with COL MIN were killed during that period (Fig 4A). To test whether decreased virulence was due to impaired ability of COL MIN to avoid host recognition, we used a Drosophila strain mutant for the peptidoglycan receptor pgrp-sa. A functional PGRP-SA is paramount for host survival against S. aureus infection, binding to S. aureus peptidoglycan and leading to Toll pathway activation, with the consequent production of antimicrobial peptides [34,35]. Injection of COL MIN in Drosophila seml mutants (which express a non-functional PGRP-SA) resulted in killing of approximately 95% of the flies (Fig 4B), showing that the impaired ability of COL MIN to kill Drosophila wild type flies was likely due to decreased ability to avoid recognition by PGRP-SA. Accordingly, COL MIN was unable to propagate in wild type Drosophila flies, but it was able to propagate as well as COL in Drosophila seml mutants (S7 Fig). Furthermore, the inability of COL MIN to propagate in wild type Drosophila flies was not due to higher production of antimicrobial peptides, as expression of drosomycin was significantly lower when flies were infected with COL MIN than when flies were infected with COL (S7 Fig).
Even in the absence of PGRP-SA (i.e. in Drosophila seml mutants), the killing curves of COL and COL MIN were statistically separable (Fig 4B). This could be due to the slightly reduced growth rate of COL MIN or could suggest the existence of a second factor, besides enhanced recognition by PGRP-SA, which contributes to the reduced virulence of COL MIN. We therefore determined the susceptibility of COL MIN and COL to lysozyme, a PG hydrolase, and found sensitivity to be increased in COL MIN (Fig 4C), possibly due to the decreased degree of PG crosslinking [36,37], suggesting that the reduced virulence of COL MIN is due to a decreased ability to avoid host recognition and to resist host defense mechanisms such as the action of bacteriolytic enzymes.
One way to determine the minimal required composition of complex multi-protein machines is through genetic screens, which have been used in various bacterial species, including S. aureus [38], to identify all essential genes. However, those screens identify genes that are dispensable when disrupted one at a time, but not genes with redundant functions that cannot be simultaneously deleted from the genome due to synthetic lethality. To identify such genes, deletion of multiple genes in various combinations is required, a process that can be used both to identify the key proteins required for specific pathways, and for efforts to build a minimal cell that contains the smallest set of essential genes.
In this work we have identified the minimal components of the molecular machinery required for PG synthesis in S. aureus. In the late 1990´s, J. Höltje proposed a three-for-one model for synthesis of Gram-negative PG, which stated that it required the concerted action of multiple enzymes that simultaneously polymerize and hydrolyze the PG, to allow insertion of new material without compromising the integrity of the stress-bearing PG sacculus that surrounds the bacterial cell [1,2]. According to the model, this concerted action would be facilitated through the formation of multi-enzyme complexes, which would allow better coordination of the proteins involved in PG synthesis [1]. The work that led to this model was performed in Escherichia coli, a Gram-negative bacteria, which has a very thin layer of PG [39], and since then genetic, biochemical and cell biological approaches have shown that various PG synthesis enzymes from E. coli are indeed able to interact with each other (reviewed in [40]). Despite this progress we are still far from understanding the PG synthesis complex in a level of detail similar to what has been achieved for other enzyme complexes, for example those involved in the synthesis of DNA or RNA [41,42]. Even less information has been obtained for proteins synthesizing PG in Gram-positive bacteria that contain a thick layer of PG [43]. Interactions between PBPs of Gram-positive organisms have been suggested, based on bacterial two-hybrid assays [12,30], but so far, to the best of our knowledge, no interactions between Gram-positive PG synthesis enzymes have been confirmed biochemically.
In this work, we have generated a mutant S. aureus strain (COL MIN) by deleting seven of the nine known genes encoding proteins with PG synthesis activity. Remarkably, the resulting cells are not only viable, but grow almost as well as the parental strain in rich as well as minimal medium, and have normal morphology and cell size, as seen by super resolution and electron microscopy. In this mutant, the PG is less extensively cross-linked, as would be expected for a strain lacking PBP4 [31], but otherwise, its composition is similar to the wild type strain. One could expect that if a large complex requiring some or all of these enzymes was required for PG synthesis, its activity would be impaired in the COL MIN strain, which is not what was observed. Furthermore, it is often the case that different components of multi-enzyme complexes depend upon each other for correct localization, as has been shown for the cell division machinery, the divisome. This multi-protein complex is built by recruiting proteins to the division site in a specific order, with proteins crucially depending on the presence of earlier ones for correct localization [44,45]. In COL MIN, the two remaining known PG synthesis enzymes, PBP1 and PBP2, are correctly localized at the septum and therefore do not depend on other PG synthesis enzymes for correct localization.
Currently, we cannot rule out the possibility that other, as yet unidentified, proteins are involved in PG synthesis of S. aureus but show no homology to known TPases or TGases and were therefore missed in our bioinformatics analysis. In B. subtilis, which encodes no canonical monofunctional glycosyltransferases, it is possible to remove all class A PBPs, and thus all known TGase activities [46]. This could suggest that other proteins capable of synthesizing PG but lacking homology to known TGases are yet to be identified. A similar observation has been made in E. faecium, which can also survive in the absence of its three Class A PBPs, suggesting that polymerization of the PG glycan strands in this mutant is catalyzed by an unknown transglycosylase enzyme [47].
Whole genome sequencing of COL and COL MIN identified minor differences between the two strains, but no gene duplications or chromosomal rearrangements. Fifteen SNPs were identified, in the scar regions corresponding to the deleted genes, in non-coding regions and in the gene encoding the molybdenum-binding protein ModA, which is part of an ABC transporter system for the uptake of nutrients [25] and is not involved in peptidoglycan synthesis. None of the mutations were present in genes likely to act as suppressors for the lack of the seven deleted genes encoding proteins involved in PG synthesis. However, transcriptome-wide analysis of the mutant COL MIN compared to the parental strain COL, by total RNA sequencing, revealed a down-regulation of a number of PG hydrolases in response to removal of seven of the nine PG synthetic enzymes. It was previously shown that down-regulation of PBP2 expression caused a concurrent reduction in transcript level of the major autolysins Atl and Sle1 [48], and it is thought that transcriptional regulation between cell wall synthetic and hydrolytic enzymes exists. Therefore it is not surprising that we note a down-regulation of some of the other PG hydrolases of S. aureus in the absence of seven PG synthetic enzymes. RNA-Seq data also confirmed that transcription levels of PBP1 and PBP2 remain unchanged in the mutant strain COL MIN indicating that normal levels of these proteins are sufficient in order to perform normal CW synthesis in the absence of the other enzymes. None of the genes involved in synthesis of the PG precursor Lipid-II had altered transcript levels, strengthening the suggestion that PBP1 and PBP2 can function normally to build PG in the strain COL MIN.
In summary, this study has identified the minimal machinery required for PG synthesis in S. aureus. While our goal was not to disprove the existence of a multi-enzyme PG synthesis complex, our results demonstrate the plasticity of this process. Apparently, S. aureus can synthesize its PG by using just two synthesis protein, PBP1 and PBP2, which performs both transpeptidation and transglycosylation reactions. In agreement with our data, a recent report described that in Caulobacter crescentus, a Gram-negative organism, only one of the five bi-functional PBPs encoded by this organism is required for growth and normal morphogenesis [49]. This study however did not investigate the requirement of the C. crescentus monofunctional transglycosylase MtgA.
The existence of simple PG synthesis machineries is further supported by our comprehensive search for PG synthesizing proteins in sequenced bacterial genomes. Among the 1295 species analyzed, we found one free-living bacteria with even fewer proteins than we have in COL MIN, the Gram-negative termite hindgut bacterium Sphaerochaeta coccoides, which encodes for only one TPase. However, to the best of our knowledge, studies of the cell wall of this organism have not been performed. Among bacteria with characterized PG, the Gram-negative pathogenic bacterium Helicobacter pylori has the minimal set of PG synthesis enzymes, with two monofunctional TPases and one bi-functional enzyme [20]. However, the vast majority of species have a higher number of PG synthesis proteins, suggesting that optimal growth in challenging environments requires a more complex set of these proteins. Accordingly, although S. aureus can apparently survive relying solely upon the activity of PBP1 and PBP2 for PG synthesis, we have shown that non-essential PG synthesis enzymes are required for survival in more complex habitats, i.e., in a host infection model, or in the presence of antibiotics. In fact, COL MIN is exquisitely sensitive to antibiotics that target enzymes with TPase or TGase activities and it is also severely affected in its ability to establish a successful infection and kill Drosophila flies. Further support for the role of PBPs in virulence comes from a recent study showing that inhibition of PBPs by nafcillin reduces virulence of MRSA in a murine subcutaneous infection model [50]. The specific role of each of the cell wall synthesis enzymes missing in COL MIN for survival within the host is not known. However, we have shown that the inability of COL MIN to successfully kill Drosophila flies is essentially reversed if the host lacks the peptidoglycan recognition protein PGRP-SA, which indicates that impaired virulence is most likely due to alterations in the cell surface that affect host recognition of the peptidoglycan. Some contribution to impaired virulence may come from increased susceptibility to the bacteriolytic effects of lysozyme, a PG hydrolase expressed by many host organisms.
By removing seven of the nine PG synthesis enzymes present in S. aureus cells, we have highlighted the redundancy in this process in rich and minimal medium, but not in more challenging environments such as in the presence of cell-wall targeting antibiotics or within the host.
The bacterial strains used in this study are listed in Table D in S1 Text and details of their construction are outlined in the supplementary materials and methods (S1 Text). Plasmids and primers used in this study are listed in Tables E and F in S1 Text, respectively. S. aureus strains were grown at 37°C in Tryptic soy broth medium (TSB; Difco) or on Tryptic soy agar (TSA; Difco) supplemented with appropriate antibiotics when required (erythromycin 10 μg/ml, chloramphenicol 10 μg/ml or tetracycline 5 μg/ml; Sigma-Aldrich) or with 0.5 mM isopropyl-β-D-thiogalactopyranoside (IPTG; VWR). For growth studies in minimal media cells were grown in SSM9PR minimal media containing 1 x M9 salts, 2 mM MgSO4, 0.1 mM CaCl2, 1% glucose, 1% casaminoacids, 1 mM Thiamine-HCl and 0.05 mM nicotinamide at 37°C. E. coli strains were grown at 30°C or 37°C in Luria-Bertani broth medium (LB broth; Difco), on LB agar (Difco), supplemented with 100 μg/ml ampicillin (Sigma-Aldrich) or 50 μg/ml kanamycin (Sigma-Aldrich), 40 μg/ml 5-bromo-4-chloro-3-indolyl-β-D-galactopyranoside (X-gal; VWR) and 0.5 mM IPTG when required.
Analysis of the distribution of PBPs and enzymes containing PBP-like domains was performed using domains SCOP-56601 and SCOP-159832 in the Superfamily database version 1.75 [51], as proxies for transpeptidases and transglycosylases respectively. A maximum likelihood tree was constructed using the software AMPHORA2 [52] and PhyML [53] following the procedure described by Wu et al [54].
Genomic DNA was extracted from individual cultures of the parental, intermediate and minimal strains, and sequenced using the Illumina HiSeq system at Beijing Genomics Institute or the Illumina MiSeq system at Instituto Gulbenkian de Ciência, Oeiras, Portugal. 300-bp paired end reads with over 100x average coverage were generated. Sequence reads were assembled with SeqMan NGen 11 software using the COL genome (NCBI Accession NC_002951.2) as a reference. The variations that were detected in the parental COL strain were filtered to identify the mutations occurring in the intermediate and minimal strains. Low quality variations with read frequencies below 50% were removed from the dataset.
Overnight cultures from isolated colonies were diluted 1/200 into fresh medium and incubated at 37°C with aeration to exponential phase (OD600 ≈ 0.6). Cells were harvested, and RNA was extracted using the Qiagen RNA Easy Kit. Integrity of RNA samples was evaluated using a 2100 Bioanalyzer (Agilent Technologies). Total RNA samples (40 μg) were sent to GATC Biotech AG, Germany for library preparation (including rRNA depletion and DNase treatment) and sequencing of libraries was performed using an Illumina HiSeq platform (paired end, 2 x 125bp read length). The RNA-Seq reads were aligned to the COL reference genome (NC_002951) using Bowtie [55], generating genome/transcriptome alignments. Cufflinks was used to process the raw data, identifying and quantifying the transcripts from the preprocessed RNA-Seq alignment-assembly. Cuffdiff was used to compare the transcripts from COL and COL MIN to determine the differential expression levels between the two samples.
Growth of parental and mutant strains in liquid culture was analyzed by diluting overnight cultures 1/200 into fresh medium (TSB). Cultures were incubated at 37°C with shaking and OD600 was monitored. Growth of strains with pbp1 or pbp2 under the control of the IPTG-inducible promoter Pspac was monitored by first incubating strains overnight at 37°C in TSB medium supplemented with appropriate antibiotics and 0.5 mM IPTG. Cells were harvested the following day, washed three times with fresh TSB lacking IPTG and used to inoculate media with and without IPTG. Cultures were incubated at 37°C with agitation and the OD600 was recorded.
Determination of the MIC to an array of antibiotics was performed in TSB by micro-dilution in 96-well plates. Overnight cultures of parental and mutant strains were added at a final cell density of 5x103 CFU/ml to wells containing 2-fold dilutions of each antibiotic. Plates were incubated at 37°C for 24 or 48 hours and the MIC was recorded as the lowest concentration of antibiotic that inhibited bacterial growth. Results are the average of six independent MIC determination experiments.
Peptidoglycan was prepared from exponentially growing cells as previously described [56]. Muropeptides were prepared by digestion with mutanolysin and glycan strands were isolated from purified peptidoglycan by sequential digestion with recombinant lysostaphin (1 μg/ml, Sigma) and purified pneumococcal amidase (LytA, 50–100 μg/ml) essentially as previously described [32] and detailed in the supplementary materials and methods (S1 Text).
Parental and mutant strains were labeled with DNA dye Hoechst 33342 (1 μg/ml, Invitrogen) and the cell wall dye Van-FL (Invitrogen) and visualized by fluorescence Structured Illumination Microscopy (SIM). Immunolabelling of COLΔspa and COL MINΔspa cells with antibodies specific for PBP1, PBP2 and FtsZ was performed essentially as previously described [5]. Electron microscopy analysis of parental and mutant strains was performed as previously described [57]. All microscopy procedures are detailed in the Supplementary materials and methods (S1 Text).
S. aureus cells (overnight culture) were harvested by centrifugation, washed once with phosphate saline solution (PBS, 10 mM Na2PO4/ 150 mM NaCl, pH 6.5), and adjusted to an OD600 of 0.4 in 50 ml of PBS. Cell suspensions were split in two and incubated with or without lysozyme (final concentration 300 μg/ml; Sigma) for 6 h with shaking at 30°C. Bacterial lysis was monitored by following the OD600 and the percentage of bacterial lysis was calculated by using the following formula: (ODT/ODT0) × 100, where ODT0 indicates OD of the culture for time zero and ODT is the OD of the culture after incubation with lysozyme at a certain time point.
Isogenized 25174 Drosophila flies (Bloomington Drosophila Stock Center) were used as wild-type background and PGRP-SAseml flies [58,59] were used as a loss of function strain for the Drosophila PGRP-SA receptor. All flies were kept on maize malt molasses food in bottles and reared at 25°C before infection.
To infect flies, S. aureus COL and COL MIN strains were cultured in TSB for 16 hours; cells were harvested by centrifugation (4000 rpm for 7 minutes) and washed in sterile phosphate buffered saline (PBS). Washed bacterial cells were again centrifuged and re-suspended in PBS to an optical density of approximately 0.330 (Thermo Scientific NanoDrop 1000 spectrophotometer). The inoculants containing S. aureus COL and COL MIN strains were further diluted 450-fold in PBS. Thirty CO2-anaesthetized female 25174 or PGRP-SAseml flies (aged 2–4 days) were infected with 13.2 nl of bacterial cells suspension (or with PBS control), directly injected into the haemolymph through the dorsolateral region of the thorax, using a micro-injector (Drummond Scientific Nanoinject II). The number of viable bacteria cells injected per fly was approximately 120, as calculated from plating homogenates of six injected flies, previously ground in TSB medium. Flies were kept at 30°C post-infection and transferred to fresh vials every 2 days. For scoring survival, the number of dead flies was recorded every 12 hours over a 4-day period. The experiment was repeated independently three times. Estimated survival curves were plotted from the raw data sets and the Log-rank (Mantel-Cox) test was used to determine significance between the curves. For clarity in display, 95% confidence intervals have been omitted from the graphs. The data was analyzed using GraphPad Prism 5 (GraphPad Software, Inc.); P<0.05 for two estimated survival curves was considered significant.
|
10.1371/journal.ppat.1005461 | Surfactant Protein-D Is Essential for Immunity to Helminth Infection | Pulmonary epithelial cell responses can enhance type 2 immunity and contribute to control of nematode infections. An important epithelial product is the collectin Surfactant Protein D (SP-D). We found that SP-D concentrations increased in the lung following Nippostrongylus brasiliensis infection; this increase was dependent on key components of the type 2 immune response. We carried out loss and gain of function studies of SP-D to establish if SP-D was required for optimal immunity to the parasite. N. brasiliensis infection of SP-D-/- mice resulted in profound impairment of host innate immunity and ability to resolve infection. Raising pulmonary SP-D levels prior to infection enhanced parasite expulsion and type 2 immune responses, including increased numbers of IL-13 producing type 2 innate lymphoid cells (ILC2), elevated expression of markers of alternative activation by alveolar macrophages (alvM) and increased production of the type 2 cytokines IL-4 and IL-13. Adoptive transfer of alvM from SP-D-treated parasite infected mice into naïve recipients enhanced immunity to N. brasiliensis. Protection was associated with selective binding by the SP-D carbohydrate recognition domain (CRD) to L4 parasites to enhance their killing by alvM. These findings are the first demonstration that the collectin SP-D is an essential component of host innate immunity to helminths.
| Infections by parasitic worms are very common, and controlling them is a major medical and veterinary challenge. Very few drugs exist to treat them, and the parasites can develop resistance to these. In order to find new ways to control worm infections, understanding how our immune system responds to them is essential. Many important parasitic worm infections move through the host lung. In this study we show that a major secreted protein in the lung, Surfactant Protein D (SP-D), is essential for immunity to a parasitic worm infection. We found that this protein binds to worm larvae in the lung to help the immune system kill them. Infecting mice that do not express SP-D with worms demonstrates SP-D is important in this immune response. These mice are unable to launch an effective anti-worm immune response and have many more worms in their intestine compared to mice that do express SP-D. We also show that if we increase SP-D levels in the lung the mouse has better immunity to worms. Together this shows for the first time that SP-D is very important for immunity to worm infections.
| Surfactant Protein (SP)-D is a constitutively expressed C-type lectin, which has a well recognized role in innate pulmonary immunity against viruses, bacteria and fungi, as well as in maintaining pulmonary homeostasis [1]. Direct SP-D interactions with immune (such as alveolar macrophages [2]) and non-immune cells [3] can protect against immune pathologies such as chronic obstructive pulmonary disorder (COPD): SP-D-deficient (SP-D-/-) mice develop spontaneous chronic lung inflammation and emphysema [4], which can be prevented by recombinant SP-D replacement [5]. SP-D binding of pathogens and allergens is also important for preventing or reducing the onset of pathology following infections such as respiratory syncytial virus (RSV) and influenza, and also protects against airway inflammation [6, 7].
SP-D is primarily secreted by alveolar epithelial type II (ATII) cells [1]; ATII cells also secrete type 2 associated alarmins such as IL-33, which are important for immunity to helminth infections [8]. SP-D also both controls and is controlled by type 2 immune responses; the canonical type 2 cytokines IL-4 and IL-13 enhance pulmonary SP-D concentrations, yet in the absence of SP-D CD4+ TH2 cytokine levels are raised [9]. SP-D therefore is likely to play an important role in limiting overzealous type 2 responses and immune-associated pathology in the lung. Control of TH2 associated immune pathologies can also be achieved by induction of regulatory innate immune cell phenotypes, such as alternatively activated macrophages (AAM) [10]. SP-D can also interact directly with myeloid cells to enhance antigen or pathogen clearance by macrophages, and also to regulate potentially damaging macrophage-driven inflammatory responses [5, 11].
Helminth infections are likely to have contributed to the evolution of both type 2/TH2 immunity and associated mechanisms that regulate the strength of this response [12, 13]. Host protective immunity against helminths is typically TH2-dependent and is initiated by parasite interaction with epithelial cells, including ATII cells [8]. Regulation of the magnitude of this TH2-mediated immune response by, for example, regulatory T cells is important for preventing immunopathology [14, 15]. How innate host factors, such as C-type lectins, are induced by helminth infection to control infection or regulate host immunity is not well understood. Some C-type lectins have been associated with helminth infection and host immunity to them. For example, Dectin-2 contributes to S. mansoni driven inflammasome activation [16]. Only one previous report has identified any interaction between SP-D and helminths; specifically that SP-D binds to fucose residues on the tegument of Schistosoma mansoni [17] however, this study did not address if this interaction contributed to host immunity.
In the study presented here we demonstrate that infection with the experimental model nematode Nippostrongylus brasiliensis induced a striking type 2-dependent increase in the levels of host SP-D. This induction of SP-D was associated with an increase in type-2 anti-parasite immune responses. Moreover, we found that immunity to infection required direct interaction of SP-D with both the fourth stage (L4) larvae and host alveolar macrophages, driving the latter to an enhanced AAM phenotype. SP-D therefore represents a previously un-described but pivotal mechanistic contributor to host immunity to helminth infection.
Type 2 cytokine-associated increases in SP-D levels have previously been shown in bronchoalveolar lavage (BAL) and serum of mice following challenge with a range of antigens and pathogens [1], but not helminths. Since the lung is an important site for immunity to N. brasiliensis infection [18, 19], we examined if host immunity to N. brasiliensis infection increased pulmonary and systemic levels of SP-D. Analysis of BAL (Fig 1A) and serum (S1A Fig) of N. brasiliensis-infected mice showed a temporal relationship between SP-D levels and progression and resolution of N. brasiliensis infection. The highest levels of SP-D were found at the peak of infection; namely day 7 post primary infection in both BAL and serum, highlighting an association with host protective immunity to N. brasiliensis. SP-D production has been shown to be dependent on IL-4, IL-13 and STAT6 [9]. Immunity to N. brasiliensis results in enhanced host secretion of IL-4 and IL-13, with IL-13 being essential for resolution of infection [20]. We investigated the requirements of IL-4 and IL-13 for SP-D production in response to N. brasiliensis infection. WT, IL-4/IL-13-/- mice (Fig 1B) and IL-4Rα-/- mice (Fig 1C) were infected with N. brasiliensis and, at 5 days post-infection, SP-D levels in BAL fluid and serum (S1B Fig) were quantified. Significantly higher SP-D levels were found in WT mice when compared to both IL-4/IL-13-/- and IL-4Rα-/- mice.
To test the association between elevated SP-D levels and immunity to N. brasiliensis we infected wild type C57/BL-6 and SP-D-deficient (SP-D-/-) mice [21] with the parasite and examined mice at days 9 and 16 post-infection (PI). At day 9 PI SP-D-/- mice had high worm burdens while wild type mice had resolved the infection, and by day 16 PI SP-D-/- mice had resolved the infection (Fig 2A). Levels of the cytokine IL-13, essential for resolution of N. brasiliensis infection, were significantly reduced at day 9 PI in the intestine, but not the lung, of SP-D-/- mice when compared to wild type mice (Fig 2B). Equivalent levels of the alarmins IL-25, IL-33 and TSLP were detected in the lung and intestine of both WT and SP-D-/- mice at days 9 and 16 PI (Fig 2C).
We also quantified the numbers and proportions of innate lymphoid cells (ILCs) and alveolar macrophages at day 9 post infection to identify if SP-D was required for the development of these cells, which are essential for optimal resolution of N. brasiliensis infection. Numbers and proportions of ILCs, which are required for resolution of N. brasiliensis infection [22], were significantly reduced at day 9 PI (Fig 2D). Proportions and numbers of alveolar macrophages (AlvM) was significantly increased in SP-D-/- mice when compared to wild type mice (Fig 2E). However, expression of a hallmark of alternative activation, resistin-like molecule (RELM) alpha/FIZZ1 (RELM-α), within AlvM was significantly decreased in SP-D-/- mice compared to wild-type by day 9 post-infection (Fig 2E). Additionally, analysis of total RELM-α protein levels in the lung revealed a significant reduction of RELM-α levels in SP-D-/- mice when compared to WT, although total levels of YM1 were equivalent in both groups (Fig 2F and S2 Fig). To establish if expression of SP-D in the lung alone was sufficient to confer protection we infected mice capable of doxycycline-inducible expression of SP-D only in the lung (CCSP-rtTA, (tetO)7-rSP-D,SP-D−/−). Significantly reduced N. brasiliensis numbers in the intestine in doxycycline treated mice (SP-Don) when compared to untreated mice (SP-Doff) demonstrated that SP-D expression in the lung is a major component of host immunity against N. brasiliensis (S2 Fig). However, numbers of worms in SP-Don mice were significantly higher than in WT mice indicating possible SP-D mediated inputs to immunity at other sites.
SP-D control of infection is associated with its ability to opsonize and enhance immune recognition of pathogens. We assessed the ability of a recombinant homotrimeric fragment of human SP-D (rfhSP-D) to directly bind to N. brasiliensis by immunofluorescence using third stage larvae (L3), fourth stage larvae (L4), and adult worms. rfhSP-D binding to the surface of parasites was only seen for L4 (Fig 3A and S3 Fig). Moulting of N. brasiliensis L3 to L4 takes place in the lungs of the host. This would suggest that elevated pulmonary SP-D levels seen in the lung (Fig 1) could induce protective immunity by coating L4 parasites and enhanced immune recognition of these parasites both in the lung and intestine. We tested this by infecting mice intranasally with rfhSP-D-coated L4, prior coating with rfhSP-D did not result in reduced parasite viability (S3 Fig). Intranasal infection of mice with rfhSP-D-coated L4 resulted in reduced worm burdens when compared to mice that received L4 alone (Fig 3B). Examination of host alarmin responses revealed no effect on IL-33 or IL-25, but a significant reduction in TSLP (Fig 3C). Recipients of rfhSP-D-coated L4 had increased numbers of ILC2 while AlvM populations showed a heightened expression of the markers of alternative activation RELM-α chitinase-3-like protein Chil3 (Ym1) (Fig 3D).
Having established in a loss-of-function system that SP-D is required for immunity to N. brasiliensis and that rfhSP-D binding of L4 parasites could enhance host immunity to the infection we next tested if heightened levels of pulmonary SP-D prior to N. brasiliensis infection conferred enhanced immunity against the parasite. Naïve mice treated with rfhSP-D for 4 days did not display detectable levels of IL-13 or IL4 in the lung. Numbers and proportions of ILC2s were also equivalent (S4 Fig). At day 5 post-infection, intra-nasal administration of rfhSP-D prior to N. brasiliensis infection (Fig 4A) significantly reduced intestinal burdens of adult N. brasiliensis in rfhSP-D-treated mice compared to bovine serum albumin (BSA)-treated controls (Fig 4B). Protection was associated with increased pulmonary type 2 cytokines IL-4 and IL-13 but not IL-33 at day 5 post infection (Fig 4C). Additionally, type 2 innate lymphoid cells (ILC2) required to control N. brasiliesis infection were increased in numbers and proportions in the lung of rfhSP-D treated mice. Moreover, a higher percentage of ILC2s produced IL-13 in rfhSP-D treated mice when compared to untreated mice (Fig 4D).
SP-D is known to regulate alveolar macrophage function [23, 24]. SP-D-/- mice have a loss of homeostatic regulation of macrophage function [25, 26] which can be rescued by treatment with recombinant rat SP-D [11]. Host control of N. brasiliensis recall infection is dependent on IL-4Rα-dependent macrophage polarization to the alternatively activated phenotype in the lung [27]. AAMs are also key effector cells for controlling other helminth infections including Heligmosomoides polygyrus [28] and Schistosoma mansoni [29]. As SP-D enhanced immunity to N. brasiliensis in the lung, we hypothesized that this may be due, at least in part, to an effect on polarisation of AlvM to AAM.
Intranasal administration of rfhSP-D resulted in an increased expression of the AAM markers YM1 and Relm-α in AlvMs when compared to BSA-treated controls at day 5 post-infection (Fig 5A). To test if these SP-D-dependent enhanced AAM response contributed to increased protection against N. brasiliensis, we isolated AlvMs from mice which had been infected with N. brasiliensis and treated with rfhSP-D or not, and transferred these cells intra-nasally into naïve mice. Recipients of rfhSP-D-treated macrophages had reduced intestinal worm burdens at day 5PI when compared to recipients of untreated macrophages (Fig 5B and S5 Fig).
It is typically considered that polarization to alternative activation by macrophages following an N. brasiliensis infection is a result of them responding to elevated levels of the cytokines IL-4 and IL-13 [30]. We directly tested in vitro whether rfhSP-D enhanced alternative activation of AlvMs isolated from naïve mice. Naïve AlvMs were polarized to AAM by ex vivo culture with recombinant IL-4/IL-13 in the presence or absence of rfhSP-D. Co-culture with rfhSP-D increased Ym1 and RELM-α expression when compared to macrophages treated with IL-4/IL-13 alone (Fig 5C). Collectively, these data show that SP-D can enhance AAM-dependent immunity to N. brasiliensis.
SP-D typically confers protection against infection by binding of the carbohydrate recognition domain (CRD) to the pathogen and, improving opsonisation and neutralization [31]. We tested if such an interaction contributed to SP-D-induced protection against N. brasiliensis. The ability of the rfhSP-D CRD domain to bind antigen can be blocked with maltose. We used this approach to block the ability of rfhSP-D to bind via the CRD to N. brasilieinsis. Mice treated with maltose-blocked rfhSP-D had higher worm burdens than mice treated with rfhSP-D alone; moreover, mice treated with rfhSP-D blocked with maltose did not demonstrate enhanced ILC2 and AAM responses seen in mice treated with rfhSP-D alone (Fig 6A).
We next tested if rfhSP-D opsonisation of L4 N. brasiliensis could enhance killing by alveolar macrophages. Macrophages isolated into serum-free medium from the lungs of N. brasiliensis infected mice were added to L4 N. brasiliensis in the presence or absence of rfhSP-D. In the presence of rfhSP-D and alveolar macrophages the ratio of live to dead parasites decreased when compared to parasites cultured with macrophages only (Fig 6B).
To date it has never been demonstrated if SP-D or any other member of the collectin sub-family of C-type lectins can directly mediate innate protection against a parasitic helminth. In this study we show for the first time that SP-D is an important component of innate immunity to helminth infection.
We find that levels of SP-D in the lung and serum increase significantly and rapidly, in a manner analogous to an alarmin response, following infection with N. brasiliensis. This is in agreement with other reports where SP-D levels increase in response to acute lung stress following pulmonary challenges with, for example, lipopolysaccharide [32], bleomycin [33] and ovalbumin [34] Similarly, SP-D levels increase in the BAL following infections by Aspergillus fumigatus [35], Actinobacillus pleurpneumoniae or Staphylococcus aureus [36]. Our data represents the first known report of increased SP-D levels in response to a helminth infection.
Elevated SP-D production can also be driven by the type 2/TH2 cytokines IL-4 and IL-13, and in turn SP-D can impart negative feedback control of type 2/TH2 responses; indeed in the absence of SP-D TH-2 cytokine levels are raised [9]. Our findings expand on this understanding by demonstrating that production of SP-D following helminth infection is significantly dependent on key protective immune responses against N. brasiliensis; namely IL-4/IL-13 signaling via IL-4Rα. Previous demonstrations of increased type 2 immunity in SP-D-/- mice [9] may have suggested that immunity to N. brasiliensis infection [37] would have been enhanced in SP-D-/- mice. This was not the case and we account for this by exploring other roles for SP-D in controlling N. brasiliensis infections. Our results demonstrated that SP-D deficiency impaired innate type-2 responses associated with immunity to N. brasiliensis infection [14]; moreover, we also found that elevated SP-D enhances these responses. This suggests that SP-D is an important modulator of protective ILC2 and alveolar macrophage responses against N. brasiliensis.
Opsonisation of pathogens and allergens by SP-D enhances host neutralization of them in the lung, primarily by improved recognition by host cells, such as alveolar macrophages [38]. It is well established that SP-D opsonization enhances innate immunity against a range of pathogens, such as bacteria, fungi and viruses [39]. We demonstrate that SP-D can also act as an interface between the L4 parasite and alveolar macrophages (which are key effector cells for controlling the parasite in the lung [27]). Moreover, direct interaction of SP-D with alveolar macrophages enhanced their polarization to an alternative phenotype. These findings demonstrate a pivotal role for alveolar macrophages in mediating the effects of SP-D via macrophage alternative activation dependent parasite immunity.
In accordance with studies demonstrating the requirement of the carbohydrate recognition domain (CRD) in mediating pathogen binding and enhanced phagocytosis and clearance by neutrophils and macrophages [31], our studies also demonstrate that SP-D-dependent clearance of N. brasiliensis is dependent on the CRD. Furthermore, SP-D binding through the CRD promotes innate type 2 responses including ILC2 induction and alternative activation of alveolar macrophages. We also show that SP-D can directly enhance L4 killing by alveolar macrophages following exposure to N. brasiliensis infection. These data clearly show that SP-D can act as an opsonin of L4 N. brasiliensis to enhance parasite killing by alveolar macrophages.
Our results therefore clearly demonstrate that binding of SP-D to the L4 lung stage of N. brasiliensis promotes parasite clearance via induction of innate type 2 responses including alternative activation of macrophages but also enhanced ILC2 expansion. To the best of our knowledge this is the first description of SP-D influencing ILC2 biology and we suggest that the decreased type 2 cytokine levels in the lungs may be a function of a loss of an interaction between ILC2 and SP-D in SP-D-/- mice.
Additionally, our in vivo studies show that SP-D deficiency increased the proportions of alveolar macrophages in the lung, key cells that maintain lung homeostasis and promote parasite clearance [10, 27]. Our observation of increased alveolar macrophage numbers is in agreement with other studies showing increased numbers of alveolar macrophage numbers in SP-D-/- mice [40]. However, reduced expression of the markers of alternative macrophage activation such as RELM-α within alveolar macrophages suggests that SP-D-deficiency does not favour induction of resolving macrophage populations in response to N. brasiliensis infection. Moreover, we show that in the absence of SP-D total levels of RELM-α in the lung were significantly reduced. Like SP-D, RELM-α regulates type 2 immunity; RELM-α-/- mice develop heightened pathology following experimental S. mansoni infection [41]. Therefore, in addition to having impaired induction of AAM to act directly on the parasite SP-D-/- mice lack appropriate induction of other mediators of immune regulation which may have a wider impact on host control of infection induced pathology [15].
Induction of type 2 immunity to helminths is significantly dependent on epithelial cells (including ATII cells, the main cellular source of SP-D) secreting cytokines such as IL-33 [8] along with other immune modulators such as RELM-ß [42], trefoil factor 2 (TFF2) [43] and TSLP [44]. Our findings show that helminth induced SP-D is an additional major player in the host epithelial response to helminths. Balanced type 2 immunity is characteristic of effective host control of parasitic helminth infections and also reduced susceptibility to allergic disease [14]. Our findings therefore may have broader relevance to understanding innate immune control of diseases associated with poor control of type 2/Th2 immunity. Moreover, as helminth-induced SP-D is able to modify innate cell function and directly control lung inflammation, our studies set a precedent for placing SP-D in a central role of mediating parasite-associated protection from, for example, allergy and pulmonary viral infection
In conclusion, we show for the first time in both gain of function and loss function approaches that SP-D is required for immunity against N. brasiliensis. This enhanced immunity is coincident with an increased induction of cells associated with the resolution of infection; namely ILC2 and alternatively activated macrophages. Thus, helminth induction of SP-D is essential for host resolution of helminth infection.
6-10-week-old BALB/c, C57/BL6, IL-4/13-/-[45], IL-4Rα-/-[46], SP-D-/- and CCSP-rtTA, (tetO)7-rSP-D,SP-D−/− [21, 47] mice were obtained from colonies maintained by the University of Cape Town specific-pathogen-free animal facility. The authors are grateful to the laboratories of J Whitsett and S Hawgood for use of SP-D transgenic mice originally generated in their laboratories. Section 20 dispensation to carry out animal work at UCT was granted nationally by the South African Government Department of Agriculture Fisheries and Food and institutionally by the UCT Health Sciences Animal Ethics Committee (Project licence 012/054) to be in accordance with guidelines laid down by the South African Bureau of Standards. All researchers were accredited by the South African Veterinary Council. Dispensation to carry out animal research at Imperial College was approved by the Imperial College Animal Welfare and Ethical Review Body and granted by the UK government Home Office; as such all research here was carried under a specific project licence (PPL70/6957).
Mice were infected subcutaneously with 500 N. brasiliensis L3 larvae and killed at various times post infection as previously described [48]. Intra-nasal infection with rfhSP-D coated L4 N. brasiliensis was carried out using techniques adapted from Harvie et al [19]. Briefly, L4 N. brasiliensis were isolated 2 days post- N. brasiliensis infection from lung tissues. L4 N. brasiliensis were then incubated with rfhSP-D or bovine serum albumin (BSA) control for 1 hour at 37°C. L4 N. brasiliensis infection of mice was carried out by intranasal administration of 250 viable L4 worms in 50 μl to lightly anaesthetized mice. Adult worm burdens were determined by removing the small intestine and exposing the lumen by dissection. The intestines were incubated at 37°C for 4 hrs in 0.65% sodium chloride (NaCl) to allow the worms to migrate out, after which the numbers of worms were counted under a dissecting microscope.
Mice were lightly anaesthetized and treated with 20 μg of rfhSP-D or BSA control (diluted in PBS) intra-nasally in a volume of 50 μl on day 0, 1, 2 and 3. For blocking of rfhSP-D CRD head region, 20 mM Maltose was incubated with rfhSP-D in the presence of 1 mM calcium chloride (CaCl2) for 1 hr at 37°C. Mice were killed on days indicated in results.
Mice underwent bronchoalveolar lavage (BAL) by administration of 1 ml sterile PBS containing 0.25 mM Ethylenediaminetetraacetic acid (EDTA) intra-tracheally. Lungs were lavaged three times. BAL fluid was centrifuged at 1200 rpm for 5 mins and the supernatant was frozen at -80°C.
The left lobe of the lung was snap frozen in liquid nitrogen and subsequently stored at -80°C until analysis. To prepare lung homogenates, 400 μl of lysis buffer [49] was added to lung tissue prior to mechanical homogenization. Homogenates were centrifuged at 14000 rpm for 20 mins and the protein concentrations of the supernatants determined using bicinchoninic acid (BCA) assay (Pierce, Rockford, IL).
Approximately 500 μl of blood was collected by cardiac puncture and the serum isolated by centrifugation (4000 rpm for 20 mins).
Whole lungs were removed from individual mice, finely cut and digested in Iscove’s modified Eagle medium (IMDM) (Invitrogen) containing 50 U/ml collagenase type I (Invitrogen) and 13 μg/ml DNase (Roche) at 37°C for 90 mins. Digested lung tissue were pushed through 70 or 100 μm nylon cell strainer (Becton Dickson, New Jersey) and subjected to red cell lysis.
1 x 106 single cell suspensions from individual lungs were stained in MACS buffer with anti-Siglec-F PE (E50/2440) and anti-CD11c APC (HL3) antibodies to stain for AlvMs and eosinophils. To stain for ILC2, the antibodies anti-Lineage PE (CD3, CD19, CD11b, FceR1, Ter119, CD4, CD8, B220, Ly6G/6C), anti-CD45 AF700, anti-CD90 PacBlu, anti-CD127 PE-Cy7 or APC (SB/199), anti-ICOS Biotin or FITC (7E.17G9) anti- Sca-1 V450 (D7) and anti-T1/ST2 FITC PerCP-Cy5.5 (DJ8) were used. Anti-FcR (2.4G2) was used to block non-specific binding of immunoglobulins to the FCγII/III receptors.
For intracellular staining, cells were stained with surface markers, fixed in 2% paraformaldehyde before being permeabilized with buffer containing saponin. Cells were subsequently stained with anti-Ym1 biotin (ECF-L) and anti-RELM-α (E19). For intracellular staining of ILC2 cells were fixed and permeabilized with Fix/perm (eBioscience) and stained with IL-13 PE-Cy7. Cells were acquired using a FORTESSA Flow cytometer (BD Biosciences) and data analyzed using FlowJo software (Tree star, inc., Ashland, Oregon, USA). Antibodies were purchased from BD Pharmingen, eBioscience and Biolegend.
Whole lung and intestine homogenates were used to quantify cytokines, and BAL fluid or serum from N. brasiliensis infected mice were analyzed for SP-D content by ELISA. 96-well flat-bottom plates (Nunc Maxisorp; Thermo Fisher Scientifica, Roskilde, Denmark) were coated overnight at 4°C with 50 μl of appropriate coating antibody diluted in 1X PBS. Plates were washed four times in Tris-Buffered Saline containing % Tween (TBST), blocked with 200 μl blocking buffer overnight at 4°C. Following this, appropriate dilutions of samples and standards were prepared, loaded into wells and incubated overnight at 4°C. The plates were further washed and 50 μl of appropriate biotinylated secondary antibodies were added and incubated at 37°C for 3 hrs. 50 μl of Streptavidin-coupled horseradish peroxidase (HRP) was then added after washing and plates were incubated for 1 hr at 37°C. The plates were developed with 3,3’,5,5’-Tetramethylbenzidine (TMB) microwell peroxidase substrate system, and the reaction was stopped with 1 M H3PO4. The plates were read at an absorbance of 450 nm using a VersaMax microplate reader (Molecular Devices Corporation, Sunnyvale, CA, U.S.A). All antibodies were from BD Pharmingen, San Diego, CA.
Mice were treated with 20 μg of rfhSP-D or BSA at D0, 1, 2, 3, 6 and 7 post-infection. Single-cell suspensions of pooled lungs were prepared at day 8 post-infection and AlvMs were stained with anti-CD11c APC-conjugated and anti-Siglec-F PE conjugated monoclonal antibody (MAb) (BD Pharmingen) before they were isolated (> 95% purity) as CD11c+Siglec F+Autoflourescenthigh using a FACS Aria cell sorter (Becton Dickinson), purity was also confirmed by microscopic analysis (S5A Fig). 1 x 105 macrophages were then transferred intra-nasally into naïve BALB/c mice 24 hrs prior to N. brasiliensis infection.
Naïve AlvMs (CD11C+Siglec-F+AutoFlourescenthigh) were isolated from single cell suspensions of lung tissue by FACS Aria as described above and plated in duplicates at 4x105 cells per well. Cells were stimulated with either recombinant mouse IL-4/IL-13, IL-4/IL-13 + 20 μg/ml of rfhSPD or left untreated. The cultures were incubated for 60 hrs at 37°C. Thereafter, cells were washed and stained for alternative activation markers, YM1 and Relm- α as described above, and acquired with LSRFORTESSA (BD Biosciences).
Mice were infected with 500 L3 N. brasiliensis and lungs isolated at day 7 post-infection. Single cell suspensions were stained for Siglec-F and CD11c and live cells isolated into serum free media using a FACS Aria, as described above. L4 were isolated from lungs as described previously. Experiments were carried using serum free media. L4 were either left untreated or incubated with 20 μg/ml SP-D for 1 hr before addition to 4x105 macrophages. After 48 hrs larvae were washed, counted and analysed for movement by bright field microscopy and a sequence of 20 images/min were taken. These were then analysed by SD overlay using Fiji software. Live/dead ratios were calculated using total, moving and dead numbers.
L3, L4 and adult stage larvae of N. brasiliensis were fixed overnight in 2% paraformaldehyde at 4°C. The larvae were extensively washed using 1X PBS containing 0.2% BSA and 1 mM CaCl2. Non-specific binding was blocked by incubation of the larvae in 0.2% BSA in PBS for 1 hr at room temperature. Thereafter, the larvae were incubated with 20 μg/ml rfhSP-D in PBS containing 0.2% BSA and 1mM CaCl2 for 1 hr at 32°C. After extensive washing, the larvae were incubated with biotinylated rabbit anti-rfhSP-D antibody (HYB 246.04, Antibody shop) used at 1/200 and left overnight at 4°C. To detect the SP-D binding, the organisms were subsequently incubated with Streptavidin cy3 (1/500) for 2 hrs at room temperature (RT). All sections were viewed with Axiovert LSM 510 Meta NLO microscope (Zeiss).
Data were expressed as mean ± standard deviation and analyzed using Mann-Whitney nonparametric T test or ANOVA with a 95% confidence interval. p values are represented as p< 0.05 (*), p<0.01 (**) and p<0.005 (***).
|
10.1371/journal.pntd.0001449 | IFN-γ Production Depends on IL-12 and IL-18 Combined Action and Mediates Host Resistance to Dengue Virus Infection in a Nitric Oxide-Dependent Manner | Dengue is a mosquito-borne disease caused by one of four serotypes of Dengue virus (DENV-1–4). Severe dengue infection in humans is characterized by thrombocytopenia, increased vascular permeability, hemorrhage and shock. However, there is little information about host response to DENV infection. Here, mechanisms accounting for IFN-γ production and effector function during dengue disease were investigated in a murine model of DENV-2 infection. IFN-γ expression was greatly increased after infection of mice and its production was preceded by increase in IL-12 and IL-18 levels. In IFN-γ−/− mice, DENV-2-associated lethality, viral loads, thrombocytopenia, hemoconcentration, and liver injury were enhanced, when compared with wild type-infected mice. IL-12p40−/− and IL-18−/− infected-mice showed decreased IFN-γ production, which was accompanied by increased disease severity, higher viral loads and enhanced lethality. Blockade of IL-18 in infected IL-12p40−/− mice resulted in complete inhibition of IFN-γ production, greater DENV-2 replication, and enhanced disease manifestation, resembling the response seen in DENV-2-infected IFN-γ−/− mice. Reduced IFN-γ production was associated with diminished Nitric Oxide-synthase 2 (NOS2) expression and NOS2−/− mice had elevated lethality, more severe disease evolution and increased viral load after DENV-2 infection. Therefore, IL-12/IL-18-induced IFN-γ production and consequent NOS2 induction are of major importance to host resistance against DENV infection.
| Dengue fever and its severe forms, dengue hemorrhagic fever and dengue shock syndrome, are the most prevalent mosquito-borne diseases on Earth. It is caused by one of four serotypes of Dengue virus (DENV-1–4). At present, there are no vaccines or specific therapies for dengue and treatment is supportive. Host response to infection is also poorly understood. Here, using a DENV-2 strain that causes a disease that resembles the severe manifestations of Dengue in humans, we demonstrate that IFN-γ production is essential for the host to deal with infection. We have also shown that IFN-γ production during DENV infection is controlled by the cytokines IL-12 and IL-18. Finally, we show that one of the mechanisms triggered by IFN-γ during host response to DENV infection is the production of Nitric Oxide, an important virustatic metabolite. Mice deficient for each of these molecules present marked increase in DENV replication after infection and more severe disease. Altogether, this study demonstrates that the IL-12/IL-18-IFN-γ-NO axis plays a major role in host ability to deal with primary DENV infection. These data bear relevance to the understanding of antiviral immune responses during Dengue disease and may aid in the rational design of vaccines against DENV infection.
| Dengue fever (DF) and its severe forms, dengue hemorrhagic fever (DHF) and dengue shock syndrome (DSS), are mosquito-borne diseases caused by one of four serotypes of Dengue virus (DENV-1–4). Fifty to 100 million cases of DF are estimated annually mostly in tropical and subtropical regions of the world [1]–[3]. According to the World Health Organization (WHO), around 500,000 patients develop the severe forms of dengue and 20,000 deaths are estimated to occur each year. DHF is defined by the WHO as fever with hemorrhagic manifestations, thrombocytopenia, and hemoconcentration or other signs of plasma leakage [2]. Treatment of DF and of the severe forms of dengue infection is largely supportive. The large number of infected individuals, the lack of clinical or laboratory markers that indicate which patients will develop severe disease and the lack of specific treatment place an enormous burden on health systems of low income countries [2].
The pathogenesis of DENV infection remains poorly understood and involves a complex interplay of viral and host factors. Risk factors for severe disease include age [1], [4], viral serotype [1], [5] and genotype [1], [6], [7], and the genetic background of the host [1], [8], among others. Retrospective and prospective human studies have demonstrated that secondary infection by a heterologous serotype is the single greatest risk factor for DHF/DSS [9]–[11]. However, severe disease may also occur after primary infection [5], [12], [13]. In both cases, there appears to be a correlation between disease severity and viral load [9]–[13]. In addition, the immunopathogenesis of DENV probably involves the effects of cytokines on both infected and bystander immune cells, hepatocytes, and endothelial cells [2], [3], [13]. There are several studies which show enhanced levels of IFN-γ in dengue-infected humans but the precise role of IFN-γ in clinical dengue is somewhat controversial [14]–[16]. There are studies which suggest that levels of this cytokine may correlate positively with disease in humans [16], but other studies have shown that increased IFN-γ production correlated with higher survival rates in DHF patients [15]. In experimental systems, non-adapted viruses usually are unable to reach high viral loads, except in mice deficient for IFN receptors, suggesting that IFN-γ and its receptors are necessary for host resistance to Dengue infection [17]–[19]. However, the major cell types producing IFN-γ, mediators controlling production of this cytokine and major effector mechanisms triggered by IFN-γ are not known.
Optimal IFN-γ production in various infections models in mice is controlled by cytokines, especially IL-12 and IL-18 [20], [21]. The IFN-γ produced may then upregulate inducible nitric oxide synthase (NOS2), resulting in high levels of NO production by dendritic cells and macrophages [22]. NO is known to possess potent antiviral activities [22]. Therefore, in order to examine the role played by these molecules during dengue disease we conducted infection experiments in mice infected with an adapted strain of DENV-2. This unique DENV-2 strain was chosen because it was previously shown to induce in immunocompetent mice a disease that resembles severe dengue cases in humans [23]–[25], what does not happen with most non-adapted strains usually utilized in experimental settings [2], [3]. We show that disease is more severe and there are higher viral loads after DENV-2 infection of IFN-γ-deficient mice. Furthermore, we demonstrate that the combined action of IL-12 and IL-18 is necessary for optimal IFN-γ production and control of DENV-2 infection. Finally, we show that IFN-γ controls expression of NOS2 and NO production after DENV-2 infection and that NO production is crucial for resistance of the murine host to infection with DENV.
This study was carried out in strict accordance with the Brazilian Government's ethical and animal experiments regulations. The experimental protocol was approved by the Committee on the Ethics of Animal Experiments of the Universidade Federal de Minas Gerais (CETEA/UFMG, Permit Protocol Number 113/09). All surgery was performed under ketamine/xylazine anesthesia, and all efforts were made to minimize suffering. The guidelines followed by this Committee are based on the guidelines of Animal Welfare Act (AWA) and associated Animal Welfare Regulations (AWRs) and Public Health Service (PHS) Policy.
Mice deficient for IFN-γ and NOS-2 were obtained from The Jackson Laboratory and were bred and maintained at the Gnotobiology and Immunology Laboratory of Instituto de Ciências Biológicas. Mice deficient for IL-12p40 were kindly provided by Dr. J. Magran through Dr. L. V. Rizzo (Instituto de Ciências Biomédicas (ICB), University of São Paulo, São Paulo, Brazil) and were bred and maintained at the Gnotobiology and Immunology Laboratory of Instituto de Ciências Biológicas. Mice deficient for IL-18 [26] were kindly provided by Dr. F.Q. Cunha and were bred and maintained at the Gnotobiology and Immunology Laboratory of Instituto de Ciências Biológicas. Mice deficient for IL-23p19 [27] were bred and maintained at the animal facility of the Transgenose Institute (CNRS, Orleans). All mice were on C57BL/6J genetic background (back-crossed at least 10 times) and wild-type control C57BL/6J (WT) mice were used, except for IL-18-deficient mice, that were on the BALB/c background and were compared to their proper WT littermates. For experiments, 7–10 weeks old mice were kept under specific pathogen–free conditions, in filtered-cages with autoclaved food and water available ad libitum.
An adapted Dengue virus 2 (DENV-2) strain was obtained from the State Collection of Viruses, Moscow, Russia [23]. Briefly, the virus had undergone two passages in the brain of BALB/c suckling mice. Five days after infection, brains of moribund mice were harvested for preparing 10% (w/v) brain suspension in modified Eagle's medium (MEM). After that, eight sequential passages through BALB/c mice of different ages (1–4 weeks old) by intraperitoneal (i.p.) injection were performed. Two sequential passages were carried out for each age group of. After each passage, the brains of the moribund mice were harvested for preparing 10% brain suspension and then used for the next passage. The last passage of DENV-2 strain P23085 was performed in neonatal mice to produce stocks which were stored as 10% brain suspension at −70°C. Sequences of portions of E and NS1 genes of the adapted virus were deposited previously at GenBank under the accession number AY927231 [22]. Virus adaptation was performed in a biosafety level-3 (BSL-3) facility of the SRC VB «Vector», Russia, Koltsovo. After adaptation, monolayers of Aedes albopictus C6/36 cell line were infected with DENV-2 strain P23085 at a multiplicity of infection (MOI) of 0.05 PFU/cell and incubated at 28°C for 5–7 days. The cultured medium was harvested after cytopathic effect was noticed and cell debris removed by centrifugation. The virus supernatant was collected and stored at −70°C until use. The cultured medium of mock-infected monolayers of Aedes albopictus C6/36 cell line was used as control of the infection. To calculate virus titer, expressed as LD50, in the harvested cultured medium, groups of ten mice were inoculated i.p. with serial dilutions of the virus and lethality recorded. The titer of our DENV-2 stock was 105 LD50/ml of suspension, as calculated in 8–10-week-old BALB/c mice. 1LD50 corresponded to 20 PFU of the adapted DENV-2 strain.
For infection experiments, the virus-containing cell-supernatant was diluted in endotoxin-free PBS and injected i.p. into mice. For the evaluation of lethality, mice were inoculated i.p. with DENV-2 virus and lethality rates evaluated every 12 h. The various other parameters were evaluated at 3, 5 or 7 days after i.p. inoculation of the virus. In all experiments using genetically deficient mice, experiments with the relevant WT controls were performed in parallel. Non-infected (NI) animals were inoculated with suspension from non-infected cell supernatant diluted in a similar manner. In the experiments involving genetically deficient mice, the NI group represents the pooled results obtained from the analysis of deficient mice and WT non-infected mice. Results were pooled for ease of presentation.
In some experiments IL-18 was neutralized by daily i.p. injection of 250 µg of recombinant human IL-18BP per animal (hIL-18 bp), starting 1 hour after DENV-2 inoculation until day 4 after virus inoculation. The dose was chosen based in a previous study [28]. Control animals received PBS. The hIL-18 bp isoform was a kind gift of Dr. Amanda Proudfoot from Merck-Serono Pharmaceuticals (Geneve, Switzerland).
Murine bone marrow cells were isolated from femurs and were differentiated into myeloid DCs after culturing at 2×106 cells/ml for 10 days in RPMI supplemented with 10% FBS and 4% J558L cell-conditioned medium as a source of GM-CSF as described [29]. DCs were plated in 96-well microculture plates (at 2×105 cells/well in DMEM supplemented with 2 mM l-glutamine and 2×10−5 M 2-ME) and for infection, cells were incubated with 50 µL of the cell supernatant suspension containing DENV-2 at 0,01 MOI in the presence or not of recombinant murine IFN-γ (100 U/ml). Negative controls were stimulated with sterile cell supernatant obtained from mock infected cells.
Mice were assayed for viral titers in spleen. For virus recovery in spleen, the organ was collected aseptically and stored at −70°C until assayed for DENV-2 virus. Tissue samples were weighed, grounded by using a pestle and mortar and prepared as 10% (w/v) homogenates in minimal essential medium (MEM) without fetal bovine serum (FBS). Viral load in the supernatants of tissue homogenates assessed by direct plaque assays using LLC-MK2 cells cultured in agarose overlay. Briefly, organ homogenates were diluted serially and a 0.4 ml volume placed in duplicate into each of 6-wells of LLC-MK2 cell monolayers and incubated for 1 h. An overlay solution containing 2× MEM and 1% agarose in equal volumes was added to each well and the cultures incubated for 7 days. Cultures were stained with crystal violet for enumeration of viral plaques. Cell monolayers incubated with tissue homogenates of not infected mice were used as control for the assay. The results were measured as plaque forming units (PFU) per gram of tissue weight. The limit of detection of the assay was 100 PFU/g of tissue.
The concentration of cytokines (TNF-α, IFN-γ, IL-6, IL-12p40, IL-12p70 and IL-18) in serum or tissue samples was measured using commercially available antibodies and according to the procedures supplied by the manufacturer (R&D Systems, Minneapolis, except for IL-18, manufactured by BD Pharmingen). Serum was obtained from coagulated blood (15 min at 37°, then 30 min a 4°C) and stored at −20°C until further analysis. One hundred milligrams of tissues (liver and spleen) was homogenized in 1 ml of PBS containing anti-proteases (0.1 mM phenylmethilsulfonyl fluoride, 0.1 mM benzethonium chloride, 10 mM EDTA and 20 KI aprotinin A) and 0.05% Tween 20. The samples were then centrifuged for 10 min at 3000 g and the supernatant immediately used for ELISA assays. The detection limit of the ELISA assays was in the range of 4–8 pg/ml.
Cell-free culture medium was obtained by centrifugation and assayed for nitrite content, determined by the Griess method [30]. For this assay, 0.1 ml of culture medium was mixed with 0.1 ml of Griess reagent in a multiwell plate, and the absorbance at 550 nm read 10 min later. The NO2− concentration (µM) was determined by reference to a NaNO2 standard curve.
Blood was obtained from the brachial plexus in heparin-containing syringes at the indicated times. The final concentration of heparin was 50 u/ml. Platelets were counted in a Coulter Counter (S-Plus Jr). Results are presented as number of platelets per µl of blood. For the determination of the hematocrit, a sample of blood was collected into heparinized capillary tubes and centrifuged for 10 min in a Hematocrit centrifuge (HT, São Paulo, Brazil).
Aspartate transaminase activity was measured in individual serum samples, using a commercially available kit (Bioclin, Belo Horizonte, Brazil). Results are expressed as the U/dL of serum.
Total RNA was isolated from Spleen of mice for evaluation of NOS2 mRNA expression. RNA isolation was performed using Illustra RNAspin Mini RNA Isolation Kit (GE Healthcare). The RNA obtained was resuspended in diethyl pyrocarbonate treated water and stocked at −70°C until use. Real-time RT-PCR was performed on an ABI PRISM 7900 sequence-detection system (Applied Biosystems) by using SYBR Green PCR Master Mix (Applied Biosystems) after a reverse transcription reaction of 2 µg of total RNA by using M-MLV reverse transcriptase (Promega). The relative level of gene expression was determined by the comparative threshold cycle method as described by the manufacturer, whereby data for each sample were normalized to hypoxanthine phosphoribosyltransferase and expressed as a fold change compared with non-infected controls. The following primer pairs were used: hypoxanthine phosphoribosyltransferase, 5′-GTTGGTTACAGGCCAGACTTTGTTG-3′ (forward) and 5′-GAGGGTAGGCTGGCCTATAGGCT-3′ (reverse); and nos2, 5′- CCAAGCCCTCACCTACTTCC -3′ (forward) and 5′- CTCTGAGGGCTGACACAAGG -3′ (reverse).
Spleen cells were evaluated ex vivo for extracellular molecular expression patterns and for intracellular cytokine expression patterns. Briefly, spleens were removed from infected mice at the indicated timepoints. Then cells were isolated, and immediately stained for surface markers, fixed with 2% formaldehyde and then permeabilized with a solution of saponin and stained for 30 min at room temperature, using conjugated anti-IFN-γ monoclonal antibodies. Preparations were then analyzed using a FACScan (Becton Dickinson), and 50 000 gated events on total lymphocyte/monocyte population were acquired for later analysis. Figure S1A shows the gating strategy utilized for IFN-γ+ population analysis in CD4+ cells. Briefly, lymphocyte/monocyte population was isolated in gate R1. At this region, the cell population positive for the surface marker of interest was isolated (R2) and among cells in this region, IFN-γ+ cells were obtained (R3). Analogous strategies were utilized for the other several populations studied. The antibodies used for the staining were rat immunoglobulin controls, anti-CD4-PE, anti-CD8-PE, anti-NK1.1-PE, anti-CD3- PE-Cy5 and anti-IFN-γ-FITC (all from Biolegend Inc). Analysis was conducted using the software Flow Jo 7.2 (Tree Star Inc).
A portion of liver was obtained from killed mice at the indicated time points, immediately fixed in 10% buffered formalin for 24 hours and tissues fragments were embedded in paraffin. Tissue sections (4 µm thick) were stained with hematoxylin and eosin (H&E) and examined under light microscopy or collected in serial sections on glass slides coated with 2% 3-aminopropyltriethylsilane (Sigma Aldrich, St. Louis, MO). The latter sections were deparaffinized by immersion in xylene, and this was followed by immersion in alcohol and then incubation with 3% hydrogen peroxide diluted in Tris-buffered saline (TBS) (pH 7.4) for 30 minutes. The sections were then immersed in citrate buffer (pH 6.0) for 20 minutes at 95°C for antigen retrieval. The slides were then incubated with the rabbit polyclonal anti-NOS2 (N-20, sc-651, Santa Cruz Biotechnology, Santa Cruz, CA) diluted 1∶100; at 4°C overnight in a humidified chamber. After washing in TBS, the sections were treated with a labeled streptavidin-biotin kit (LSAB, K0492, Dako, Carpinteria, CA). The sections were then incubated in 3,3′-Diaminobenzidine (K3468, Dako) for 2 to 5 minutes, stained with Mayer's hematoxylin and covered. Negative controls were obtained by the omission of primary antibodies, which were substituted by 1% PBS-BSA.
Results are shown as means ± SEM. Differences were compared by using analysis of variance (ANOVA) followed by Student-Newman-Keuls post-hoc analysis. Differences between lethality curves were calculated using Log rank test (Graph Prism Software 4.0). Results with a P<0.05 were considered significant.
An initial set of experiments were carried out to assess the kinetics of IFN-γ production and major IFN-γ producing cell types after DENV-2 infection. As shown in Figure 1, there was an increase in serum and splenic levels of IFN-γ from the 5th day of infection (Figure 1A). Levels of IFN-γ enhanced further at day 7 in both serum and spleen (Figure 1A). In spleen, IFN-γ staining was detected in about 10% of total cells in the 5th day after inoculation and reached about 15% at the 7th day post infection (Figure 1B and Figure S1B). CD3−NK1.1+ NK cells and CD3+NK1.1+ NKT populations presented increased proportions of IFN-γ staining at the 5th day post infection (Figure 1B and Figure S1E and S1F). In addition, there was increase in expression of IFN-γ on all cell populations analyzed at day 7 after infection (Figure 1B). Significantly, over 30% of CD4+ T cells, 25% of CD8+ T cells, 40% of CD3−NK1.1+ NK cells and CD3+NK1.1+ NKT cells were IFN-γ+ at day 7 after infection (Figure 1B and Figures S1C–F). When the gate was set at IFN-γ+ cells, the majority of IFN-γ+ cells were CD8+ T cells (30±3%) and CD4+ T cells (25±1%).
To investigate the role played by IFN-γ during DENV infection, WT and IFN-γ-deficient (IFN-γ−/−) mice were inoculated DENV-2 and lethality rates and disease course evaluated. As seen in Figure 1C, 100% of IFN-γ−/− mice were dead before the seventh day of infection, and only 15% of WT mice had succumbed to infection. This early lethality of IFN-γ−/− mice was characterized by more severe manifestation of disease after DENV infection. Three days after infection, IFN-γ−/− mice already presented reduced platelets counts (Figure 1D), and at the 5th day of infection, there was marked thrombocytopenia (Figure 1D) and significant increase in hematocrit values (Figure 1E) in IFN-γ−/− mice when compared to WT mice. In addition to the alterations seen in hematological parameters, there was enhanced production of pro-inflammatory cytokines after infection. As shown in Figures 1F and 1G, there were no detectable levels of TNF-α and IL-6 in serum of WT mice at day 5 after DENV-2 infection. However, both cytokines were significantly elevated in serum of infected IFN-γ−/− mice (Figures 1F and 1G). Infected-IFN-γ−/− mice showed hepatic injury, as assessed by increased AST activity in plasma of IFN-γ−/− mice in the 5th day of infection (Figure 1H). There was also marked changes in liver architecture. WT mice inoculated with DENV-2 had little changes in liver, as assessed by histology. In contrast, there were signs of congestion and hepatocyte degeneration and necrosis in infected IFN-γ−/− mice (Figure 1I). In addition to the greater disease severity observed, IFN-γ−/− mice presented greater viral replication after infection than in WT mice. At the 3rd day of infection, IFN-γ−/− mice presented a 10 fold increase in DENV-2 viral loads in spleen and DENV-2 titers in spleen of infected-IFN-γ−/− mice were above 1.5 log greater than in infected-WT mice in the 5th day of infection (Figure 1J). Therefore, the data depicted here show IFN-γ is expressed and plays an important role in host defense against DENV infection.
Our next objective was to evaluate the roles of IL-12 and IL-18 in controlling IFN-γ production by the murine host during DENV infection. After DENV-2 infection, there were detectable levels of both IL-12p70 and IL-12p40 in the spleen of WT mice already in the 3rd day of infection (Figure 2A). The concentration of both cytokines was increased in the 5th and remained above background levels at the 7th day of infection (Figure 2A). This early production is consistent with a putative role of IL-12 in inducing IFN-γ production. Consistently with the latter possibility, there was a drastic reduction in IFN-γ production after DENV-2 infection of IL-12p40−/− mice, which are deficient for both IL-12 and IL-23 production (Figures 2B and 2C). In keeping with the relevance of IFN-γ during dengue infection and reduced IFN-γ production, there was enhanced lethality rates (Figure 2D), increased thrombocytopenia (Figure 2E) and enhanced hemoconcentration (Figure 2F) after DENV-2 infection of IL-12p40−/− mice. There were higher concentrations of TNF-α (Figure 2G) and IL-6 (Figure 2H) in spleen and more severe hepatic injury in IL-12p40−/− than WT mice after infection (Figure 2I and 2J). Finally, IL-12p40 deficiency resulted in greater loads of DENV-2 in spleen at the 7th day after infection, when compared with WT-infected mice (Figure 2K). The reduction of IFN-γ production and the more severe disease seen in IL-12p40−/− mice seem to be specifically due to IL-12 deficiency as IL-23p19−/−-deficient mice produced similar amounts of IFN-γ after DENV-2 infection (Supplementary Figure S2A) and presented a disease of similar intensity (Figure S2B and S2C) and unaltered viral loads (Figure S2 D) when compared to infected-WT mice.
Another cytokine shown to induce IFN-γ production during infections is IL-18 [21]. In the present study, IL-18 concentrations rose rapidly in liver at the 3rd day of DENV-2 infection, but returned to basal levels in the subsequent timepoints evaluated (Figure 3A). There was marked reduction of IFN-γ production in spleen and serum of DENV-2-infected IL-18−/− mice when compared with WT infected mice (Figure 3B and 3C, respectively). Available IL-18−/− mice were in the BALB/c background which we have previously shown to be more susceptible to DENV2-induced disease and lethality [24]. Indeed, all WT mice in the BALB/c background were dead by day 10 of DENV-2 infection using an inoculum that caused little lethality in C57Bl/6 mice (compare Figures 3D and 1C). All IL-18−/− mice also succumbed to infection but mice died earlier than WT controls after DENV-2 infection (p = 0.0237) (Figure 3D). Although the degree of thrombocytopenia was similar in both strains of mice (Figure 3E), hemoconcentration was greater in IL-18−/− than WT infected mice (Figure 3F). Levels of TNF-α (Figure 3G) and IL-6 (Figure 3H) and severity of liver injury (Figure 3I and 3J) occurred to a greater extent in spleens of IL-18−/− than WT infected mice (Figure 3G and 3H). Significantly, enhanced clinical disease and earlier deaths were accompanied by elevation in viral loads in spleen of IL-18−/− mice (Figures 3K).
The phenotype of either IL-12−/− or IL-18−/− mice were not as severe as the phenotype of IFN-γ−/− mice. For example, whereas viral loads were already approximately 2 log greater at day 5 in IFN-γ−/− mice, this was not the case in IL-12−/− or IL-18−/− mice (Figures 2J and 3J). Indeed, IFN-γ production was not abolished in IL-12−/− or IL-18−/− mice and viral loads were only significantly different from WT at day 7 after infection (see Figures 2J and 3J). In order to block simultaneously the action of both IL-12 and IL-18, IL-12p40−/− mice were treated with IL-18 bp at doses shown to block IL-18 action [28]. Treatment of IL-12p40−/− mice with IL-18 bp also resulted in total abrogation of IFN-γ levels in serum (Figure 4A) or spleen (Figure 4B) of infected mice. Treatment of IL-12p40−/− with IL-18 bp also resulted in marked enhancement of viremia already at day 5 after infection (Figure 4C), results which are similar to those obtained in IFN-γ−/− mice (Figure 1I) and substantially different from results observed at day 5 in IL-12p40−/− mice or mice treated with IL-18 bp alone (Figure 4C). Moreover, treatment of IL-12p40−/− with IL-18 bp resulted in thrombocytopenia, which was similar to that observed in IL-12p40−/− or IL-18 bp-treated mice (Figure 4D), and hemoconcentration, which was greater than in the other groups (Figure 4E). Levels of IL-6 in plasma were also further enhanced by the treatment of IL-12p40−/− mice with IL-18 bp than in either condition alone (Figure 4F). The enhanced viral load and greater disease severity already at day 5 resulted in greater lethality rates in IL-12p40−/− mice treated with IL-18 bp than in either condition alone or WT mice (Lethality rate at day 7: WT mice, 0%; IL-18bp-treated mice, 0%; IL-12p40−/− mice, 33%; IL-12p40−/− mice+IL-18 bp, 83%, n = 6). In concert, the data presented above suggest that IL-12 and IL-18 act together to induce optimal IFN-γ production during dengue infection in mice.
Nitric Oxide production by phagocytes is a well known effector mechanism induced by IFN-γ during host response to infections [22]. To assess whether this pathway is relevant in host response to DENV infection, we evaluated NOS2 expression after DENV-2 infection. As shown in Figure 5A, there was increase in NOS2 mRNA expression in spleen already at day 5 day but expression rose rapidly at day 7 after DENV2 infection of WT mice (Figure 5A). Evaluation of NOS2 staining in the liver by immunohistochemistry showed significant NOS2 expression, virtually only in infiltrating leukocytes, at day 7 after infection (Figure 5B, C). Consistently with the ability of IFN-γ to induce NOS2, there was no production of NO by dendritic cells infected with DENV-2, in vitro (Figures 5D). However, treatment of dendritic cells with IFN-γ prior to infection resulted in production of significant amounts of NO (Figure 5D). In addition, expression of NOS2 was greatly decreased in spleen of IFN-γ−/− mice after DENV-2 infection (Figure 5E). As IL-12 and IL-18 cooperate for optimal induction of IFN-γ (results above), we evaluated whether treatment of IL-12p40−/− mice with IL-18 bp would also results in reduced NOS2 expression in spleen. As seen in Figure 5E, concomitant absence of both IL-12 and IL-18 led to impaired NOS2 expression in spleen that was quantitatively similar to results obtained in IFN-γ−/− mice (Figure 5E).
To assess the role played by NOS2-induced NO during DENV infection, NOS2−/− mice were inoculated with DENV-2 and lethality rates and hematological alterations monitored. As shown in Figure 6A, NOS2−/− mice were markedly susceptible to DENV infection, as all knockout animals but none of WT mice were dead by the 10th day of infection. Thrombocytopenia (Figures 6B) was more intense earlier but hemoconcentration was similar in both groups (Figure 6C). There was enhanced splenic production of TNF-α (Figure 6D) and IL-6 (Figure 6E) and greater hepatic injury (Figure 6F and 6G) after DENV-2 infection of NOS2−/− than WT mice. Importantly, viral loads in spleen after DENV-2 infection were significantly greater in NOS2−/− than WT mice (Figures 6H). Of note, all alterations seen in NOS2−/−-infected mice were not due to reduction in IFN-γ production after infection. Indeed, IFN-γ levels in spleen and serum were similar in WT and NOS2−/− infected mice (Figures 6I and 6J). Therefore, NOS2-derived NO production is driven by IFN-γ and is essential for host protection during DENV primary infection.
The major findings of the present study can be summarized as follows: 1) IFN-γ production is essential for host resistance to DENV infection. NK and NKT cells are the sources of IFN-γ during the early periods of infection and are followed by CD4+ and CD8+ T cells, which are the main producers at the peak of host response to infection; 2) production of IL-12 and IL-18 precedes IFN-γ and optimal IFN-γ production relies on the combined action of IL-12 and IL-18; and 3) IFN-γ is essential for NOS2 induction and NOS2 plays an important role in controlling virus replication. These studies, therefore, indicate that IL-12/IL-18-induced IFN-γ production and consequent induction of NOS2 are essential for murine host response to DENV infection.
Previous studies support a protective role played by IFN-γ during host response to DENV infection. For example, Shresta and coworkers have shown that IFN-γ receptor-deficient mice were more susceptible to DENV-induced lethality than WT-infected mice, despite no differences in viral loads in several target organs between both groups [17]. The increased susceptibility was enhanced further when type I IFN receptor was also absent, and deficiency in both cytokine receptors resulted in disseminated viral replication [17]. In this respect, IFN receptors-deficient mice (AG129 strain) are known to be permissive for replication of DENV clinical isolates in peripheral tissues and CNS, and represent a well established experimental model of DENV infection [17]–[19]. In the present work, we have demonstrated that IFN-γ is produced as early as the fifth day of infection in WT mice and lack of IFN-γ action culminated in early lethality to a sublethal inoculum. These data establish IFN-γ as essential for host control of DENV replication and resistance to infection. The correlation between increased IFN-γ production and higher survival rates in DHF patients [15] also supports this idea.
Of note, enhanced viral replication in IFN-γ-deficient mice was associated with more severe disease manifestation, as showed by enhanced hematological alterations and hepatic injury. More severe disease was also noticed in DENV-infected AG129 mice, characterized by paralysis and elevated hematocrit [17]. Importantly, Gunther and colleagues have demonstrated in a human challenge model of DENV infection that only sustained IFN-γ production was associated with protection against fever and viremia during the acute phase of illness [31]. These data suggest that IFN-γ is important to prevent worsening of disease. In humans, epidemiological studies have shown that a substantial number of patients with severe disease have evidence of a previous infection with a distinct serotype [1]–[3], [9]–[11], [32]. Several hypotheses have been raised to explain this immune-mediated enhancement of disease severity. For example, it has been hypothesized that subneutralizing levels of antibodies facilitate the entry of viral particles in permissive cells (a phenomenon termed antibody-dependent enhancement - ADE), enhancing viral load, and exacerbating disease manifestation [33]. Experimental DENV models support this hypothesis and suggest that disease severity is directly associated with enhanced viral replication during infection [34], [35]. Of note, infected IFN-γ-deficient mice, as well as IL-12p40−/− and IL-18−/− infected mice, presented elevated viral loads, in parallel with elevated hematocrits, thrombocytopenia, and liver injury. Therefore, we may suggest that the worse outcome seen in mice with reduced IFN-γ production after infection is due to inability in control of DENV replication, leading to viral burden and enhancement of disease.
Mice in which IFN-γ production was decreased or deficient had a significant increase in levels of pro-inflammatory mediators after DENV infection. Indeed, both TNF-α and IL-6 production were enhanced in DENV-2 infected IFN-γ−/−, IL-12p40−/−, and IL-18−/− mice, when compared with WT controls. Increased levels of these cytokines have been associated with severity of dengue manifestation in humans [36]–[38]. Hence, enhanced TNF-α release by T cells during secondary stimulation with DENV antigens was found in hospitalized patients with more severe disease evolution [39]. In addition, the ratio of TNF-α-producing to IFN-γ-producing T cells among peripheral blood mononuclear cells from dengue-vaccine recipients was shown to be greater after in vitro stimulation with antigen from heterologous dengue serotypes [39], suggesting that increased amounts of TNF-α alters response to infection and may result in more-severe disease manifestation. Findings in murine experimental models support this idea [40]. Altogether, these findings in humans suggest that IFN-γ production is associated with protective responses to DENV infection and that severe disease may occur due to absence of proper IFN-γ release and to enhanced TNF-α production during response, although it remains to be shown if enhanced TNF-α production seen in DENV infected IFN-γ−/− mice was due to T cells or to any other cellular population.
Interestingly, enhanced viral load have also been associated with increased pro-inflammatory response during mouse experimental infection by West Nile virus [41], another important flavivirus that is pathogenic to humans. The latter findings support the hypothesis that increased virus replication in the absence of IFN-γ production leads to increased pro-inflammatory mediators response. TNF-α blockade in experimental models of DENV infection resulted in prevention of disease [19], [23] and TNF-α action has been implicated in increased vascular permeability after infection in experimental settings [13]. Of note, inhibition of other pro-inflammatory mediators produced in the evaluated experimental model of DENV infection, including PAF and MIF, is associated with reduced production of TNF-α and IL-6 and milder disease manifestation, reduced hypotension and vascular permeability after DENV infection [13], [24], [25]. Hepatic injury was also enhanced in IFN-γ−/− mice infected with DENV. Data from our laboratory suggest that enhanced liver injury during experimental DENV infection involves both productive viral infection of hepatocytes and immunopathological mechanisms, such as enhanced leukocyte arrest and activation in hepatic tissue (our unpublished data, manuscript in preparation). Therefore, the elevation of pro-inflammatory cytokine production and consequent liver injury seen in the absence of IFN-γ appears to account for the worse outcome after DENV infection in mice.
Several studies have demonstrated the IFN-γ-inductive role played by IL-12 and IL-18 during experimental models of viral infections [20], [21], [42]. Here, we have shown that IL-12 and IL-18 were produced early after DENV infection. The kinetics of production of these cytokines was compatible with their inductive role of IFN-γ production. In support of the latter possibility, IL-12p40−/− and IL-18−/− mice presented marked reduction in IFN-γ production after DENV infection. In addition, absence of one of these cytokines led to worsening of dengue disease, despite a small delay in peak of DENV-induced alterations. Of note, only during simultaneous blockade of both IL-12 and IL-18, there was complete abrogation of IFN-γ production. Interestingly, IL-12−/− mice treated with IL-18 bp presented marked enhancement of splenic viral loads already at the 5th day post DENV-2 infection and disease seen in these mice was very similar to that found in infected IFN-γ−/− mice. Thus, IL-12 and IL-18 act synergistically to induce IFN-γ production during DENV infection. Of note, IL-18 production has been shown to be dependent on inflammasome complex activation [43], suggesting that this molecular scaffold may play a role in the control of IFN-γ production and in host resistance to DENV infection.
IL-18 is known to augment IL-12-induced IFN-γ production by T and NK cells [20], [21], [42], [44], and absence of IFN-γ in infected mice is known to abolish both NK cell and CTL responses during viral infections [42], [44]. Our data suggest that, upon infection, NK and NKT cells are the cell populations involved in early IFN-γ production and that CD8+ and CD4+ T cells are the main IFN-γ producers at later moments of response to infection (7th day). IFN-γ production by CD4+ T cells during experimental DENV infection has been previously demonstrated [45]. In addition, CD8 T cell activation has been associated to protection to DENV primary infection in mice [46], [47]. Our data showing a significant increase in IFN-γ+ NK and NKT cells and the finding that IFN-γ−/− mice succumb very early to infection suggest a important role for these cell populations in mediating resistance to DENV infection during its initial phases. Of note, NK cell activation early after experimental DENV infection has been previously demonstrated [44]. Interestingly, increased percentages of NK cells and of activated NK cells were also associated with milder DF, whereas reduced cell counts, low percentages and lack of activation markers (comparable to healthy controls) were associated with evolution to DHF in patients [48], [49]. Altogether, these observations suggest that sequential and coordinated IFN-γ production by these lymphocytes populations during DENV infection is an event of extreme importance for host resistance to disease.
However, it remains to be shown the antigenic specificity of these IFN-γ-producing lymphocytes in the studied experimental settings. In addition, whether these cells are poly-functional and secrete other cytokines or present other effector functions remain to be studied. In this regard, it has been demonstrated that development of subclinical secondary infection in school children is associated with increased proportions of DENV-specific TNF-α, IFN-γ and IL-2-producing CD4+ and CD8+ T cells [50], suggesting that poly-functional responses correlate with protection to severe disease manifestation. On the contrary, cytokine-producing T cells (especially TNF-α and/or IFN-γ) were associated with DHF development in patients and these DHF associated, cytokine-producing T cells were shown to be negative for CD107a staining, suggesting that these lymphocyte populations represent mono-functional or oligo-functional T cells [51]. Therefore, assessment of the pattern of T cell cytokine production and of the mechanisms controlling such polyfunctionality (whether IL-12 and or IL-18 are involved in such control) may provide important information regarding protective versus pathogenic responses to DENV infection and may bear relevance during development of vaccinal strategies. At the moment, these subjects have been matter of ongoing analysis in our experimental infection model.
Apart from promotion of NK and CTL responses, IFN-γ seems to be important for viral clearance by induction of NO production. It has been shown that NOS2 expression is increased upon DENV infection in humans and that this expression in peripheral blood monocytes of DF patients was found to correlate with the late acute phase of disease and preceded the clearance of DENV from monocytes [52]. Hence, NO production was associated with less severe form of dengue disease in humans [53]. Here, we demonstrate that NOS2 expression is increased during DENV infection and that this expression is controlled by IFN-γ production, once IFN-γ−/− and IL-12p40−/− mice treated with IL-18 bp presented reduced NOS2 expression. In addition, IFN-γ stimulation was necessary for NO production by DENV-infected DCs, in vitro. Importantly, blockade of NOS2 action was associated with enhanced viral loads after infection, and more severe disease manifestation, even in the presence of high levels of IFN-γ. Of note, NO is able to inhibit DENV replication in human cells in vitro [54], [55], an effect associated with inhibition of DENV associated polymerase activity [54]–[56]. Thus, NOS2-mediated NO production is pivotal for resistance to DENV infection and this seems to be a major pathway involved in IFN-γ-mediated resistance to disease. However, in the absence of NOS2, animals die with a slower kinetics than IFN-γ−/− mice, suggesting that mechanisms in addition to NOS2-mediated NO production may be relevant for IFN-γ-mediated host protection to infection. This could involve the presence of CTL responses and NK cells, but not NKT cells, which seem to play detrimental role in experimental DENV infection [57]. These IFN-γ-dependent and NOS2-independent mechanisms are currently being investigated in our laboratory.
However, other studies have demonstrated a pathogenic role for NO during DENV infection. Utilizing human cell lines and experimental mouse infection, it has been shown that overproduction of NO could lead to endothelial cell damage, and cross-reactive antibodies against endothelial cells, present during DENV infection, were found to induce cell damage in an NO-dependent manner [58]. For example, Yen and coworkers have found that tissue hemorrhage after experimental DENV infection was dependent upon reactive nitrogen species production by endothelial cells. This event was associated with increased endothelial cell apoptosis during infection [59]. Although NOS2 inhibition resulted in reduced hemorrhage, viral replication was not evaluated. In addition, the increased hemorrhage displayed after NO production seemed to be an endothelial cell-associated phenomenon and was potentiated by TNF-α and reactive oxygen species (ROS). On the contrary, IFN-γ-mediated NO inhibition of viral replication was demonstrated especially in leukocytes population both in human and mouse settings [52]–[56]. Our results showed that NOS2 staining during DENV-2 infection in the present model was mainly associated to leukocytes. These findings suggest that NO may have a dual role during DENV infection and that this is associated with the cell populations involved in NO production and on the presence of additional inflammatory mediators. NO production by infected leukocytes may be associated to control of viral replication and prevention of disease evolution, while NO production by endothelial cells, especially in the presence of TNF-α and ROS, would favor cell death and more severe disease manifestation. Additional experiments evaluating cell-specific NOS2-deficient mice will help in answering the latter hypothesis and aid in defining other roles of NO in the context of experimental dengue.
In conclusion, we have demonstrated that IFN-γ production is essential for host resistance to DENV infection. IFN-γ production upon infection is controlled by concomitant production of IL-12 and IL-18 and the IFN-γ-dependent mechanisms associated to resistance to dengue disease involve NOS2 up-regulation and consequent NO production. In the absence of these molecules, there is enhancement of viral burden and more severe manifestation of dengue disease. Thus, IFN-γ induction helps to orchestrate immune response maturation, control of viral replication and regulation of inflammatory response during host response to DENV infection, defining the outcome of dengue disease. Despite extrapolation of this experimental scenario to human infection requires further investigation, we may suggest that strategies that improve the production of IFN-γ-mediated immunity by the host could be useful during the control of primary infection by Dengue virus.
|
10.1371/journal.pcbi.0040008 | Noise Propagation and Signaling Sensitivity in Biological Networks: A Role for Positive Feedback | Interactions between genes and proteins are crucial for efficient processing of internal or external signals, but this connectivity also amplifies stochastic fluctuations by propagating noise between components. Linear (unbranched) cascades were shown to exhibit an interplay between the sensitivity to changes in input signals and the ability to buffer noise. We searched for biological circuits that can maintain signaling sensitivity while minimizing noise propagation, focusing on cases where the noise is characterized by rapid fluctuations. Negative feedback can buffer this type of noise, but this buffering comes at the expense of an even greater reduction in signaling sensitivity. By systematically analyzing three-component circuits, we identify positive feedback as a central motif allowing for the buffering of propagated noise while maintaining sensitivity to long-term changes in input signals. We show analytically that noise reduction in the presence of positive feedback results from improved averaging of rapid fluctuations over time, and discuss in detail a particular implementation in the control of nutrient homeostasis in yeast. As the design of biological networks optimizes for multiple constraints, positive feedback can be used to improve sensitivity without a compromise in the ability to buffer propagated noise.
| Biological circuits need to be sensitive to changes in environmental signals but at the same time buffer rapid fluctuations (noise) that might be imposed on this input. In this paper, we analyze the interplay between sensitivity to signals and the ability to buffer noise. Previous studies reported that negative feedback attenuates noise. We show, however, that this ability comes at the expense of an even more dramatic reduction in sensitivity. In fact, when comparing systems of the same sensitivity, a system with negative feedback is more amenable to noise than a system without such feedback. We searched for small biological circuits that can buffer noise while maintaining high sensitivity, and found that positive feedback exhibits this property. This ability of positive feedback to buffer noise reflects its slowed-down dynamics. We discuss general requirements for the function of positive feedback as a noise-filtering device and describe a particular implementation that appears to function in yeast nutrient homeostasis. Our study emphasizes the need to consider multiple constraints when analyzing the design logic of biological networks.
| Cells sense and process information using biochemical networks of interacting genes and proteins. Typically, a signal is sensed at a specific point of the network (input) and is propagated to modulate the activity or abundance of other network components (output). Reliable information processing requires high sensitivity to changes in the input signal but low sensitivity to random fluctuations in the transmitted signal. Since the detection of signal is inherently stochastic [1], and the microenvirnment of the cell is also fluctuating randomly, understanding the principles of noise propagation in biochemical and genetic networks is of interest [2–5].
Linear (unbranched) cascades present the simplest instance of biochemical networks. Recent studies have shown that such cascades display an interplay between sensitivity to changes in input signal and the ability to buffer stochastic fluctuations [6–9]. Indeed, an increase in the sensitivity toward input signals results also in elevated sensitivity to noise in the input. A key question is whether network connectivity, e.g., the presence of positive or negative feedbacks, can modulate this interplay, reducing propagated noise while maintaining high sensitivity. Previous studies argued that negative feedbacks can buffer noise relative to linear cascades [10–12]. These studies, however, did not consider the associated changes in signaling sensitivity.
In general, the fine line that separates “noise” from “signal” is established functionally. Nevertheless, in many systems such as the sensing of temperature, nutrient levels, ligand concentration, etc., the signal is interpreted as a long-term change in the input, whereas noise is characterized by rapid stochastic fluctuations. In this study, we focus on this particular class of systems.
We explore for gene circuits that can buffer propagated noise while maintaining signaling sensitivity. We consider a large set of networks that are differentially designed but are equally sensitive to long-term changes in the input, and compare their ability to buffer propagated noise. Systematic analysis of all three-gene circuits revealed that negative feedback amplifies propagated noise. In contrast, positive feedback appears to be a necessary element for buffering such noise. Analytical analysis demonstrated that positive feedback contributes to noise buffering by slowing down the dynamics, thus providing a longer averaging time. A detailed analysis of a recurrent network design, found in systems controlling nutrient homeostasis, suggests that it functions as a noise-reduction device based on the principles identified in our analysis.
To begin analyzing the effect of network architecture on the interplay between sensitivity and noise buffering, we considered a two-component cascade with a negative feedback loop. This cascade is composed of an input node, n0, which activates an output node, n1. The output node feeds back to repress its own expression (Figure 1A). Formally, this system is described by
where β1 denotes the maximal rate of n1 production,
denotes the rate of n1 degradation, and h0, h1 are Hill coefficients. Note that n0 and n1 are normalized by their respective dissociation constants from the gene promoter.
We consider an input signal n0(t) = 〈n0〉 + σ0(t) which fluctuates around some mean level 〈n0〉. The fluctuating component σ0(t) has a zero mean and some autocorrelation time τ0. Figure 1B depicts the temporal fluctuations of n1 for a system with a strong negative feedback (Hill coefficient h1 = 4). The analogous dynamics for a system that lacks such feedback (h1 = 0) is also shown. Consistent with previous studies [10–12], output noise is lower in the presence of negative feedback. Nevertheless, negative feedback also significantly lowers the sensitivity of the system to a two-fold change in the level of the mean input (Figure 1C).
To rigorously quantify the interplay between the sensitivity of the input–output relation and the buffering of propagated noise, we define two measures for the sensitivity and noise-amplification of the system. The steady-state sensitivity is captured by the susceptibility, s, [3,13,14] (also termed gain [9]) defined as the relative change in output following a change in the input:
with all quantities measured at steady state. The measure for noise amplification
is defined as the ratio between the output and input noise:
As before, all quantities are measured at steady state. Both s and
depend on the different parameters of the system, including the Hill coefficients and mean input levels.
Figure 1D depicts the noise amplification versus susceptibility for different levels of mean input. The case of no feedback (h1 = 0) is compared to that of increasing feedback cooperativity (h1 = 1 and 2). Again, a clear interplay between susceptibility and noise buffering is observed, with systems that are more sensitive to changes in the input level being also more vulnerable to noise. Notably, this interplay seems to be more severe in the presence of negative feedback. Thus, for a given level of susceptibility, propagated noise is amplified to a greater extent in the presence of a negative feedback. This result is consistent with the theoretical arguments for a two-node model [3]: with negligible intrinsic noise, and when controlling for sensitivity, negative feedback enhances, rather then represses, propagated noise.
Our analysis implies that negative feedback cannot be used to buffer against rapidly varying propagated noise in systems which require a sensitive response to long-term changes in their input. To identify network architectures that can buffer noise while maintaining sensitivity, we characterized systematically the relation between susceptibility and noise-buffering of all three-node circuits (Figure 2A). A three-node circuit is composed of an input node (n0), an output node (n2), and an intermediate node (n1), connected via activating or repressing interactions (arrows). We allowed for all incoming or outgoing arrows, with the exception of the input n0 which could affect both n1 and n2 (outgoing arrows), but was not subject to feedback regulation (no incoming arrows). Each arrow was assigned a positive sign (activation) or a negative sign (repression), thus leading to a total of 324 networks (Text S1, Section IV). To ensure controlled comparison between networks, we assume that degradation is not regulated and that all proteins degrade at the same rate (e.g., by dilution).
Each specific circuit supports a range of dynamic behaviors, depending on the precise value of the interaction parameters. Following the formalism presented by Paulsson [3,13], we used the Fluctuation Dissipation Theorem (FDT [15]) to derive an analytical formula for susceptibility and noise-amplification (Methods). Briefly, the system of equations that describe the dynamics of the three-node network was
The degradation terms were assumed to be first-order, and we considered τi = 1 to maintain a mathematically controlled comparison. This system of equations was linearized around the steady state. The linearization process excludes from the analysis possible noise-filtering mechanisms that show very sharp functions such as AND or OR gates [16] as well as oscillations or transitions between multiple steady states (such as the study on positive feedback loops and noise in [17]). Note also that the steady state n0 = n1 = n2 = 0 is not formally part of our analysis, because it renders the relative fluctuation η infinite.
Following linearization, the combined effect of all interaction parameters (i.e., Hill coefficients and saturation levels) is captured by the elasticity [3,13]:
where
is the rate of generating ni (e.g., via transcription) and
is its degradation rate. With these definitions, the susceptibility of the output was (Text S1, Section I)
The absolute value facilitates a comparison between systems that increase or decrease their output when the input goes up.
Noise amplification was found by solving the matrix equation [3,13]
where the matrix η is composed of the normalized noise terms, the matrix M is related to the elasticities and time scales τi, and the matrix D contains a single term corresponding to noise input from n0. In the construction of D, we assume that the sole noise source is fluctuations in n0, and that these fluctuations die out exponentially with a time scale τ0 = 1 (autocorrelation time of one unit). The exact terms of Equation 7 are defined in Section II of Text S1 and in [3,13].
A particular choice of elasticity values for all arrows of the network defines a single point in the
plane, and the general interplay between sensitivity and noise buffering was derived by considering a large number of different elasticity values (Figure 2A, Methods).
As expected, noise amplification in linear (unbranched) cascades is precisely proportional to the susceptibility (Figure 2B). This case of no feedback provides the reference for comparison for other network architectures. Consistent with the analysis above, in the case of a negative feedback, all the points appear above the reference line (Figure 2C) implying an increase in noise for a given level of susceptibility. Noise amplification at constant susceptibilities is observed also for the coherent (Figure 2D) and incoherent (Figure 2E) feed-forward loops (FFLs) [16,18], probably reflecting the addition of nonsynchronous noise components mediated through the intermediate node. In contrast, for positive feedback the points in the
plot appear below the reference line (Figure 2F). Thus, for a given level of sensitivity, positive feedback buffers propagated noise.
To further characterize the properties of all three-node circuits, we calculated for each network the fraction of parameter sets that produce a stable steady state (Text S1, Section III), following the paradigm of [19]. We then calculated the fraction of stable parameter sets (Methods, Section IV of Text S1) that display high susceptibility and low noise (Figure 2G). Notably, networks that were stable throughout the entire parameter range provided poor noise buffering for a given susceptibility. None of these circuits contained positive feedback loops. In sharp contrast, the circuits that enhanced noise buffering were amenable to instability and were all composed of positive feedback loops. Taken together, among the networks tested, positive feedback appears to be required for buffering propagated noise while maintaining sensitivity.
To better understand the reason underlying the ability of positive feedback to buffer propagated noise for a given susceptibility, we used the analytical description of a two-component system with an input n0 and an output n1, as was derived by Paulsson in [3,13] using the FDT approach [15]. In this framework (and while neglecting intrinsic noise), noise amplification was shown to be given by [3,13]:
where s = −H10/H11 denotes the susceptibility; τ0, τ1 denote the degradation time scales of n0 and n1, respectively; and H10, H11 are the elasticities (as defined in Equation 5 above). In the absence of feedback, H11 = 1. Negative feedback of n1 on itself implies H11 > 1, whereas positive feedback implies 0 < H11 < 1 (if H11 falls below zero, instability arises). As was shown by Paulsson [3], negative feedback impairs noise buffering at a given susceptibility by effectively accelerating the dynamics and reducing averaging of fluctuations over time. Similarly, positive feedback enhances noise buffering for constant susceptibility by slowing down the dynamics and allowing for better time-averaging of fluctuating components.
Notably, Equation 8 suggests that positive feedback is not the only way to reduce H11 [3,13]. Moving beyond the first-order degradation assumed in our study, H11 can also be decreased if the degradation is independent of n1 (
in Equation 5). This, in fact, is likely to be the case for nondividing microorganisms. Hence, time averaging would also be improved if the degradation is close to zero-order and the synthesis is not influenced by n1.
While positive feedback appears to be important for buffering propagated noise (when sensitivity is controlled for), such a mechanism needs to comply with several requirements. First, the feedback loop itself should produce low internal noise because intrinsic noise is not buffered. Second, the effective elasticity H11 (Equation 8) should be of intermediate magnitude: when it is too high (H11 → 1) the effect of the positive feedback is negligible, but when it is too low (H11 → 0) the system is on the verge of instability, and the steady state will no longer resist small fluctuations. Finally, to avoid decrease in susceptibility due to saturation effects, H11 must be maintained constant over a large range of parameters.
A class of mechanisms that complies with the above requirements is based on a combination of positive and negative feedbacks. Fast-acting negative feedback functions to ensure stability, while positive feedback provides the required noise buffering. A specific example for such a network is involved in nitrogen homeostasis in yeast [20–22] (Figure 3A). Here, a transcription factor (Gat1p), which is activated by nuclear Gln3p, feeds back to enhance its own transcription, and in addition induces a transcriptional repressor (DAL80) that competes with Gat1p for the same DNA binding sites. This competition effectively weakens the positive feedback and ensures stability. Denoting the input signal to the system by n0, the output Gat1p by n1 and the repressor Dal80p by n2, the system can be modeled by the following two differential equations:
Here αi and βi denote the degradation and transcription rate constants, respectively, and l is a low rate of basal transcription required to prevent the shutdown of the system, n1 = n2 = 0. We will neglect this factor in subsequent analysis. The Kij coefficients in the protein production terms are dissociation constants, with n2/Ki2 describing the competitive inhibition of Dal80p. The Hill coefficient of n2 binding to its own promoter is 2 because Dal80p binds as a dimer [21,22]. The Hill coefficient of n2 binding to the n1 promoter is set to 1 to enhance noise buffering and susceptibility (although a value of 2 would still increase noise averaging).
For the system described by Equations 9 and 10 to operate as a sensitive noise buffer, it must work in a regime where all interactions are unsaturated. Hence, all the binding constants of the repressor, Ki2, must be small, and all binding constants of the activator, Ki1, must be large. In this regime, Equation 9 and 10 reduce to
and
Finally, if n2 responds more rapidly than n1 and n0 (H22α2 ≫ H11α1, 1/τ0), then it can be assumed to be at quasi-steady state, and Equations 11 and 12 are combined to
The power law dependence of the transcription rate on n1 results in an almost-constant elasticity
= 1/3 (
+ in Equation 5). Hence, this network can buffer noise and maintain susceptibility for a large range of concentrations at which it remains unsaturated. A more rigorous analysis of the system is presented in Section VI of Text S1.
Detailed simulations confirm that this system can indeed buffer propagated noise, as compared to a loop-free system with the same levels of susceptibility (Figure 3B and 3C). Furthermore, the noise buffering capacity and the susceptibility of this system are maintained over a large range of input levels (Text S1, Section VI).
The ability to distinguish input signals from stochastic fluctuations is crucial for reliable information processing. Yet, being processed by the same computation device, signal and noise are inherently coupled. It thus comes as no surprise that increasing the ability to buffer propagated noise comes typically at the expense of reducing the sensitivity toward the input signal. We study this interplay in the context of a special class of systems where the signal is retained for long time periods, whereas the noise fluctuates rapidly. Such systems are ubiquitous in the adjustment of cells to aspects of their extracellular environment.
Previous studies reported that negative feedback buffers gene expression noise [10–12]. Nonetheless, when considering propagated noise that originates upstream of the feedback loop, this noise filtering merely reflects the reduction in the ability of the system to respond to changes in its input. Moreover, when parameters are chosen to preserve system sensitivity, negative feedback in fact amplifies, rather than reduces, propagated noise. By the same token, positive feedback, which appears to both increase the sensitivity of the system to changes in its input and to amplify intrinsic noise, reduces propagated noise when susceptibility (steady-state sensitivity) is controlled for.
Analytical analysis [3,13] revealed that noise propagation depends on two factors: the sensitivity to changes in input (susceptibility) on the one hand, and the averaging time [3,13] on the other hand. In the absence of feedback, this averaging time depends only on degradation rate. However, both negative and positive feedbacks impact this averaging time: negative feedback accelerates the dynamics [23] and consequently it reduces time averaging and does not buffer against noise. In contrast, positive feedback delays the kinetics leading to attenuation of propagated noise. If we view the feedback modules as low-pass frequency filters [12,24,25] and define a critical frequency [25] above which fluctuations are eliminated, then negative feedback increases this critical frequency, allowing more propagated noise to pass, whereas positive feedback decreases this frequency, thus reducing the amount of noise.
Whereas our study illustrates the effect of positive feedbacks, additional mechanisms could be used for reducing propagated noise by similarly increasing the averaging time. Such mechanisms include long linear cascades; cascades with an intermediate component that has a relatively large half-life [3,26]; or scenarios where both synthesis and degradation are essentially zero-order (from the definition of H11 in Equation 5). Finally, we note that systems that exhibit time delays together with bistability were not included in our screen but could also attenuate noise [17].
Positive feedbacks did not emerge as a recurrent network motif in several of the transcriptional networks analyzed [19]. One possibility is that designing the proper feedback that will maintain stability while providing noise buffering is evolutionarily difficult for the small size networks considered in these studies [19], due to the requirement it imposes on the extent of nonlinearities (Hill-coefficients) of the interactions. A simple realization of this concept, however, can easily be implemented by somewhat larger networks, as exemplified by the coupled positive–negative feedback we described. This and similar implementations function over a broad range of parameters and do not require strict tuning. Further analysis will be required to assess the abundance of this positive feedback–based noise-reduction scheme in different biological systems.
All simulations were based on the Gibson-Bruck [27] modification of Gillespie [28] algorithm. Input noise was implemented via transcription from a low copy mRNA with a short half life. No other mRNAs were explicitly considered. Simulation parameters are detailed in Section VII of Text S1. Simulations were carried out using Dizzy [29].
The interaction parameters of each arrow in each network are captured by the interaction elasticities Hij. The susceptibility (Equation 2) for each network is calculated from the elasticities using Equation 6 (for a general derivation of the susceptibility, see Text S1, Section I). The noise amplification is connected to the elasticities through the solution of Equation 7. Definitions for the terms in Equation 7 appear in Section II of Text S1. Equation 7 was solved symbolically for all three node networks using Maple (MapleSoft, Waterloo Maple). Solutions for specific network architectures are shown in Table S1.
The values for the elasticities were randomly assigned to each network. To control for similar distribution of positive and negative interactions, we defined the synthesis elasticity Sij by Hii = 1 − Sii and Hij = −Sij. When i enhances the synthesis of j, then Sij > 0, and vice versa. Positive feedback of i on itself implies Sii > 0 and vice-versa. The synthesis elasticity values were sampled from a uniform distribution between zero and four and assigned to the arrows in each network. Different sampling ranges did not have a significant effect on the conclusions (Text S1, Section IV). We sampled 20,000 random sets of parameters for each circuit. The time constants were held fixed at a value of one, but different values did not change the results (Text S1, Section IV).
Stability criteria were established via the sign of the eigenvalues of the interaction matrix (Text S1, Section III). |
10.1371/journal.pbio.1001219 | Hyperactive Neuroendocrine Secretion Causes Size, Feeding, and Metabolic Defects of C. elegans Bardet-Biedl Syndrome Mutants | Bardet-Biedl syndrome, BBS, is a rare autosomal recessive disorder with clinical presentations including polydactyly, retinopathy, hyperphagia, obesity, short stature, cognitive impairment, and developmental delays. Disruptions of BBS proteins in a variety of organisms impair cilia formation and function and the multi-organ defects of BBS have been attributed to deficiencies in various cilia-associated signaling pathways. In C. elegans, bbs genes are expressed exclusively in the sixty ciliated sensory neurons of these animals and bbs mutants exhibit sensory defects as well as body size, feeding, and metabolic abnormalities. Here we show that in contrast to many other cilia-defective mutants, C. elegans bbs mutants exhibit increased release of dense-core vesicles and organism-wide phenotypes associated with enhanced activities of insulin, neuropeptide, and biogenic amine signaling pathways. We show that the altered body size, feeding, and metabolic abnormalities of bbs mutants can be corrected to wild-type levels by abrogating the enhanced secretion of dense-core vesicles without concomitant correction of ciliary defects. These findings expand the role of BBS proteins to the regulation of dense-core-vesicle exocytosis and suggest that some features of Bardet-Biedl Syndrome may be caused by excessive neuroendocrine secretion.
| Bardet-Biedl syndrome, BBS, is a rare human genetic disease caused by mutations in many genes. The BBS phenotype is very complex; it is principally characterized by early-onset obesity, progressive blindness, extra digits on the hands and feet, and renal problems. BBS patients may also suffer from developmental delay, learning disabilities, diabetes, and loss of the sense of smell. This complexity suggests that BBS proteins function in a variety of tissues, causing defects in many organs. A unifying theme for the diverse features of BBS emerged when BBS genes were identified and their protein products were found to function in the cilium, a sensory structure found in many cell types. Since then, the various manifestations of BBS have been attributed to the loss of ciliary function in the corresponding tissues. This notion was also supported by the finding that mutations in several genes required for proper cilia formation and function reproduce some of the features seen in BBS patients. Here, we have further investigated the defects found in Caenorhabditis elegans strains carrying mutations in BBS genes (bbs mutants). We find that not only do they display sensory deficits associated with loss of ciliary function, but they also exhibit increased release of multiple peptide and biogenic amine hormones contained in dense-core vesicles of ciliated sensory neurons. Importantly, limiting this excessive hormonal release without correcting the ciliary defects of bbs mutants was sufficient to restore normal body size, feeding, and metabolism to these mutants. Moreover, we show that although non-bbs ciliary mutations can mimic some of the phenotypes of bbs mutants, these effects can be attributed to distinct spatial and molecular mechanisms. Our findings indicate that C. elegans bbs mutants exhibit features of both ciliary and endocrine defects and suggest that some of the clinical manifestations of human BBS may result from excessive endocrine activity, independently of the loss of ciliary function.
| Bardet-Biedl Syndrome is a rare, multigenic, pleiotropic disorder characterized by defects in many tissues including the eyes, kidneys, central nervous system, and reproductive organs [1]. The clinical presentations of BBS include broadly prevalent disease conditions such as obesity and retinopathy. Mutations in at least 14 genes cause BBS [1]–[3]. Of these, seven interact to form a complex termed the BBSome [4]. The mammalian BBSome associates with ciliary membranes and interacts with a guanine nucleotide exchange factor (GEF) for Rab8, a Rab GTPase implicated in cilia formation and function [4]. The BBSome is thought to act as a coat protein for cilia-bound vesicles, promoting the sorting and delivery of signaling molecules to the cilium [5]–[7]. Given that cilia are sensory organelles enriched in signaling molecules, the molecular functions of the BBSome have led to the current paradigm for understanding Bardet-Biedl Syndrome in which defects in cilia formation and signaling have been advanced as a unifying explanation for the wide range of phenotypes seen in BBS patients [3],[8]–[11]. In support of this notion, conditional knockouts in components of intraflagellar transport, IFT, a multi-protein machinery required for building the cilium, have been shown to result in some of the phenotypes present in bbs-deficient mice such as obesity and kidney disease [12]–[14].
In C. elegans, the seven BBSome components are conserved and are exclusively expressed in the 60 sensory neurons of this animal [8]. Various subsets of these neurons have been implicated in sensation of environmental cues that affect animal growth, metabolism, and feeding behavior [15]–[17]. Disruptions of the C. elegans BBSome cause altered rates of IFT, defects in the structural integrity of sensory cilia, and diminished behavioral responses to various sensory cues [8],[9],[18],[19]. Thus, in C. elegans as in mammals, loss of BBSome components leads to phenotypes that are also seen in IFT mutants. Consequently, the physiological abnormalities of C. elegans bbs mutants have also been attributed to defects in cilia formation and signaling [9],[18].
Here we report that BBSome mutants display a dramatic increase in the secretion of dense-core vesicle cargoes from ciliated sensory neurons, while mutations in IFT components generally result in reduced secretion, indicating that the consequences of bbs deficiency cannot be fully explained by impaired intraflagellar transport. We show that the enhanced secretion of bbs-deficient animals depends on the evolutionarily conserved Rab27/rabphillin/CAPS exocytosis machinery but not on Rab8, which participates in vesicular transport to the cilium. Importantly, we show that the altered size, feeding, and metabolic phenotypes of bbs mutants can be normalized to wild-type levels by abrogating the enhanced secretion of these mutants without simultaneous correction of ciliary defects. We also demonstrate that while certain phenotypes of IFT mutants mimic those caused by bbs mutations, distinct neural and molecular mechanisms underlie these phenotypes. These findings expand the role of the BBSome to the regulation of dense-core vesicle secretion and raise the possibility that enhanced neuroendocrine secretions rather than ciliary defects per se may be primarily responsible for some of the clinical features of BBS patients.
To survive and thrive in a dynamic environment, C. elegans, as in other animals, must sense changes in environmental conditions and respond through behavioral, physiological, and metabolic adaptations. C. elegans accomplishes this in part by secreting a diverse array of neuroendocrine hormones including insulin-like peptides from their various ciliated sensory neurons [20]. We had previously shown that insulin secretion from the ADL pair of ciliated sensory neurons can be assessed by expressing a fluorescently tagged insulin, DAF-28-mCherry, exclusively in this pair of neurons, and measuring the accumulation of the secreted insulins in coelomocytes, scavenger cells that non-specifically endocytose pseudocoelomic fluid [21]. It is thought that peptidergic secretions from neurons such as ADL, which are not directly in contact with the pseudocoelom, gain access to this space through the extracellular matrix and the ancillary pseudocoelom, a fluid fill cavity that bathes many of the neurons in the head of the worm. Eventual passage of secreted molecules from the ancillary pseudocoelom to the pseudocoelom is thought to be mediated by tight junctions. Coelomocytes within the pseudocoelom non-specifically endocytose pseudocoelomic fluids, concentrating them in endocytic vesicles; thus, the steady state level of circulating signaling molecules secreted from ADL can be measured by quantifying the amount of fluorescent insulins within the coelomocytes [21]–[23].
To better understand the molecular mechanisms that link environmental cues to secretion of insulin-like peptides, we examined the effects of various defects in cilia formation and function on insulin secretion from the ADL pair of sensory neurons. We found that mutations in intraflagellar transport, IFT, components such as osm-3, osm-5, che-2, or che-11 [24], which result in defective cilia, cause a ∼50% reduction in insulin secretion (Figure 1A,B,E,F) consistent with the idea that appropriate cilia function is required for detection of food-related cues and subsequent coordination of metabolic and growth pathways through insulin signaling. We were therefore surprised to find that cilia-defective osm-12 mutants [18], the C. elegans homolog of bbs-7, exhibit a dramatic 2–3-fold increased insulin accumulation in coelomocytes (Figure 1A–D). Mutations in bbs-1, bbs-5, bbs-8, and bbs-9, encoding other components of the BBSome, also caused enhanced accumulation of the insulin reporter in coelomocytes (Figure 1C,D), suggesting that the entire BBSome complex regulates insulin release from ADL sensory neurons.
The increased secretion was not limited to ADL neurons, as bbs mutants also exhibited increased secretion when the fluorescent insulin reporter was expressed exclusively in the ASI pair of sensory neurons (Figure 1E,F). Additionally, BBSome mutants exhibited a ∼2–3-fold elevated secretion of DAF-7, a neurally expressed TGF-β ligand and FLP-21, an FRMF-amide neuropeptide expressed under the ADL specific, srh-220, promoter (Figure 1G,H), indicating that BBSome mutants have elevated release of various dense-core vesicle cargoes from multiple ciliated sensory neurons.
To assess whether the enhanced accumulations of the fluorescent reporters in coelomocytes indicate functional increases in the release of insulins and neuropeptides, we assayed phenotypes associated with hyperactive insulin and neuropeptide signaling. Insulin and TGF-β signaling pathways regulate whether C. elegans grow reproductively or enter a hibernating dauer stage [25]. Reductions in either of these parallel pathways initiate dauer entry, while favorable conditions promote dauer exit and reproductive growth. Accordingly, overexpression of either ins-4 or daf-28, encoding distinct insulins, partially suppresses the constitutive dauer formation phenotype of TGF-β mutants [22]. We found that mutations in either bbs-7 or bbs-9 partially suppressed the dauer phenotypes of daf-7 or daf-1 mutants, encoding the TGF-β ligand and its receptor, respectively, but not that of mutants in daf-2, the insulin receptor (Figure 2A and unpublished data). Consistent with the notion that the enhanced secretion of bbs mutants mediates dauer suppression, loss of tom-1, an inhibitor of the dense-core vesicle release [26], also partially abrogated dauer formation of TGF-β mutants (Figure 2A). By contrast, mutations in IFT components such as osm-5 and che-11 failed to alter dauer formation of TGF-β mutants [27]. Therefore, BBSome mutants display enhanced insulin signaling and have the same impact on dauer formation and maintenance as animals with enhanced dense-core vesicle secretion and as such are distinct from other ciliary defective mutants.
To functionally assess neuropeptide signaling in bbs mutants, we examined the NPR-1 signaling pathway. FLP-21 is one of two FRMF-amide neuropeptides that activates NPR-1, a neuropeptide-Y-like receptor that regulates C. elegans oxygen sensation and social feeding [28]. Overexpression of FLP-21 can partially suppress the hypomorphic npr-1(215F) but not the null mutation of npr-1 as the hypomorph can still bind to FLP-21, albeit with a lower affinity [28]. We found that mutations in either bbs-7 or bbs-9 also partially suppress the social feeding phenotype of npr-1(215F) but not that of the null mutant, suggesting that BBSome mutants, as our reporter assay indicated (Figure 1G,H), have increased FLP-21 release (Figure 2B). The hypersecreting tom-1 mutants similarly strongly suppressed the hypomorphic allele of npr-1 but only weakly suppressed the null allele (Figure 2B). These findings indicated that the enhanced secretion of neuropeptides as assessed by fluorescent reporters indeed correlate to functional increases in neuropeptide signaling in BBSome-deficient animals.
To determine if the effects of the BBSome on dense-core vesicle secretion are cell-autonomous, we expressed wild-type bbs-1 cDNA under the control of various sensory neuron promoters in bbs-1(ok1111) mutants. This bbs-1 transgene was fully functional as judged by its capacity to fully rescue the small body size, feeding, and dye-filling defect of bbs-1 mutants (Table S1; unpublished data). Expression of bbs-1 under its own promoter or that of ocr-2, which drives expression in ADL in addition to a few other sensory neurons (see Table S1 for expression pattern), abrogated the enhanced secretion of bbs-1 mutant animals (Figure 3A,B). By contrast, use of the tax-4 promoter to drive expression in numerous sensory neurons excluding ADL had no effect on the secretion phenotype of bbs-1 mutants (Figure 3A,B). Expression exclusively in ADL neurons showed partial but highly significant reduction in the enhanced insulin secretion of bbs-1 mutants, suggesting that the BBSome regulates secretion at least partially cell autonomously (Figure 3A,B). The partial suppression could be due to differences in promoter strength or regulation of secretion from other sensory neurons. Similar results were obtained when a functional bbs-7 cDNA was expressed in a bbs-7 mutant: two independently generated transgenic lines showed partial but highly significant reductions of the enhanced insulin secretion of bbs-7 mutants when expressed exclusively in ADL (Figure 3C,D). Finally, use of a heat shock inducible promoter to induce expression of wild-type bbs-7 in late fourth larval stage bbs-7 mutants when the developmental programs for ciliated neurons have been completed led to a partial but significant abrogation of enhanced secretion (Figure 3E,F). Thus, the requirement of the BBSome for wild-type secretion is, in part, cell autonomous and not a consequence of altered neural development.
As the cilia and BBS proteins have been shown to regulate transcription through the TCF/LEF1 transcription factor [29], we investigated whether the observed enhanced secretions could be secondary consequences of increased transcription in bbs mutants. While expression of FLP-21-mCherry translational reporter fusion using the ADL-specific srh-220 promoter resulted in increased accumulation of the secreted FLP-21-mCherry in coelomocytes (Figure 1H), expression levels of the Psrh-220::gfp transcriptional reporter were similar in wild type and bbs-9 mutants (Figure S3A,B). Furthermore, transcript levels in several bbs mutants, as measured by semi-quantitative RT-PCR, did not show any significant increase in transcription of srh-220, daf-28, flp-21, daf-7, or several other neuropeptide genes endogenously expressed in ADL and ASI neurons (Figure S3C). Thus, enhanced secretion levels of insulin, TGF-β ligand, and neuropeptides and the corresponding increase in their associated signaling pathways are unlikely to be secondary consequences of enhanced transcription of insulin and neuropeptide genes.
To better understand the enhanced secretions of bbs mutants, we next investigated the requirements of various components of the dense-core secretion machinery. Regulated release of dense-core vesicles is triggered by Ca+2 entry leading to activation of UNC-31/CAPS, which promotes vesicle fusion with the plasma membrane [30]. We previously showed that UNC-31 is required for insulin release from ADL [21]. Therefore, we investigated whether enhanced secretion of dense-core vesicle contents seen in bbs mutants was dependent on UNC-31. We found that secretion levels of bbs-7; unc-31 double mutants were indistinguishable from unc-31 mutant alone (Figure 4A,B), indicating that the BBSome is a negative regulator of Ca+2 stimulated release of dense-core vesicles and does not affect the constitutive basal release seen in unc-31 mutants. UNC-31 acts in parallel to TOM-1, the C. elegans tomosyn [26]. Tomosyns negatively regulate both dense-core and synaptic vesicle fusions by binding to and inhibiting syntaxins, an essential component of the vesicle fusion machinery [26]. As in bbs mutants, tom-1 mutants exhibited increased secretion from ADL (Figure 3A,B) and the tom-1 and bbs-7 secretion phenotypes were additive (Figure 4A,B). The finding that loss of unc-31 fully abrogated the hypersecretion of bbs mutants cannot distinguish whether the BBSome regulates secretion through mechanisms that directly modulate the unc-31 pathway from indirect mechanisms that ultimately depend on UNC-31 mediated secretion. However, since the hypersecretion of bbs mutants was fully abrogated by unc-31 but additive with that of tom-1 mutants, these findings suggest that UNC-31 and the BBSome act in a common release pathway in parallel to TOM-1.
Dense-core vesicles are trafficked from their site of formation in the Golgi to the plasma membrane by Rab GTPases and their regulators. Rab GTPases are a highly conserved family of proteins involved in various aspects of vesicle fusion and transport. Previous studies have indicated that the BBSome regulates Rab8 localization and activity in mammalian cells [4]. bbs and rab-8 mutants in C. elegans show similar defects in cilia membrane morphology, suggesting that the BBSome is also likely to regulate Rab8 activity in C. elegans [31]. However, rab-8 mutants exhibited wild-type levels of secretion from ADL and loss of rab-8 did not change the hypersecretion phenotype of bbs-7 mutants (Figure 4C,D). This suggests that the BBSome might regulate dense-core vesicle secretion through different Rabs. Previous studies have suggested that Rab3 and Rab27 regulate vesicular exocytosis by promoting the movement and tethering of dense-core vesicles to the plasma membrane [32],[33]. We found that AEX-6, the C. elegans homolog of Rab27, was essential for dense-core vesicle secretion from ADL and that its loss abrogated the enhanced secretion of bbs mutants (Figure 4E,F). By contrast, loss of rab-3, another GTPase implicated in exocytosis, had no effect (Figure 4E,F). The enhanced secretion of bbs mutants was also dependent on rabphilin/RBF-1, an effector of RAB-27/AEX-6 (Figure 4G,H), and AEX-3, a RabGEF previously shown to regulate RAB-27/AEX-6 and RAB-3 (Figure 3C,D) [32]. Taken together, these data are consistent with a model whereby the BBSome acts as a negative regulator of AEX-6/RBF-1/UNC-31, the worm counterpart of mammalian Rab27/rabphilin/CAPS exocytosis machinery. Given that the mammalian BBSome binds to Rabin8, a RabGEF for Rab8, and regulates Rab8 activity [4], the BBSome might regulate dense-core vesicle secretion by regulating the activity of AEX-3, the RabGEF for AEX-6/Rab27.
To further investigate the effects of bbs deficiency on dense-core vesicles, we examined subcellular localization of IDA-1-GFP, a tyrosine phosphatase-like receptor involved in protein secretion that has been used as a maker of dense-core vesicles [34]. The overall expression levels of IDA-1-GFP was similar in wild type and bbs mutants, and in both cases GFP puncta were clearly visible along axons as well as dendrites terminating in cilia (Figure S5 and unpublished data). While the IDA-1-GFP reporter is largely excluded from the cilia of wild type animals, it was prominently visible in those of bbs mutants (Figure S5). Presence of the marker in cilia of bbs mutants could suggest either mis-sorting of proteins normally localized to dense-core vesicles or the appearance of intact dense-core vesicles, which are normally excluded from cilia, within these structures.
One mechanism that has been proposed to ensure proper segregation of cilia-localized membrane proteins from other plasma membrane proteins is the proteinaceous structure at the base of the cilia known as the transition zone [35]. The transition zone is thought to provide a barrier that helps exclude non-ciliary proteins from the cilia [35]. Mutations in components of the transition zone underlie ciliopathies such as Merkel-Gruber syndrome and nephronophthisis, which share some clinical features such as renal abnormalities with Bardet-Biedl syndrome [2]. Unlike bbs mutations, however, mutations in genes encoding components of the transition zone, such as mksr-1 and mksr-2 did not cause enhanced neuroendocrine secretion (Figure S2).
The finding that loss of BBSome components enhances secretion of dense-core vesicles prompted us to examine the relationship between hyperactive neuroendocrine signaling and ciliary defects of bbs mutants. Since losses of various IFT components result in reduced secretion, we first investigated the requirement of functional cilia for the enhanced secretion of bbs mutants. We found that mutations in each of che-2 and che-11, which cause reduced secretion levels (Figure 1A,B), completely abrogate the enhanced secretion phenotype of bbs mutants (Figure S2). These findings further validate the notion that the excess secretion phenotype of bbs-deficient animals is not simply a consequence of defective cilia but rather that some level of normal ciliary function is required in order for the secretion phenotype of bbs mutants to be manifested.
We next investigated whether ciliary defects of bbs mutants were dependent on hypersecretion of dense-core vesicles. To assess structural integrity of cilia in various mutants, we used a dye-filling assay. Subsets of C. elegans sensory neurons have their ciliated endings located near the surface of the animals and are directly exposed to the environment for chemosensation. Upon immersion of whole animals in a fluorescent dye solution, these neurons become visible as the dyes can enter these neurons through their cilia (Figure S1) [36]. Since animals with defective cilia formation such as the IFT mutants fail to dye fill (Figure S1B) [36], this assay has been used to assess cilia integrity and proper opening of the cilia to the external environment. As previously reported, we found that BBSome mutants also failed to dye fill (Figure S1) [18],[37],[38]. Interestingly, while mutations in the Rab27/rabphilin/CAPS exocytosis machinery abrogated the enhanced secretion of bbs mutants, they did not restore structural integrity to cilia in these animals as judged by the inability of the double mutants to dye fill (Figure S1C,D). These findings suggested that the structural and functional defects of cilia in bbs mutants are unlikely to be a consequence of hypersecretion of dense-core vesicles.
To determine if ciliary structural deficiencies that manifest as dye-filling defects are a pre-requisite for hyperactive secretion, we examined the hypersecretion tom-1 mutants and found that they exhibited normal dye filling (Figure S1B). We also found that while bbs-5 mutants display similarly enhanced levels of secretion as other BBSome mutants (Figures 1C,D and 5A), they dye fill like wild-type animals (Figure S1A). Additionally, bbs-5 mutants displayed normal expression of the serotonin synthesis enzyme, tph-1 (Figure S4C,D), whereas expression of this enzyme was shown to be elevated in many cilia-defective mutants [39] (Figure S4C,D), further suggesting the presence of functional cilia, in bbs-5 mutants.
Together these findings suggest that the enhanced secretion phenotype of bbs mutants depends on retention of some level of ciliary function and that ciliary structural defects are not a pre-requisite for enhanced release of dense-core vesicles. In turn, enhanced release of dense-core vesicles can be abrogated without concomitant correction of ciliary structural abnormalities of bbs mutants.
As various organism-wide phenotypes are controlled by neuroendocrine signals, we asked whether some of the phenotypes seen in C. elegans bbs mutant might be due to enhanced neuroendocrine signaling in these mutants. Human BBS patients are short statured, bbs mutant mice are born small [13],[14],[40], and bbs mutant C. elegans have a reduced body size (Figure 5), suggesting that size regulation might be a conserved function of the BBSome. We found that tom-1 mutants, similar to BBSome-deficient animals, exhibited reduced adult body size (Figure 5B), suggesting that hypersecretion without ciliary defects can cause small body size. Considering that some IFT mutants also exhibit small body sizes (Figure 5D) [41], we sought to distinguish whether the small body size of bbs mutants could be attributed to either hypersecretion or ciliary defects. To do so, we suppressed the enhanced secretion of BBSome mutants with mutations in the dense-core vesicle exocytosis pathway (Figure 5B). Loss of rbf-1, which suppressed the enhanced dense-core vesicle release of bbs mutants, fully restored normal body size to bbs-7 mutants (Figure 5B,D). By contrast, rbf-1 mutation failed to change the small body of IFT mutants, suggesting that IFT and bbs mutants regulate size through different pathways (Figure 5D). Similarly, loss of unc-31 and aex-6, but not that of rab-3, conferred nearly wild-type body size to bbs-7 mutants (Figure 5B). Importantly, in all of these cases normalizations of body sizes were achieved without functional correction of ciliary defects, as the double mutants remained dye-filling defective (Figure S1C,D, unpublished data). Furthermore, tom-1 mutants that exhibited an additive secretion phenotype with BBSome mutants caused a further reduction in the size of bbs mutant (Figure 5B). Together, these findings suggest that the small size of BBSome mutants is due to enhanced secretion of dense-core vesicle cargoes and distinct from IFT mutants.
In further support of the notion that distinct mechanisms underlie the small sizes of BBSome and IFT mutants, we also found that neurons that function in size regulation in bbs mutants are distinct from the neurons that function in size regulation in IFT mutants. Normal body size can be restored to che-2 IFT mutants by expressing wild-type che-2 cDNA in a subset of ciliated neurons using the tax-4 promoter [42]. By contrast, expression of functional bbs cDNAs using a tax-4 promoter failed to alter the small body size of bbs mutants (Figure 5C and Table S1). The body-size phenotype of bbs mutants could, however, be reverted to wild-type levels when bbs cDNAs were expressed using an ocr-2 promoter, which directs expression to a distinct subset of neurons than those expressing tax-4. This result indicates the spatial requirements of the IFT machinery and the BBSome can be attributed to distinct, non-overlapping neurons in the regulation of body size (Figure 5C, see Table S1 for expression pattern). Thus, despite phenotypic similarities, the size phenotypes of BBSome and IFT mutants are based on distinct cellular and molecular mechanisms.
As in the case of small body size, we found that a metabolic phenotype exhibited by bbs mutants is mimicked by the hyperactive secretion mutant tom-1 but not by IFT mutants. Specifically, bbs mutants were previously reported to exhibit increased accumulation of Nile Red and bodipy-labeled fatty acids, vital dyes used as proxies for assessment of metabolic state in intact animals [17],[38],[43]–[45]. As in body size, tom-1 mutants exhibited a Nile Red phenotype that was reminiscent of that seen in BBSome mutants, while IFT mutants such as osm-5 and che-11 exhibited wild-type patterns of Nile Red staining (Figure 6A,C). As in body size and release of dense-core vesicles, tom-1 and bbs-7 mutants exhibited additive Nile Red phenotype (Figure S4A,B). Furthermore, mutations in the Rab27 exocytosis pathway, which abrogated the enhanced secretion of bbs mutants, restored wild-type levels of Nile Red staining to bbs-7 mutants (Figure 6B,D), suggesting that enhanced secretion underlies the Nile Red phenotype of bbs mutants.
In mice, the obesity seen in the setting of bbs deficiency has been attributed to hyperphagia [13]. C. elegans BBSome mutants similarly showed an altered food intake behavior (Figure 7). Food intake behavior in C. elegans is assessed by counting the pumping rate of the pharynx, an organ for ingesting bacteria [16],[46]. Pumping rate is modulated by the availability of food supplies, food quality, and prior feeding experience of the animals [47]. Although under plentiful food conditions, wild type and bbs-deficient C. elegans display similar pumping rate [38], we found that bbs-deficient animals, unlike wild-type animals, continue to exhibit rapid pharyngeal pumps even when food supplies are exhausted (Figure 7). This rapid pumping rate is unlikely to be merely a consequence of defective cilia as many other cilia-defective mutants exhibit wild-type pumping rate in the presence or absence of food (Figures 7E, S4C) [41]. As in the case of body size and vital dye staining patterns, reducing secretion of dense-core vesicles without concomitant correction of ciliary defects was sufficient to restore wild-type rates of food intake in bbs mutants (Figure 7A).
One molecular mechanism known to modulate pumping rate is serotonin signaling [48]. In C. elegans as in mammals, serotonin production is dependent on the tryptophan hydroxylase enzyme, tph-1. This enzyme is expressed in only a few neurons, only one pair of which, the ADF neurons, is ciliated [48]. To investigate whether abnormal serotonin signaling may underlie the enhanced pumping rate of food-deprived bbs mutants, we expressed functional bbs-1 cDNA using a tph-1 promoter in bbs-1 mutants. bbs-1 reconstitution was sufficient to fully restore wild-type pumping rate to bbs-1 mutants (Figure 7B, Table S1). Similarly, all other promoters used to reconstitute wild-type bbs-1 that express in ADF neurons fully rescued the feeding phenotype, whereas expression of bbs-1 using promoters that do not target ADF failed to alter the enhanced pumping rate of bbs-1 mutants (Figure 7B,C, see Table S1 for expression pattern). Similar results were obtained when bbs-7 cDNA was expressed in the ADF neurons of bbs-7 mutants (Table S1). Consistent with the rescue data, loss of tph-1 fully abrogated the enhanced pumping rate of bbs mutants (Figure 7D). These findings suggest that the elevated pumping rate of BBSome mutants is due to excessive serotonin signaling initiated from ADF neurons.
Since elevated expression of tph-1 in ADF neurons has been reported in numerous cilia-defective mutants including bbs (Figure S4C,D) [39], we sought to distinguish whether the excess serotonin signaling of bbs-deficient animals could be attributed to excessive production of serotonin or to increased release of this biogenic amine through dense-core vesicles. Despite elevated tph-1 expression, cilia-defective mutants such as che-2, che-11 and osm-5 exhibited wild-type pumping rates in the presence or absence of food (Figures 7E, S4C), suggesting that increased tph-1 expression does not necessarily correlate with excessive serotonin signaling. In contrast, the hypersecretion mutant tom-1 mimicked the bbs mutant feeding phenotype in that tom-1 mutants exhibited elevated pumping rate when removed from food (Figure 7E). Furthermore, bbs-5 mutants, in which cilia are relatively normal as assessed by dye filling (Figure S1A), exhibited wild-type level of tph-1 expression in ADF neurons yet displayed elevated pumping in the absence of food (Figure S4C,D). Thus, the elevated pumping rate of bbs mutants is likely due to excessive release of serotonin rather than increased synthesis via elevated expression of tph-1.
Taken together, our data indicate that while some ciliary-defective mutants display size, metabolic, and feeding phenotypes that are reminiscent of those seen in the bbs mutants, the underlying molecular and cellular bases of these phenotypes are likely to be distinct. In turn, mutations that cause enhanced secretion of dense-core vesicles without gross structural defects of cilia mimic several physiological phenotypes seen in bbs mutants.
Increased levels of circulating leptin and insulin in bbs-deficient mice and humans have been reported [1],[11],[14],[49]. In the case of insulin, the increase is generally assumed to be a secondary consequence of obesity and insulin resistance. However, our data raise the possibility that elevated levels of these circulating hormones may be in part a consequence of enhanced dense-core secretion rather than secondary outcomes of obesity. To further support this possibility, we employed Min6 cells, a mouse pancreatic β-cell line that is ciliated and expresses BBSome components. Treatment of Min6 cells with siRNAs targeting each of three distinct BBSome subunits, Bbs5, Bbs7 and Bbs9, resulted in ∼1.5–2-fold increase in secreted insulin compared to control siRNA (Figure 7F). Importantly, since these experiments were conducted in cell lines, the enhanced secretion could not be merely attributed to a secondary consequence of organismal obesity. Consistent with this interpretation, excess circulating level of leptin is evident in bbs-deficient mice before the onset of weight gain [11].
Although disruptions in either the IFT machinery or the BBSome can lead to similar structural and functional defects, here we show that they elicit opposite effects on dense-core vesicle secretion. Specifically, defects in the BBSome cause elevated secretions of dense-core vesicles, while IFT defects generally cause reduced secretions. Our findings indicate that this excess secretion is largely cell autonomous, is not a consequence of altered development, and is dependent on the C. elegans counterpart of Rab27/rabphillin/CAPS exocytosis machinery but distinct from the Rab8/Rabin8 vesicular transport machinery that helps target membranes and cargoes to the cilium [4]. Therefore, we propose that the role of the BBSome complex in vesicular transport within ciliated cells should be expanded from ciliary functions [4],[5],[8],[10],[18],[19] and melanosome movement [4],[50] to also include dense-core vesicle exocytosis.
The current paradigm for understanding the myriad manifestations of bbs deficiency is largely framed in the context of defective ciliary functions arising from a failure to properly sort receptors and other signaling molecules to the cilia [5],[6],[8],[10],[11],[18],[19]. As such, mutations in the IFT machinery and the BBSome are thought to cause similar ciliary defects albeit with different levels of severity. This view emerged, in part, by the observations that mutations in IFT components result in organism-wide phenotypes that resemble those of BBSome deficiency [12]. Our findings here indicate that while bbs deficient animals share some of the defects of IFT mutants, the consequences of these mutations on dense-core vesicle secretion and resultant phenotypes are dramatically different from one another. More importantly, our findings challenge the view that similar mechanisms underlie all of the phenotypic similarities caused by BBSome and IFT mutants. For instance, while mutants in both IFT and BBSome components share defects in body size, proper body size regulation requires these components in distinct subsets of sensory neurons. Moreover, the small body sizes of bbs mutants could be reverted to wild type upon abrogation of enhanced secretion of dense-core vesicles, while similar manipulations had no effects on the small body sizes of IFT mutant. Similarly, it has been reported that BBSome and IFT mutants displayed distinct phenotypes of a specific learning behavior in C. elegans [51]; our findings suggested that this difference is likely to reflect a role of dense-core vesicle release in this behavior. Finally, consistent with the notion that some of the phenotypic consequences of bbs deficiency are due to hypersecretion of dense-core vesicles, we found that tom-1 mutants, which cause hypersecretion without obvious ciliary defects, exhibit the size, metabolic, and feeding abnormalities similar to bbs mutants.
We do not currently know the precise mechanisms through which bbs deficiencies cause hypersecretion of dense-core vesicles. Although the enhanced secretion phenotype of bbs mutants can be manipulated independently of their ciliary defects, we found that wild type IFT activity is required for manifestation of the hypersecretion phenotype. These findings are consistent with several models. In one model, bbs mutants may mislocalize or excessively accumulate receptors that modulate dense-core vesicle release within the cilia [9]. Given that cilia bound vesicles and dense-core vesicles both arise at the trans-Golgi membrane, it is possible that as a coat-protein, the BBSome not only sorts the appropriate cilia-localized proteins but also prevents the inappropriate accumulation of other molecules in this organelle. Consistent with this scenario, we found that accumulation of IDA-1-GFP, a marker of dense-core vesicles normally excluded from the cilia, accumulates in cilia of bbs mutants (Figure S5). An alternative but not mutually exclusive possibility is that the BBSome may regulate dense-core vesicle release as a separate function from its role in sorting membranes and cargoes to the cilium. Given that the BBSome regulates cilia-bound vesicles through binding Rabin8 [4], it is possible that the BBSome regulates the Rab27/rabphillin/CAPS exocytosis pathway through binding specific regulatory components of this machinery such as AEX-3, the RabGEF for RAB-27/AEX-6. The idea that the transport machinery associated with cilia formation and function may not be restricted to these organelles and may have evolved from existing cellular machinery has been suggested by the recent demonstration that IFT proteins also play a role in polarized receptor trafficking to a cellular region reminiscent of the cilia in non-ciliated cells [52]. Additionally, the BBSome mediates the movement of melanosomes on microtubules tracks, an organelle not directly associated with the cilia, suggesting that the BBSome might have cilia-independent functions [50]. The genetic epistasis studies presented here do not distinguish between these possibilities. Whether the effects of bbs deficiency on dense-core vesicle secretion are cilia-independent functions of the BBSome or secondary consequences of ciliary abnormalities, our findings challenge the existing paradigm that phenotypic manifestations of the bbs deficiency are recapitulated by IFT mutations. The most salient insight to emerge from our findings is that several of the organismal phenotypes seen in bbs mutants are shared with hypersecretion tom-1 mutants and can be reverted to wild-type by abrogating the hypersecretion of these mutants without concomitant correction of ciliary defects. Although a number of defects seen in bbs patients are likely to be consequences of mis-localized ciliary proteins, we believe that the etiologies of some features of BBS merit re-evaluation in light of the role of this complex in dense-core vesicle secretions. While the specific hormonal pathways that underlie the behavioral and physiological abnormalities of C. elegans may or may not play similar roles in mammals, excessive hormonal secretions are likely to contribute to disease manifestations in BBS patients. For instance, obesity and hyperphagia of BBS patients may be a direct consequence of or exacerbated by increased secretion of appetite-promoting hormones, and hypertension in these patients could be caused by increased release of catecholamines. The finding that hypersecretion in bbs deficiency can be abrogated without correcting concomitant ciliary defects may open the door to new strategies for treating patients with Bardet-Biedl syndrome.
Strains were constructed by standard C. elegans methods [53]. Strains obtained from the C. elegans Genetics Center or the National Bioresourse Project were backcrossed at least 4× with wild type before phenotypic assessment. Transgenic animals were generated by injecting the plasmid of interest at 100 ng/µl with 20–50 ng/µl of Punc-122::GFP or Pmyo-3::GFP as co-injection marker. Plasmids were construct by Gateway cloning as described in [21].
Coelomocyte uptake assays were performed as described in [21] with the exception that 20–40 animals were imaged for each strain at sub-saturating exposure with the shutter half open to split fluorescence between the eyepiece and the camera.
∼30 sychronized adults for Ptph-1::GFP and L4 of Psrh-220::GFP expressing animals were imaged at 40× magnification on a Zeiss Axioplane microscope at a sub-saturating exposure. The neuron of interest was circled and GFP fluorescent within that neuron quantified with OpenLab software. The minimum fluorescence is subtracted from the mean for each cell. The mean and standard error of the mean was determined for each strain and and normalized to wild type.
Synchronized mid L4 animals were heat shocked on plates at 37°C for an hour and allowed to recover at 20°C for 4 h before phenotypic assessment.
DIC images of 20–40 animals were taken at 5× magnification and the outline of each animal was traced by OpenLab software. The mean and standard error of the perimeter was calculated for each strain and normalized to wild type.
∼200 synchronized L1 animals for each strain were plated onto five plates and incubated at 25°C for 2 d before the percentage of dauer animals was determined along with standard error of the mean.
The assay was preformed as described in [28] on six plates per strain. The percentage of animals feeding in groups was determined along with standard error of the mean.
∼300 synchronized L4 worms were washed into 1.5 ml eppendorf with S-basal and incubated with 1 µl of 10 mg/µl of DiI in 1 ml of S-basal for ∼2 h rotating at room temp. Worms were than pelleted and washed once with S-basal and plated onto an OP50 seeded plate for ∼1.5 h before imaging at 16× on a Zeiss Axioplan microscope.
Pumping rate was measured essentially as described in [16] with the exception that pumping rates were determined in 10-s intervals for 10 animals for each genotype. The average pumping rate and standard error of the mean were determined and normalized to wild type.
∼200 synchronized L1 animals were plated onto a 6 cm plate containing 0.05 µg/ml of Nile Red and incubated at 20°C for 3 d. Gravid animals were imaged using a TRITC filter at 16× magnification at a sub-satrating exposure with the shutter half open to split fluorescence between eyepiece and the camera on a Zeiss Axioplan microscope. The first pair of intestinal cells was traced using the DIC image in OpenLab and the mean Nile Red fluorescent minus the background was determined for 10–20 animals for each genotype. The average fluorescent along with the standard error of the mean was determined for each strain and normalized to wild type.
Min6 cells were treated with siRNAs from Qiagen with Hiperfect (Qiagen) for 4 d in triplicates. Cells were rested in Krebs-Ringer bicarbonate buffer for 2 h and then allowed to release insulin into the media for an hour. Insulin was measured by an ELISA kit (Mercodia) in duplicates and normalized to the total insulin obtained after cell lysis. The mean and standard error was calculated and expressed as a percentage of the control siRNA.
∼10,000 synchronized L1 for each genotype were grown to L4 at 20°C. RNA extraction, cDNA prepartion, and RT-PCR were preformed as described in [54] with primers and syber green PCR mix from Qiagen. The data were standardized to the actin gene, act-1. Data from two independent growths were aggragated for statistical analysis.
All statistical analyses were performed using a two-tailed student t test.
|
10.1371/journal.pbio.2005293 | Loss of RXFP2 and INSL3 genes in Afrotheria shows that testicular descent is the ancestral condition in placental mammals | Descent of testes from a position near the kidneys into the lower abdomen or into the scrotum is an important developmental process that occurs in all placental mammals, with the exception of five afrotherian lineages. Since soft-tissue structures like testes are not preserved in the fossil record and since key parts of the placental mammal phylogeny remain controversial, it has been debated whether testicular descent is the ancestral or derived condition in placental mammals. To resolve this debate, we used genomic data of 71 mammalian species and analyzed the evolution of two key genes (relaxin/insulin-like family peptide receptor 2 [RXFP2] and insulin-like 3 [INSL3]) that induce the development of the gubernaculum, the ligament that is crucial for testicular descent. We show that both RXFP2 and INSL3 are lost or nonfunctional exclusively in four afrotherians (tenrec, cape elephant shrew, cape golden mole, and manatee) that completely lack testicular descent. The presence of remnants of once functional orthologs of both genes in these afrotherian species shows that these gene losses happened after the split from the placental mammal ancestor. These “molecular vestiges” provide strong evidence that testicular descent is the ancestral condition, irrespective of persisting phylogenetic discrepancies. Furthermore, the absence of shared gene-inactivating mutations and our estimates that the loss of RXFP2 happened at different time points strongly suggest that testicular descent was lost independently in Afrotheria. Our results provide a molecular mechanism that explains the loss of testicular descent in afrotherians and, more generally, highlight how molecular vestiges can provide insights into the evolution of soft-tissue characters.
| While fossils of whales with legs demonstrate that these species evolved from legged ancestors, the ancestral state of nonfossilizing soft-tissue structures can only be indirectly inferred. This difficulty is also confounded by uncertainties in the phylogenetic relationships between the animals concerned. A prime example is the case of testicular descent, a developmental process that determines the final position of testes, which occurs in most placental mammals but is absent from several afrotherian lineages. Here, we discovered that afrotherians possess remnants of genes known to be required for testicular descent. These “molecular vestiges” show that testicular descent was already present in the placental ancestor and was subsequently lost in Afrotheria. Our study highlights the potential of molecular vestiges in resolving contradictory ancestral states of soft-tissue characters.
| In placental mammals—the eutherian crown group consisting of the clades Afrotheria, Xenarthra, and Boreoeutheria [1]—optimal testicular function requires a temperature that is lower than the body temperature. To achieve this, the testes are located outside of the abdominal cavity in a scrotum in many species such as primates, most rodents, lagomorphs, most carnivores, and most terrestrial artiodactyls [2, 3]. Alternatively, testes are located in the lower abdomen in dolphins, true seals, pangolins, and other mammals. In these species, testicular cooling is achieved by vascular countercurrent heat exchanger systems, as observed in dolphin [4]; direct cooling with blood from the hind limbs, as observed in seals [5]; or testicular cooling may not be necessary, as these species have lower body temperatures [2, 6, 7].
The position of the testes in the lower abdomen or in the scrotum is the result of a developmental descent process (S1 Fig). During mammalian development, testes initially form at a position near the kidneys in the embryo. Testicular descent into the scrotum occurs in two phases: first from the abdomen to the inguinal canal and second through the inguinal canal into the scrotum [8, 9]. The first transabdominal phase is governed by the growth and reorganization of the gubernaculum, a ligament that connects the lower pole of the testes and inner ring of the future inguinal canal [8–10]. Migration of the testes is caused by the swelling of the distal gubernaculum, which anchors the testis to the inguinal canal, while the abdominal cavity enlarges. The second inguinoscrotal phase is dependent on androgen signaling and requires the elongation of the gubernaculum, which migrates into the scrotum [8–10]. The involved signaling and mechanics make testicular descent a difficult and complex developmental process. Failure in any of the descent phases results in a pathological condition called cryptorchidism (absence of testes from the scrotum), which is a congenital birth defect observed at an appreciable frequency in human males (2%–4% at birth [11]) and other animals (up to 10% in male dogs [12], 2% in male cats [13], 2%–8% in male horses [14]).
Almost all placental mammals exhibit either partial descent (only the transabdominal phase), which results in ascrotal testes located in the lower abdomen, or complete descent (transabdominal and inguinoscrotal phase), which results in scrotal testes [2, 3]. A notable exception is Afrotheria, in which five of the six main lineages (represented here by the lesser hedgehog tenrec, cape golden mole, cape elephant shrew, manatee, elephant, and rock hyrax) do not show any testicular descent and have testes positioned at their initial abdominal position near the kidneys [2, 3, 15–17]. This lack of any testicular descent is termed testicondy. The aardvark is the only afrotherian exhibiting descended but ascrotal testes [2, 3, 18]. A schematic illustration of the different position of testes in mammals is shown in S1 Fig.
Since Afrotheria represent one of the three main clades of placental mammals (together with Xenarthra and Boreoeutheria), two different evolutionary scenarios could explain testicondy in several afrotherian lineages. First, if testicondy is the ancestral condition in placental mammals, then testicular descent was gained two or three times (depending on the phylogeny) in Xenarthra, Boreoeutheria, and the aardvark lineage. Second, if testicular descent is the ancestral condition in placental mammals, then testicular descent was lost once or more often (again depending on the phylogeny) in five of the six afrotherian lineages. Since soft-tissue structures like testes and the transient gubernaculum ligament are typically not preserved in the fossil record, the evolution of such soft-tissue structures can only be inferred by analytical methods such as parsimony, extant phylogenetic bracketing, or maximum likelihood [19–22], all of which rely on the given phylogenetic tree. Consequently, resolving whether testicondy or testicular descent is the ancestral condition in placental mammals requires accurate knowledge of the underlying phylogeny.
Unfortunately, although integrative approaches using both morphological and molecular characters have brought major advances in our understanding of mammalian phylogeny [23–26], there is still no final consensus on the relationships between (and sometimes within) the main clades of placental mammals. In particular, the placental root and branching pattern of the clades Afrotheria, Xenarthra, and Boreoeutheria are still debated [26–28] (S2 Fig), and an analysis of rare genomic events raised the concrete possibility of a near-simultaneous split [29]. Furthermore, the phylogeny within Afrotheria is not well resolved, because of conflicting evidence for the position of the aardvark (the only nontesticond afrotherian lineage) and the relationships between manatees, elephants, and hyraxes [30–34] (S3 Fig).
Given these phylogenetic uncertainties, it is probably not surprising that two different studies reached opposite conclusions about whether testicondy is the ancestral or derived state for placental mammals and for Afrotheria. Werdelin and Nilsonne [2] inferred that testicular descent in placental mammals and Afrotheria is the ancestral condition (testicular descent was subsequently lost). However, their results were based on a phylogeny in which Afrotheria were nested within Boreoeutheria, which is not supported by current phylogenies. More recently, Kleisner and colleagues [3] reexamined the evolution of testicular descent in the context of current phylogenies and came to the opposite conclusion that testicondy in placental mammals and Afrotheria is the ancestral phenotypic character.
Here, we sought to resolve this debate whether testicondy or testicular descent is the ancestral condition in placental mammals and in Afrotheria by using molecular evidence. First, we reasoned that if testicular descent is ancestral, then testicond afrotherian lineages may have lost key genetic information that is necessary for testicular descent. Such a loss of genetic information may be detectable by comparative genomics analysis. Second, we reasoned that if testicular descent is ancestral and if aardvarks are nested within Afrotheria, then testicondy would have evolved independently several times. This is expected to leave a signature of independent loss of the genetic information that is necessary for testicular descent. By analyzing the evolution of two key genes (relaxin/insulin-like family peptide receptor 2 [RXFP2] and insulin-like 3 [INSL3]) that are required for gubernaculum development and function in 71 placental mammals, we found that both genes have loss-of-function mutations only in several testicond afrotherian species. The absence of shared inactivating mutations and our age estimates for the loss of RXFP2 further suggest that testicondy evolved independently in afrotherian lineages at different time points. Together, these results provide not only a molecular mechanism that explains the loss of testicular descent in afrotherian lineages but also shows that testicular descent is the ancestral state for placental mammals and Afrotheria.
To determine if testicondy is the ancestral or derived condition for placental mammals and for Afrotheria, we examined two key genes that are necessary and sufficient for the development of the gubernaculum: INSL3 and RXFP2. INSL3 encodes a relaxin-like hormone that is secreted by Leydig cells of the testes and binds specifically to the transmembrane receptor encoded by RXFP2, which is highly expressed in gubernacular cells [35–39]. The INSL3-RXFP2 ligand-receptor pair promotes gubernacular cell proliferation and stimulates the swelling reaction [40–42]. Both genes are necessary for gubernacular function, as knockout of RXFP2 [37, 39] or INSL3 [40, 43, 44] in mice results in the absence of the gubernaculum and no testicular descent, which in turn leads to spermatogenesis defects and male infertility. Despite the fact that RXFP2 is also expressed in postmeiotic spermatogenic cells, surgically correcting the position of undescended testes in global INSL3 knockout mice or a knockout of RXFP2 that is restricted to male sperm cells results in normal spermatogenesis and fertility [39, 43], suggesting that both genes are dispensable for spermatogenesis and germ cell survival in adult male mice.
To investigate the evolution of RXFP2 and INSL3 in placental mammals, we made use of existing genome alignments between humans and 68 other mammals [45]. In addition, we further computed a genome alignment between human and the most recent genome assemblies of the rock hyrax and Hoffmann’s two-toed sloth (Materials and methods). Inspecting the genomic loci that correspond to human RXFP2 and INSL3 allowed us to examine both genes in all 70 mammals, even in the absence of gene annotations for most of these species.
To investigate if testicond Afrotheria lost the genetic information necessary for testicular descent, we first examined the coding region of RXFP2 and INSL3 in seven afrotherians with available genomes (aardvark, lesser hedgehog tenrec, cape golden mole, cape elephant shrew, manatee, elephant, and hyrax). Our genome alignments revealed that four testicond lineages (tenrec, cape golden mole, cape elephant shrew, and manatee) have several mutations in RXFP2 that inactivate its reading frame. These gene-inactivating mutations create premature stop codons, shift the reading frame, disrupt the splice site dinucleotides, and delete entire exons (Fig 1A). Furthermore, three out of these four species (tenrec, cape elephant shrew, manatee) also have inactivating mutations in the INSL3 gene (Fig 2A). Since these mutations affect several exons and destroy functional protein domains in INSL3 (A- and B-chain, Fig 2A), it is highly unlikely that the remnants of these genes encode a functional protein. Importantly, reciprocal-best BLAST hits and conserved gene order clearly show that these remnants are “molecular vestiges” that correspond to the RXFP2 and INSL3 genes (S4 Fig). In analogy to vestigial organs, these molecular vestiges imply the presence of once functional RXFP2 and INSL3 orthologs that were subsequently lost in several afrotherians during evolution.
To confirm that these inactivating mutations are real and do not represent genome assembly or alignment errors, we used a multistep validation approach. Since genome alignments do not take reading frame and splice site information into account, we first sought to rule out the possibility that inactivating mutations are a consequence of alignment ambiguities. To this end, we realigned all coding exons with the Coding Exon-Structure Aware Realigner (CESAR), an exon alignment method that produces an alignment with consensus splice sites and an intact reading frame whenever possible [49, 50]. CESAR confirmed that all affected exons exhibit inactivating mutations (Figs 1 and 2A). Second, to validate that these mutations are not sequencing or assembly errors, we investigated raw sequencing reads from the Sequence Read Archive (SRA) [51]. For both RXFP2 and INSL3, we found that all genomic loci containing an inactivating mutation are supported by at least 10 sequencing reads, while not a single read aligns to a putative sequence, in which the inactivating mutation was reversed to its ancestral state. We further confirmed the presence of two frameshifting mutations in RXFP2 in the lesser hedgehog tenrec by PCR and Sanger sequencing (S5A and S5B Fig). Overall, this shows that the inactivating mutations shown in Figs 1 and 2A are not sequencing errors or artefacts arising from genome assembly or alignment issues. Finally, we investigated whether hitherto undetected functional copies of RXFP2 or INSL3 exist in afrotherians that may have arisen by lineage-specific duplications. By performing ultra-sensitive genome alignments, we only detected a single orthologous locus for RXFP2 and INSL3. In addition, we found alignments to RXFP1, a paralog of RXFP2 that exists in all placental mammals (S6 Fig), showing that these alignment parameters are sufficiently sensitive to even detect more ancient gene duplications. Together, this excludes the possibility that afrotherians possess another functional duplicated copy of RXFP2 or INSL3.
If RXFP2 and INSL3 are truly lost, we further expect that they evolve neutrally in the lineages with inactivating mutations. Indeed, using RELAX [52], we found that RXFP2 evolves under relaxed selection in all four gene-loss species (adjusted P values < 3.2e−5, S1 Table). For INSL3, no significant evidence for relaxed selection was found, likely because large deletions in this short 131-residue protein in tenrec, cape elephant shrew, and manatee (Fig 2A) severely reduced alignment length. Therefore, we inspected the two protein domains that are necessary for the function of the mature INSL3 hormone. Similar to insulin, the preprohormone INSL3 is processed into an A- and B-chain peptide. The A- and B-chain then forms a heterodimer that is stabilized by two disulfide bonds between the A- and B-chains and one disulfide bond within the A-chain [47]. We found that tenrec, cape elephant shrew, and manatee have deletions that overlap the A- and B-chain and affect residues that are critical for INSL3 structure and function (Fig 2B). Together, our results conclusively show that the remnants of RXFP2 and INSL3 cannot encode functional proteins in several testicond afrotherian lineages.
Since INSL3 lacks clear inactivating mutations in the cape golden mole, we examined the residues that are important for INSL3 structure and function. We found that the Cys at position 10 in the A-chain that forms a disulfide bond with Cys at position 15 [47] is mutated to a Tyr in the cape golden mole (Fig 2B). Furthermore, the Lys at position 8 in the B-chain (Fig 2B), a residue that is important for receptor activation [48], is deleted in this species. This suggests that, while INSL3 still has an intact reading frame in the cape golden mole, it accumulated mutations that most likely render the encoded protein nonfunctional.
Interestingly, both RXFP2 and INSL3 lack any gene-inactivating mutations in the elephant and the rock hyrax, two afrotherians that are also testicond [15, 17]. While the elephant has a 2-bp deletion in the last exon of RXFP2, this merely truncates the C-terminus by 25 residues and is not an indication of loss (see section RXFP2 and INSL3 are intact in all nontesticond placental mammals and S7 Fig). Furthermore, RELAX estimates a Ka/Ks value of 0.31 and 0.33 for elephant and rock hyrax, respectively, which is slightly but not significantly higher than the Ka/Ks value of 0.27 observed for other mammals. Thus, there is no significant evidence for relaxed selection in these two lineages. We also scanned both genes for amino acid mutations that were only observed in human cryptorchidism patients (V18M, P49S, W69R, P93L, R102C, R102H, R105H, N110K in INSL3 and T222P in RXFP2 [8]). Whereas elephant INSL3 exhibits the R102H mutation, this mutation is observed in many other nontesticond mammals, and cell line experiments have shown that this mutation does not affect INSL3 activity [38]. Similarly, elephant RXFP2 has a T222A (Thr to Ala) mutation at a position where a mutation from Thr to Pro renders RXFP2 nonfunctional [37, 53]. However, the T222A mutation that is present in elephant is also observed in the nontesticond aardvark and pangolin, and experiments have shown that mutating this Thr to Ala does not affect RXFP2 function [53]. The rock hyrax does not exhibit any of the mutations observed in human cryptorchidism patients. Based on these evidences, elephant and rock hyrax RXFP2 and INSL3 may encode functional proteins.
So far, our analysis suggests that the function of RXFP2 and INSL3 is only compromised in several testicond afrotherians. Therefore, we examined both genes in the aardvark, the only afrotherian exhibiting partial testicular descent [2, 18], and found that RXFP2 and INSL3 are intact and evolve under selection (S1 Table) and that INSL3 lacks any mutations of critical amino acids (Fig 2B). To further investigate the relation between loss of RXFP2 and INSL3 and testicondy, we examined both genes in 64 other nontesticond mammals. While the genome alignment showed a few putative inactivating mutations in some species, a detailed manual inspection revealed that these are assembly errors (S8 Fig), assembly gaps (S9 Fig), and alignment ambiguities (S10 Fig). For RXFP2, we further found that the N-terminus is 17 amino acids longer in human, chimpanzee, bonobo and gorilla (S11 Fig) and that several species have small truncations and elongations of the C-terminus without affecting the transmembrane domains of the receptor (S7 Fig). Since RXFP2 does not evolve under relaxed selection in any of these 64 mammals (S1 Table), these length variations are not an indication of gene loss but support previous observations that N- and C-termini of proteins are evolutionarily less constrained [49, 54]. Together, this shows that both genes are intact and under selection in all other nontesticond placental mammals.
Interestingly, we observed no inactivating mutations in RXFP2 and INSL3 that are shared among any testicond afrotherian species, suggesting that the loss of these genes happened independently, after these species split from their common ancestors. Since RXFP2 and INSL3 are expected to evolve neutrally after the loss of testicular descent, an estimate of how long these genes have been evolving neutrally provides an estimate for when testicondy occurred. To this end, for the four branches in the phylogenetic tree leading to the four gene-loss species, we estimated the portion of the branch where the gene evolved under selection and the portion where it evolved neutrally, as described in [55, 56]. Since large parts of INSL3 are deleted in tenrec, cape elephant shrew and manatee (Fig 2A) and since exon 1 overlaps assembly gaps in several other species (Fig 3), we focused on RXFP2, for which each gene-loss species provides at least 594 bp in the codon alignment, to obtain robust estimates.
As shown in Fig 4 and S2 Table, we estimate that each lineage lost the RXFP2 gene at different time points. Consistent with the large number of inactivating mutations, RXFP2 appears to be lost first in the cape elephant shrew around 66–83 million years ago (Mya). For the lesser hedgehog tenrec, we estimate that RXFP2 loss happened around 50–59 Mya. Consistent with this estimate, we found that the same gene-inactivating mutation in RXFP2 exon 17 is shared with its sister species greater hedgehog tenrec (S5B Fig), suggesting that RXFP2 loss already occurred in the ancestor of both tenrec species that lived 7–14 Mya (Fig 4). The loss of RXFP2 in manatee is estimated to have happened around 43–51 Mya and thus likely predates the split of the manatee and dugong lineage 26–53 Mya. To test this, we used PCR and sequencing experiments and found that the dugong shares two stop codon mutations in different exons with the manatee (S5C and S5D Fig), confirming that RXFP2 loss predates the split of manatees and dugongs. The loss in the cape golden mole likely happened more recently (23–28 Mya), consistent with our observations that INSL3 did not yet accumulate a gene-inactivating mutation. While the absolute estimates of gene-loss times are tentative (as fossil-based time calibrations of the species divergence times are lacking), these time intervals are substantially different from each other (Fig 4). Together with the absence of shared inactivating mutations, this strongly suggests that testicondy evolved independently in Afrotheria.
The evolution of testicondy and whether testicular descent is the ancestral [2] or derived state [3] for placental mammals and for Afrotheria has been controversial, despite agreement in the phenotypic character assignment. The different conclusions are mainly due to persisting differences in the phylogeny, which affect ancestral character reconstruction. To resolve this debate, we investigated the evolution of RXFP2 and INSL3, two genes encoding a hormone receptor pair that is required for the development of the testes-descending gubernaculum ligament. We found remnants of once functional orthologs of RXFP2 and INSL3 as molecular vestiges in four testicond afrotherian lineages. Together with the presence of orthologs of both genes in other Afrotheria and other placental mammals, this shows that these genes were lost after the testicond lineages split from the afrotherian ancestor. This allows us to conclude that testicular descent is the ancestral condition in placental mammals and was subsequently lost in different afrotherian lineages. Importantly, our conclusion holds regardless of persisting phylogenetic discrepancies that involve the branching pattern of Afrotheria, Xenarthra, and Boreoeutheria at the placental root, and the phylogeny within Afrotheria (S2 and S3 Figs).
Our study also provides three lines of evidence that testicondy evolved independently in Afrotheria. First, if testicondy evolved in the common ancestor of any two testicond afrotherians, we would expect inactivating mutations in RXFP2 and INSL3 that are shared among species. However, both genes do not exhibit any shared inactivating mutations. Second, by estimating how long these genes have been evolving neutrally, we found different time intervals in which gene loss and likely testicondy evolved. These estimated time intervals may be helpful to better understand the ecological conditions under which testicondy evolved repeatedly. Third, independent evolution of testicondy is further supported by our finding that the testicond elephant and rock hyrax have intact RXFP2 and INSL3 genes that still appear to evolve under purifying selection. Since recent shifts from purifying selection to neutral evolution will not leave a detectable signature of significantly increased Ka/Ks ratios, our results suggest that elephant and rock hyrax RXFP2 and INSL3 may have come under relaxed or no selection in recent evolutionary time. In agreement with a more recent evolution of testicondy, the rock hyrax still exhibits rudiments of gubernacular structures [15]. It is possible that amino acid mutations or cis-regulatory mutations that affect the expression of RXFP2 or INSL3 are responsible for testicondy in elephant and rock hyrax. Indeed, the RXFP2 promoter region exhibits more sequence divergence in the rock hyrax than in all other nontesticond species, and the elephant sequence also shows elevated divergence compared to most but not all mammals (S12 Fig). However, the precise genomic basis of testicondy in elephant and rock hyrax remains to be elucidated.
Spermatogenesis and sperm storage require temperatures below 37 °C. Body temperatures below 35 °C, as observed in the heterothermic tenrecs and golden moles [7], may abolish the need for testicular cooling, which would explain why testicondy is tolerated in these species. However, other afrotherian species have body temperatures similar to many scrotal mammals. For example, the body temperature of elephants and elephant shrews is 36.8 °C and 37.2 °C, respectively [7, 58]. This raises the questions of how these testicond afrotherian lineages maintain normal testicular function at a body temperature of approximately 37 °C and whether these species have evolved novel cooling mechanisms. Previous anatomical studies did not reveal conclusive evidence of such cooling systems. In particular, testicond afrotherians do not possess a pampiniform plexus or a vascular countercurrent heat exchanger that could act as a comparable cooling system [17, 59, 60]. In the dugong, the epididymis has a large, folded surface and is surrounded by highly vascularized tissue [60]; however, whether this arrangement acts as a cooling system is not clear. In rock hyrax and elephant, it was observed that the part of the testicular duct that stores spermatozoa lies close to the body surface [59], but it remains unknown whether this anatomical feature has a cooling function during sperm storage. Hence, detailed comparative and functional studies of the testes, epididymis, and their associated anatomical structures are required to understand whether Afrotheria evolved novel mechanisms for sperm cooling during spermatogenesis and storage. Furthermore, it would be of great interest to learn why several afrotherian but not any other placental lineages have completely lost testicular descent. For example, it is unknown whether the loss of testicular descent is beneficial for these species, whether anatomical constraints associated with body plan or life history explain the loss of descent, or whether this phenotypic reversal is an evolutionary tradeoff for another advantageous trait. More anatomical studies that in particular compare the development of Afrotheria are required to address these questions.
More generally, our results highlight how molecular evolution can shed light on the evolution of phenotypes. While the fossil record has contributed substantially to our understanding of how hard-tissue characters such as bones, teeth, or shells evolved, the evolution of soft-tissue structures that do not fossilize can often only be inferred with analytical methods. Moreover, conclusions about the ancestry of such characters are dependent on the underlying phylogeny. Molecular vestiges offer an alternative strategy to investigate character ancestry. This strategy may be broadly applicable, since molecular vestiges are also known for other phenotypes whose ancestry is not controversial based on fossil evidence and accurate knowledge of the underlying phylogeny. For example, remnants of enamel-related genes in toothless mammals [61–63], remnants of hair development genes in hairless cetaceans [63], remnants of gastric genes in vertebrates without a stomach [64], or remnants of eye-related genes in subterranean mammals with degenerated eyes [65–68] show that these species descended from ancestors with teeth, hair, a stomach, and functional eyes, respectively. Thus, instead of investigating a soft-tissue structure directly, one can trace the evolution of genes that are crucial for the development of this structure. Molecular vestiges of such genes can then provide insights into the ancestry of soft-tissue structures, even if the phylogenetic positions of the respective species remain controversial.
To investigate the coding region of RXFP2 and INSL3 in placental mammals, we used an alignment between the human hg38 genome assembly and the genomes of 68 nonhuman placental mammals [45]. Since this alignment does not provide the rock hyrax (Procavia capensis) and Hoffmann’s two-toed sloth (Choloepus hoffmanni), two mammals for which improved genome assemblies have recently become available, we used the pipeline of lastz (version 1.03.54, parameters K = 2,400, L = 3,000, Y = 9,400, H = 2,000, and the lastz default scoring matrix) [69], chaining (parameters chainMinScore 1,000, chainLinearGap loose), and netting (default parameters) [70] to compute new genome alignments. Since several species have assembly gaps overlapping INSL3 exon 1, likely because this GC-rich region gets poor coverage in Illumina sequencing, we further computed genome alignments using updated assemblies of bonobo, domestic goat, camel, and dolphin. We also inspected the genome alignment to the recent gorGor5 assembly of the gorilla, provided by UCSC [71]. For gorilla, goat, and dolphin, the new assemblies closed these assembly gaps and revealed an intact INSL3 exon 1. For bonobo and camel, the new assemblies still have assembly gaps overlapping INSL3 exon 1. S3 Table lists all species and their assemblies that were analyzed.
We examined the colinear alignment chains (loaded in the UCSC genome browser) in order to assess if gene order around the RXFP2 and the INSL3 genes is conserved in Afrotheria. This showed that both genes and several neighboring genes occur in a conserved order in afrotherian genomes. Furthermore, existing gene annotations (Ensembl) and reciprocal-best BLAST of the human proteins confirmed that the aligning loci contain orthologous RXFP2 and INSL3 genes or their remnants.
To perform ultra-sensitive genome alignments that could reveal any hitherto undetected functional copies of RXFP2 or INSL3, we first computed genome alignments between the human genome (hg38 assembly) and each of the seven afrotherians, using highly-sensitive lastz parameters by setting K = 2,000 and L = 2,500. Subsequently, we used even more sensitive parameters to find additional local alignments by running lastz with parameters K = 1,500, L = 2,200, and W = 5 on all chain gaps (nonaligning regions flanked by aligning blocks) that are at least 30 bp and at most 500 Kb long, as described in [45, 72]. We built alignment chains from all local alignments [70] and visualized them in the UCSC genome browser [71].
We used the genome alignments to systematically search each exon for frameshifting insertions and deletions, mutations that create in-frame stop codons, and mutations that disrupt the conserved splice site dinucleotides (donor GT/GC, acceptor AG). We validated each putative mutation by the following steps: First, we realigned the exonic sequence of each exon with an inactivating mutation, using CESAR [49, 50] with default parameters. CESAR is a Hidden-Markov-Model based aligner that takes the reading frame and the splice sites into account and tries to produce an intact exon alignment (no inactivating mutations, consensus splice sites) between human and a query species, wherever possible. All exons for which CESAR confirmed the existence of at least one inactivating mutation are shown in Fig 3. S10 Fig shows an example for which CESAR found an alternative alignment that lacks inactivating mutations.
Second, we validated the remaining inactivating mutations by searching SRA for unassembled sequencing reads that have a 100% match to the genomic context comprising at least 50 bp around the mutation, as described in [73]. Furthermore, we searched SRA for a putative sequence, in which the inactivating mutation was reversed to its ancestral state. Every mutation that is supported by at least 10 sequencing reads without any hit to the putative “ancestralized” sequence was considered as real.
Third, we examined all exons that do not align in the query genome, since these are either truly deleted or an artifact arising from incomplete genome assemblies [68, 74]. We used the nearest up- and downstream aligning blocks in the alignment chain to determine the genomic locus that corresponds to this exon in the query genome. If this locus does not overlap an assembly gap, we concluded that the exon is lost (either deleted or accumulated so many mutations that it does not align anymore). If this locus overlaps an assembly gap, we conservatively considered this exon as missing sequence, as described in [63]. Indeed, as shown in S9 Fig, exons that overlapped assembly gaps in a previous assembly readily aligned with an intact reading frame in an improved genome assembly, showing that assembly gaps should not be mistaken for exon loss.
To determine if the RXFP2 coding sequence evolves neutrally in different species, we used RELAX version 2 [52]. Briefly, given an alignment, a tree topology, and branches labeled as either test or reference, RELAX determines if the gene evolves under relaxed or intensified selection in the test branches relative to the reference branches. We used RELAX’s partition descriptive method that fits three Ka/Ks classes to the test and the reference branches and also estimates an overall Ka/Ks value. To determine if RXFP2 evolves under relaxed selection in any mammal, we iteratively applied RELAX, labeling each of the 71 species as the test branch. The resulting P values were corrected for multiple testing using the Benjamini and Hochberg method (S1 Table). To exclude any bias caused by relaxed selection in the RXFP2-loss species, which might inflate the reference Ka/Ks values, we further applied RELAX, labeling only one afrotherian species as test and excluding all other afrotherian species from the reference branches. These tests confirmed that RXFP2 evolves under relaxed selection only in the four afrotherians with an inactivated RXFP2 gene but not in the three other afrotherian species (S1 Table).
To date the loss of RXFP2, we followed the procedure described in [55, 56]. For a branch along which the gene was inactivated, this method assumes that a gene evolves under a selective pressure similar to that in other species until it is inactivated. Afterward, the gene is assumed to accumulate both synonymous and nonsynonymous mutations at a neutral rate. The Ka/Ks value (K) estimated for this entire branch is then the average of the Ka/Ks value for the part of the branch where the gene was under selection (Ks) and the Ka/Ks value for the part of the branch where the gene evolved neutrally (Kn = 1), weighted by the proportion of time for which the gene was evolving under selection (Ts / T) and neutrally (Tn / T):
K=Ks×Ts/T+Kn×Tn/T,
where T represents the time since the split from the last common ancestor. Using the lower and upper bound of the confidence interval for the species divergence time T obtained from TimeTree [57] (S4 Table) and using the Ka/Ks value for mammals with a functional RXFP2 (Ks = 0.2491), one can estimate a lower and upper bound for Tn as
Tn=T×(K−Ks)/(1−Ks),
which provides an estimate of how long RXFP2 has been evolving neutrally (S2 Table).
DNA of the greater hedgehog tenrec (Setifer setosus) and the lesser hedgehog tenrec (Echinops telfairi) was kindly provided by Athanasia Tzika (University of Geneva). Dugong tissue was kindly provided by Joerns Fickel (Leibniz-Institute of Zoo and Wildlife Research, Berlin). DNA was extracted using the innuPrep DNA Mini Kit (Analytik Jena, Jena, Germany) following the manufacturer’s instructions (protocol for DNA isolation from tissue samples or rodent tails) but extending tissue lysis overnight and including RNA digestion.
Standard PCR reactions were performed in a total volume of 10 μl containing 5–20 ng DNA, 5 μl Phusion Flash High-Fidelity PCR Master Mix (Thermo Scientific, Waltham, United States of America; #F-548S), 0.25% DMSO (New England Biolabs, Frankfurt Main, Germany), and 0.5 μM of each primer (S5 Table). Cycling conditions were as follows: 35 cycles were used with denaturation at 98 °C (15 s but 2 min for the first cycle), annealing at a primer pair–specific temperature (S5 Table, 20s), and extension at 72 °C (10 s–1 min but 5 min after the last cycle).
PCR products were Sanger-sequenced after enzymatic cleanup with CleanDTR (GC biotech, Waddinxveen, Netherlands) followed by cycle sequencing with the PCR primers and a mix of the BigDye Terminator v3.1 Cycle Sequencing Kit and dGTP BigDye Terminator v 3.0 Kit (Applied Biosystems, Foster City, CA, USA). Analyses were performed on an ABI 3730xl DNA Analyser (Applied Biosystems).
We analyzed per-species sequence divergence of the promoter region of RXFP2 and INSL3. Using the human hg38 assembly coordinates chr13:31739464–31739544 and chr19:17821524–17821716, we extracted the sequence of all placental mammals, used PRANK [75] with parameters “-keep -showtree -showanc -prunetree” to align these sequences and to reconstruct the most likely sequence of the placental mammal ancestor and measured the percent sequence identity between the ancestral sequence and the sequence of every extant mammal, as described in [68, 74]. The result is visualized in S12 Fig.
|
10.1371/journal.ppat.1002872 | Convergent Evolution of Argonaute-2 Slicer Antagonism in Two Distinct Insect RNA Viruses | RNA interference (RNAi) is a major antiviral pathway that shapes evolution of RNA viruses. We show here that Nora virus, a natural Drosophila pathogen, is both a target and suppressor of RNAi. We detected viral small RNAs with a signature of Dicer-2 dependent small interfering RNAs in Nora virus infected Drosophila. Furthermore, we demonstrate that the Nora virus VP1 protein contains RNAi suppressive activity in vitro and in vivo that enhances pathogenicity of recombinant Sindbis virus in an RNAi dependent manner. Nora virus VP1 and the viral suppressor of RNAi of Cricket paralysis virus (1A) antagonized Argonaute-2 (AGO2) Slicer activity of RNA induced silencing complexes pre-loaded with a methylated single-stranded guide strand. The convergent evolution of AGO2 suppression in two unrelated insect RNA viruses highlights the importance of AGO2 in antiviral defense.
| Multi-cellular organisms require a potent immune response to ensure survival under the ongoing assault by microbial pathogens. Co-evolution of virus and host shapes the genome of both pathogen and host. Using Drosophila melanogaster as a model, we study virus-host interactions in infections by Nora virus, a non-lethal natural pathogen of fruit flies. Insects depend on the RNA interference (RNAi) pathway for antiviral defense. A hallmark of the antiviral RNAi response is the production of viral small RNAs during infection. We detected Nora virus small RNAs during infection of Drosophila, demonstrating that Nora virus is a target of the antiviral RNAi pathway. Furthermore, we show that Nora virus viral protein 1 (VP1) inhibits the catalytic activity of Argonaute-2, a key protein of the RNAi pathway. The 1A protein of Cricket paralysis virus suppresses RNAi via a similar mechanism. Importantly, whereas Nora virus persistently infects Drosophila, Cricket paralysis virus induces a lethal infection. Our findings thus indicate that two distantly related viruses independently evolved an RNAi suppressor protein that targets the Argonaute-2 protein. Altogether, our results emphasize the critical role of Argonaute-2 in insect antiviral defense, both in lethal and persistent infections.
| An efficient antiviral immune response is essential for the control or elimination of virus infection and for survival of the infected host. The immune system exerts a strong evolutionary pressure that shapes the genetic makeup of viral pathogens. Indeed, viruses evolved counter-defense mechanisms to evade, suppress or inactivate host immunity. Studying these mechanisms provides important insight in the critical steps of antiviral responses and may uncover novel components and regulators of immune pathways.
Plants, fungi, and invertebrate animals rely on the RNA interference (RNAi) pathway for antiviral defense [1], [2]. The initial trigger of an antiviral RNAi response is the recognition and cleavage of viral double-stranded RNA (dsRNA) into viral small interfering RNAs (vsiRNAs), in insects by the ribonuclease Dicer-2 (Dcr-2). These vsiRNAs act as specificity determinants of the Argonaute-2 (AGO2) containing effector nuclease complex RISC (RNA induced silencing complex). RISC maturation involves a number of sequential steps: loading of the vsiRNA into AGO2, cleavage and elimination of the passenger RNA strand, and 2′-O-methylation of the 3′-terminal nucleotide of the retained guide strand. It is thought that vsiRNA-loaded RISC subsequently cleaves viral target RNA (Slicer activity). The hypersensitivity to viral infections of AGO2 mutant flies and of AGO2 knockdown mosquitoes provides genetic support for this hypothesis [3]–[7]. Nevertheless, direct evidence supporting this model, for example by the detection of viral Slicer products, is lacking.
The evolution of viral suppressors of RNAi (VSRs) is a testament to the antiviral potential of the RNAi pathway in plants and insects. Given the central role of dsRNA and siRNAs as initiators and specificity determinants of the RNAi pathway, it is not surprising that many VSRs sequester dsRNA. For instance, the Drosophila C virus (DCV) 1A protein binds long dsRNA and shields it from processing by Dcr-2 [6]. Flock House virus (FHV) B2 displays a dual RNA binding activity: it binds long dsRNA as well as siRNAs, thereby preventing their incorporation into RISC [8]–[10]. Similarly, many plant VSRs display dsRNA binding activities, leading to the hypothesis that dsRNA or siRNA binding is a general mechanism for RNAi suppression [11], [12]. Nevertheless, other mechanisms have been reported [1]. The RNAi suppressive activity of the Cricket paralysis virus (CrPV) 1A protein, for example, relies on a direct interaction with AGO2 [13].
VSRs have been identified in dozens of plant viruses from all major virus families [1]. In contrast, VSRs have thus far been identified in only three insect RNA viruses (FHV, CrPV, and DCV). These VSRs were characterized using genetic and biochemical approaches in the model organism Drosophila melanogaster. While these viruses indeed efficiently infect Drosophila laboratory stocks and cell lines, DCV is the only natural Drosophila pathogen among these three viruses [14], [15]. Although FHV and CrPV have a remarkable broad host range in the laboratory, they were originally isolated from non-Drosophilid host species: the New Zealand grass grub (Costelytra zealandica) and field crickets (Teleogryllus oceanicus and T. commodus), respectively [16]–[19].
Since viral counter-defense mechanisms co-evolve with the antiviral immune responses of the host species, it is essential to characterize a VSR within the correct evolutionary context. We therefore set out to identify an RNAi suppressor in Nora virus, a positive sense (+) RNA virus that persistently infects Drosophila laboratory stocks as well as Drosophila in the wild [20] (D.J. Obbard, personal communication). The genome organization and phylogeny suggest that Nora virus is the type member of a novel virus family within the order of Picornavirales [20]. Here we show that Nora virus VP1, the protein product of open reading frame 1 (ORF1), suppresses RNAi in cell culture as well as in flies. In accordance, VP1 is an RNAi-dependent viral pathogenicity factor. In a series of biochemical assays, we show that both Nora virus VP1 as well as CrPV 1A inhibit Slicer activity of a pre-assembled RISC loaded with a methylated guide strand. The lack of amino acid sequence similarity between CrPV 1A and Nora virus VP1 suggests that their Slicer antagonistic activities resulted from convergent evolution, providing direct support for the critical role of AGO2 Slicer activity in antiviral defense.
Nora virus is an enteric (+) RNA virus that successfully establishes a persistent infection in flies [20]. The mechanism by which this virus establishes persistent infections is unknown. To determine whether Nora virus is a target for Dcr-2, we analyzed the presence of Nora virus small RNAs in the w1118 Drosophila strain that is widely used as a recipient strain for transgenesis. We isolated and sequenced 19–29 nt small RNAs from body (abdomen and thorax), thorax and head of adult w1118 flies. Sequence reads that perfectly matched the Drosophila genome were annotated and discarded. Of the remaining reads, 396.646 (7,8%, body), 237.265 (10,6%, thorax), and 1.099.496 (7,7%, head) matched the published Nora virus sequence (NC_007919.3), indicating that the w1118 strain was infected by Nora virus (Table 1). As RNA viruses rapidly evolve, viral small RNA sequences may have been missed in this initial matching step. We therefore reconstituted the Nora virus genome through an iterative alignment/consensus treatment of the viral small RNA sequences in our libraries [21]. The reconstituted Nora virus genome (rNora virus) differed at only 3.2% of the nucleotides from the published genome sequence. Aligning small RNAs to the rNora virus genome instead of the published Nora virus sequence resulted in an increased number of viral reads in the three libraries (∼121%, Table 1). We therefore used the reconstituted genome as a reference genome in further analyses.
In all three libraries, Nora virus-derived small RNAs were predominantly 21-nt long, the typical size of Dicer-2 products. The size distribution of small RNAs derived from the (+) RNA strand, however, were noticeably wider than those derived from the (−) RNA strand (Figure 1A). For 21-nt viral RNA reads, there was only a slight bias towards (+) small RNAs (ratio (+) RNA/total RNA ∼0.58), whereas small RNAs of other sizes were predominantly derived from the (+) strand (Figure 1B). In all three libraries, the 21-nt Nora virus-derived RNAs are distributed across the genome, covering both the (+) and (−) viral RNA strands with approximately equal numbers (Figure 1C). These data suggest that dsRNA replication intermediates of Nora virus are processed into 21-nt long siRNAs. The origin of the other size classes of viral small RNAs remains unclear. However, as the predominance of (+) over (−) small RNA reads is reminiscent of the excess of (+) over (−) viral (full-length) RNA that is typically observed in (+) RNA virus infection, they may be due to non-specific RNA degradation.
Drosophila Dcr-2 generates 21-nt duplex siRNAs in which 19 nucleotides are base-paired leaving a 2-nt 3′ overhang at each end. For each library, we collected the 21-nt RNA reads whose 5′ ends overlapped with another 21-nt RNA read on the opposite strand of the Nora virus genome. Then, for each possible overlap of 1 to 21-nt, the numbers of read pairs were counted and converted into Z-scores (Figure 1D). This analysis revealed that 21-nt Nora virus-derived RNAs in body and thorax libraries tend to overlap by 19-nt, which is a typical feature of siRNA duplex precursors. This siRNA duplex signature was observed to a lesser extent in head libraries. Very little Nora virus RNA can be detected in the head [22], yet vsiRNA levels were similar in head, thorax, and body (Table 1). The origin of the vsiRNAs in the head and the reason for the less pronounced vsiRNA signature of those small RNAs remain unclear. Altogether, our results strongly suggest that Nora virus double-stranded replication intermediates are processed by Dcr-2 into vsiRNAs that trigger an RNAi response in infected flies.
Our small RNA profiles indicate that Nora virus is targeted by Dcr-2. Nevertheless, the virus efficiently establishes a persistent infection, suggesting that it is able to evade or suppress the antiviral RNAi response. The Nora virus genome contains four open reading frames (ORFs) (Figure 2A). Nora virus ORF2 is predicted to encode the helicase, protease, and polymerase domains that together form a picornavirus-like replication cassette. ORF4 encodes three proteins that make up the Nora virus capsid (VP4A, VP4B, and VP4C) [23]. To determine whether the Nora virus genome encodes an RNAi suppressor, we analyzed the four ORFs in an RNAi sensor assay in Drosophila cell culture (Figure 2B–2D). In this assay, S2 cells are transfected with firefly (Fluc) and Renilla luciferase (RLuc) reporter plasmids and a plasmid that expresses one of the four viral ORFs. Subsequently, Fluc expression is silenced using specific dsRNA, and Fluc and Rluc activity is monitored. As expected, DCV 1A, a well characterized VSR that binds long dsRNA, efficiently suppressed RNAi, whereas the inactive DCV 1A K73A mutant was unable to do so (Figure 2C and [6]). Cotransfection of the ORF1 expression plasmid also resulted in de-repression of Fluc, suggesting that VP1, the protein product of ORF1, is a suppressor of RNAi. Expression of ORF3 and ORF4 did not affect Fluc activity (Figure 2C). However, since expression of ORF2 and the production of mature capsid proteins from ORF4 were not detectable on western blot, we cannot exclude the possibility that these protein products are able to suppress RNAi as well (Figure 2B).
Next, we tested whether VP1 inhibits the production of siRNAs by Dcr-2 or a subsequent step in the RNAi pathway. To this end, we repeated the RNAi sensor assay using a synthetic siRNA that does not require Dcr-2 cleavage for its silencing activity. Also under these conditions, Nora virus VP1 suppressed silencing of the Fluc reporter. Furthermore, VP1 suppressed RNAi to a similar extent as CrPV 1A, which was previously shown to suppress the effector stage of the RNAi machinery [13] (Figure 2D).
In Drosophila, the microRNA (miRNA) and siRNA pathways are separate processes, with Dcr-1 and AGO1 dedicated to the miRNA pathway and Dcr-2 and AGO2 to the siRNA pathway. Nevertheless, crosstalk between the miRNA and RNAi pathways occurs. Using miRNA sensor assays in S2 cells, in which Fluc expression is silenced by endogenous miRNAs or co-expressed primary miRNAs, we observed that VP1 does not suppress miRNA activity (Text S1 and Figure S1). Together, these data indicate that VP1 is able to suppress the RNAi, but not the miRNA pathway, at a step after dsRNA processing by Dcr-2.
VP1 is highly conserved among different Nora virus isolates (Figure S2). We were unable to predict a protein domain in VP1 suggestive of a mechanism of action. Furthermore, we did not obtain a significant alignment to any other protein from the non-redundant protein sequence database. To map the VSR region of VP1, we generated a series of N- and C-terminal (ΔN and ΔC) truncations and tested them in the RNAi reporter assay in S2 cells (Figures 3A and S3). With the exception of the VP1ΔN390 and VP1ΔN418 mutants, in which no protein could be detected on Western blot, all VP1ΔN and VP1ΔC constructs produced proteins of the expected size (Figure 3B). Deletion of 74 amino acids (aa) or more from the C-terminus of VP1 resulted in loss of suppressor activity (Figure 3C). This suggests that the active domain of VP1 resides in its C-terminal region. Indeed, deleting up to 351 aa from the N-terminus (VP1ΔN351), out of a total of 475 aa, did not affect VSR activity. These results show that the RNAi suppressor activity of VP1 maps to the C-terminal 124 aa.
We next evaluated the VSR activity of Nora virus VP1 in vivo using transgenic flies in which thread (th), also known as Drosophila inhibitor of apoptosis 1, can be silenced by expression of dsRNA targeting this gene (thRNAi [24], [25]) (Figure 4). Eye-specific expression of thRNAi using the GMR-Gal4 driver leads to severe apoptosis in the developing eye. As a consequence, thRNAi flies display a reduced eye size, loss of eye pigmentation, and roughening of the eye surface (Figure 4A, results are shown for AGO2321 heterozygotes; thRNAi in a wildtype background shows the same phenotype, data not shown and ref. 24). Silencing of th in the eye of thRNAi flies is fully dependent on the RNAi pathway, since the phenotype is lost in an AGO2 null mutant background (Figure 4B). These results indicate that the thRNAi sensor fly is a robust system to monitor RNAi activity in vivo.
Consistent with its RNAi suppressive activity in cell culture, expression of full-length VP1 (VP1FL) in thRNAi flies resulted in eyes with a normal size and a rescue of the rough eye phenotype (Figure 4C). The phenotype of thRNAi flies expressing the VP1ΔC74 mutant was similar to that of flies expressing GFP as a negative control, confirming that this mutant is functionally inactive (Figure 4D, E). Notably, while VP1 only partially rescued the RNAi-dependent phenotype, CrPV 1A fully reverted the thRNAi induced phenotype (Figure 4F). Whether this difference is due to a more robust RNAi suppressive activity of CrPV 1A or to a difference in expression level remains to be established.
Having established that VP1 displays RNAi suppressive activity in vitro and in vivo, we next analyzed the effect of VP1 on viral pathogenicity in adult flies. To this end, we generated recombinant Sindbis virus (SINV) expressing the functional VP1ΔN351 (SINV-VP1) or GFP (SINV-GFP) from a second subgenomic promoter (Figure 5A). Although arboviruses are a target of the RNAi pathway during infection in insects [3], [5], [26], we and others have not detected VSR activity in infections with SINV and the related alphavirus Semliki Forest virus [27], [28] (data not shown). Indeed, SINV recombinants expressing the viral RNAi suppressors FHV B2 and CrPV 1A were significantly more pathogenic than their controls in mosquitoes and Drosophila, respectively [13], [27].
We injected wildtype w1118 flies with the SINV recombinants and monitored survival over time. SINV-GFP (and the parental SINV virus, data not shown) induced only modest mortality in these flies with a fully functional RNAi response. After 36 days of infection, 73% of the SINV-GFP infected flies and all mock infected flies were still alive. In contrast, SINV-VP1 infection resulted in more severe mortality. SINV-VP1 infected flies died faster and only 9% of the flies survived the 36-days follow up period (Figure 5B). Although these results indicate that VP1 enhances viral pathogenicity, they fail to show that this effect depends on its VSR activity. Viral proteins are often multifunctional and the effect of VP1 on the course of infection might be attributed to another, as yet unknown, activity of VP1. We therefore performed recombinant SINV infections in RNAi deficient Dcr-2 mutant flies. In this genetic background, an RNAi suppressor is not expected to enhance pathogenicity of the virus. Upon infection with SINV-GFP, the Dcr-2 mutants died much faster than wild-type flies, confirming that SINV is indeed a target of the RNAi pathway. In contrast to infections in RNAi competent flies, the course of infection of SINV-VP1 and SINV-GFP was remarkably similar in Dcr-2 mutants, with 100% mortality at 22 days after infection in both cases (Figure 5C). We therefore conclude that VP1 enhances virulence of an RNA virus in vivo through its RNAi suppressive activity.
To further characterize the VSR activity of Nora virus VP1, we next analyzed the activity of VP1 in a series of biochemical assays that monitor individual steps of the RNAi pathway. To this end, we fused the active VP1ΔN284 mutant to the maltose binding protein (MBP-VP1) and purified it from Escherichia coli. We verified that MBP-VP1 fusion proteins are fully functional in VSR assays in S2 cells to exclude the possibility that MBP interferes with VP1 VSR activity (data not shown).
The ability of VP1 to suppress siRNA-initiated RNAi in S2 cells (Figure 2D) suggests that VP1 inhibits a step downstream of siRNA production by Dcr-2. In accordance, recombinant VP1 was unable to bind long dsRNA in gel mobility shift assays and could not interfere with Dcr-2 mediated processing of long dsRNA into siRNAs in S2 cell extract (Figure S4A, B). We next analyzed whether VP1 is able to bind siRNAs in a gel mobility shift assay. As a positive control, we used a fusion protein of MBP and the Rice hoja blanca virus non-structural protein 3 (NS3), which binds duplex siRNAs with high affinity [29]. Whereas NS3 efficiently bound siRNAs in our assays, we were unable to observe a shift in mobility of siRNAs after incubation with VP1, even at the highest concentrations used (Figure 6A).
Since VP1 is incapable of interfering with the initiator phase of the RNAi pathway, we next examined the effect of VP1 on the effector phase of RNAi. For this purpose, we used an in vitro RNA cleavage assay (Slicer assay) in Drosophila embryo extract [30], in which a sequence-specific siRNA triggers cleavage of a target RNA. Since the 5′ cap of the target RNA is radioactively labeled, the 5′ cleavage product can be visualized by autoradiography after separation on a denaturing polyacrylamide gel. Indeed, a cleavage product of the expected size was detected if embryo extract was incubated with a target RNA and a specific siRNA. Specific cleavage products were not generated in the presence of a non-specific control siRNA (Figure 6B, lanes 1 and 2). Recombinant VP1 protein, but not control MBP protein, efficiently inhibited the production of cleavage product (Figure 6B, lanes 3 and 4). We note, however, that a minor fraction of the target RNA is still cleaved in the presence of VP1 (Figure 6B, lane 3). Together, these experiments show that VP1 does not affect the initiator phase of the RNAi pathway, but interferes with RISC activity.
To discriminate between RISC assembly and target RNA cleavage by a pre-assembled RISC complex, we performed Slicer assays under two experimental conditions (Figure 7A). In the first approach, a purified suppressor protein is added 30 minutes before the siRNA, which allows us to analyze the effect of the VSR on both RISC loading and target cleavage. In the second approach, the embryo extract is incubated with siRNAs for 30 minutes before addition of recombinant protein. This second protocol allows a mature RISC to form prior to the addition of a VSR, thereby allowing us to assess the effect of the VSR on slicing only. As CrPV 1A was previously shown to affect the effector phase of the RNAi pathway [13], we generated recombinant GST-CrPV 1A as well as control GST. These proteins were included in our assays.
Using the first protocol, cleavage of the target RNA was suppressed by VP1 (Figure 7B, lane 3). Strikingly, VP1 was also able to inhibit target cleavage when added to an embryo lysate containing pre-loaded RISC (Figure 7B, lane 7). The observed suppression of slicing was VP1 specific, since MBP alone did not inhibit RNA cleavage (lane 4 and 8). Recombinant CrPV 1A also suppressed slicing in both experimental procedures (Figure 7B, lanes 5 and 9).
To determine if VP1 affects the protein stability of AGO2, we incubated the recombinant proteins in Drosophila embryo extract and analyzed endogenous AGO2 protein levels by Western blot. Neither VP1 nor CrPV 1A affected AGO2 protein levels in embryo lysate, indicating that these two proteins do not mediate RNAi suppression through degradation of AGO2 (Figure 7C).
To further confirm the inhibitory effect of VP1 on Slicer activity rather than RISC assembly, we performed Slicer assays using different siRNA guides. During RISC maturation, guide strands in AGO2 are 2′-O-methylated at their 3′ terminal nucleotide by the Drosophila methyltransferase Hen1 [31]. This modification protects AGO2 associated siRNAs from degradation by trimming and tailing events that occur when there is extensive base-pairing of the guide RNA with a target RNA [32]. To overcome a requirement for Hen1, an siRNA bearing a 2′-O-methylated 3′-terminal nucleotide on the guide strand was used in Slicer assays. Similar to the non-methylated siRNA, the methylated siRNA produced a specific cleavage product of the expected size (Figure 7D, lane 2). Both Nora virus VP1 and CrPV 1A inhibited the cleavage activity of RISC that was pre-loaded with the methylated siRNA (Figure 7D, lane 3 and 5). Again, the GST and MBP control proteins were unable to affect Slicer activity (Figure 7D, lane 4 and 6).
After loading of the siRNA as a duplex, AGO2 cleaves the passenger strand which is then degraded by the C3PO nuclease complex [33]. To circumvent canonical loading of RISC, we induced RISC formation with a single-stranded methylated guide RNA. Although less efficient, loading of single-stranded guide strands into AGO2 is possible via a bypass mechanism [34], [35]. Indeed, at high concentrations, methylated single-stranded guide RNA induced specific cleavage of cap-labeled target RNA (Figure 7E, lane 2). Interestingly, single-stranded guide RNA-induced target cleavage was specifically inhibited both by Nora virus VP1 and by CrPV 1A (Figure 7E, lanes 3 and 5). These results indicate that both CrPV 1A and Nora virus VP1 inhibit Slicer activity of mature RISC rather than RISC assembly.
Following maturation, RISC binds, cleaves, and releases complementary target RNA, and returns to a Slicer-competent state. Drosophila RISC is a multiple turnover complex, in which release of the cleaved target RNA is a rate-limiting step that is greatly enhanced by ATP [36]. We therefore analyzed suppression of Slicer activity under ATP-limiting conditions with a 20-fold molar excess of siRNA over target RNA. RISC was loaded in the presence of ATP, after which creatine kinase was inactivated by NEM, and ATP was depleted (−ATP) by addition of hexokinase and glucose (Figure S5). In parallel, ATP levels were restored (+ATP) after NEM treatment by adding back creatine kinase, and omitting hexokinase treatment. As expected, RISC shows a lower cleavage rate in –ATP conditions than in +ATP conditions (Figure 7F, compare lanes 3 and 5 with lanes 8 and 10). Even under –ATP conditions, Nora virus VP1 and CrPV 1A were able to inhibit Slicer activity (Figure 7F, lanes 2 and 4), suggesting that these two VSRs inhibit the catalytic target cleavage by AGO2.
The mechanisms by which RNA viruses evade sterilizing immunity and establish chronic persistent infections remain poorly understood [37]. Nora virus successfully establishes a persistent infection in Drosophila, providing an excellent model to study mechanisms of persistence. We show here that Nora virus is a target of the antiviral RNAi machinery and that it encodes a potent suppressor of RNAi. Of note, Nora virus RNA levels are unaffected by mutations in the RNAi pathway [38]. These observations therefore suggest that dynamic interactions between the antiviral RNAi response and viral counter-defense mechanisms determine viral persistence.
The production of viral siRNAs is a hallmark of an antiviral RNAi response. By detection of Nora virus-derived vsiRNAs in infected fly stocks, we provide direct evidence that Nora virus is a target of Dcr-2. Nora virus vsiRNAs are distributed across the viral genome, with similar amounts derived from the (+) and (−) RNA strands. During (+) RNA virus infection, (+) viral RNA accumulates in large excess over (−) viral RNA (∼50–100 fold). Cleavage of structured RNA elements by Dcr-2 is therefore expected to produce viral small RNAs that mirror this asymmetric distribution. Thus, similar to other RNA viruses, our results imply that Dcr-2 targets the dsRNA intermediates in Nora virus replication [2], [4], [39]–[41].
The current model proposes that the antiviral RNAi response relies on dicing of viral dsRNA and on slicing of viral target RNAs using vsiRNAs as a guide. Genetic analyses support the role of AGO2 in antiviral defense: AGO2 mutants are hypersensitive to a number of RNA virus infections [3]–[7], [42]. Yet, interpretation of this AGO2 phenotype is complicated by other cellular functions of AGO2, such as regulation of cellular gene transcription and control of transposon activity [43]–[45]. An alternative model proposes that dicing of double-stranded replication intermediates plays an important role in latent virus infection [46]. Dicing of an essential replication intermediate by Dicer-2 should theoretically be sufficient to abort a productive virus replication cycle. The convergent evolution of VSRs that suppress the catalytic activity of AGO2 in two distantly related RNA viruses, Nora virus and CrPV, underlines the essential role of AGO2 Slicer activity in antiviral defense, also in persistent infections in vivo. Importantly, these two viruses display a strikingly different course of infection – CrPV causes a lethal infection, whereas Nora virus establishes a non-lethal, persistent infection – suggesting that the interaction between a VSR and the host RNAi machinery is not the main determinant for viral pathogenicity.
Total RNA was extracted from dissected heads, bodies (abdomens and thoraxes) and thoraxes from w1118 male flies using Trizol reagent (Invitrogen), and RNA quality was verified on a Bioanalyzer (Agilent). Small RNAs were then cloned using the DGE-Small RNA Sample Prep Kit and the Small RNA v1.5 Sample Preparation Kit (Illumina) following the manufacturer's instructions. Libraries were sequenced on the Illumina HiSeq platform.
Sequence reads were clipped from 3′ adapters using fastx_clipper (http://hannonlab.cshl.edu/fastx_toolkit/). Reads in which the adapter sequence (CTGTAGGCACCATCAATCGT) could not be detected were discarded. Only the clipped 19–30 nt reads were retained. Sequence reads were first matched against the Drosophila genome (v5.37) using Bowtie (http://bowtie-bio.sourceforge.net/index.shtml). Reads not matching the Drosophila genome were then matched against the published Nora virus sequence (NC_007919.3, isolate Umeå 2007), allowing one mismatch during alignment. Viral small RNAs were then used to reconstitute a small RNA-based consensus genome sequence (rNora virus, JX220408) using Paparazzi [21] with NC_007919.3 as a starting viral reference genome. Distributions of Nora virus small RNA sizes were computed by parsing the Bowtie outputs with a python script (available upon request). Small RNA profiles were generated by collecting the 21-nt reads that matched the rNora virus sequence allowing one mismatch, and their frequency relative to their 5′ position in the rNora virus (+) or (−) genomic strand was plotted in R. siRNA duplex signatures were calculated according to an algorithm developed to calculate overlap in piRNA sequence reads [47], [48]. The distribution of siRNA overlaps was computed by collecting the 21-nt rNora virus RNA reads whose 5′ ends overlapped with another 21-nt read on the opposite strand. For each possible overlap of 1 to 21 nt (i), the number of read pairs (O) was counted and converted to a Z-score with the formula Z(i) = (O(i)-mean(O))/standard deviation (O). Small RNA sequences were deposited to the Sequence Read Archive (SRA) at the National Center for Biotechnology Information (NCBI) under accession number SRA054241.
Drosophila S2 cells were cultured at 25°C in Schneider's medium (Invitrogen) supplemented with 10% heat inactivated fetal calf serum, 50 U/mL penicillin, and 50 µg/mL streptomycin (Invitrogen). DCV was cultured and titered on S2 cells as described previously [6]. For the production of recombinant SINV, the coding sequence of either GFP or the N-terminal V5 epitope tagged VP1ΔN351 was cloned into the XbaI site of the double subgenomic pTE3'2J vector [49]. The resulting plasmids were linearized by XhoI restriction, purified and used as template for in vitro transcription using the mMESSAGE mMACHINE SP6 High Yield Capped RNA Transcription kit (Ambion). In vitro transcribed RNA was purified using the RNeasy kit (Qiagen) and transfected into BHK cells. Viral titers in the supernatant were determined by plaque assay on BHK cells.
RNAi reporter assays were performed as described previously using 25 ng pMT-GL3, 6 ng pMT-Ren, and 25 ng suppressor plasmid per well of a 96-well plate [50]. Plasmids encoding Nora virus cDNA constructs were generated as described in Protocol S1.
Flies were maintained on standard medium at 25°C with a light/dark cycle of 12 hours/12 hours. Fly stocks that were used for Sindbis virus infection and for preparation of embryo lysate were cleared of Wolbachia and endogenous virus infection (see Protocol S1). We used the following fly stocks and alleles: UAS-CrPV 1A [13], [51], AGO2321 [52], Dcr-2L811fsX [53], thRNAi [24], [25]. The coding sequences of the full-length VP1 and the inactive VP1ΔC74 mutant with an N-terminal V5 epitope tag were cloned into the pUAST vector using the SacII and XbaI restriction sites [54]. The resulting plasmids were microinjected into Drosophila w1118 embryos to generate transgenic fly lines (Bestgene Inc). Virus infections of adult female flies were performed as described previously using 5,000 PFU of recombinant SINV [6]. Survival was monitored daily. In vivo RNAi experiments were performed by crossing GMR-Gal4, UAS-thRNAi/CyO virgins (Meyer et al., 2006) with UAS-VSR/TM3 Sb flies. The eye phenotype was monitored in two- to four-day-old male F1 offspring lacking the CyO and TM3 Sb balancers.
The GST and MBP fusion proteins were purified from E. coli as described in Protocol S1. Purified recombinant proteins were dialyzed against dialysis buffer (20 mM Tris-HCl, 0.5 mM EDTA, 5 mM MgCl2, 1 mM DTT, 140 mM NaCl, 2.7 mM KCl) Recombinant proteins were stored as aliquots at −80°C in dialysis buffer containing 30% glycerol.
Gel mobility shift assays were performed as described [6]. Briefly, uniformly radio-labeled 113 nt long dsRNA (50 cps/reaction) or end-labeled siRNAs (200 cps/reaction) were incubated with purified recombinant protein for 30 minutes at room temperature. Samples were then separated on an 8% native polyacrylamide gel and exposed to a Kodak Biomax XAR film.
Dicer and Slicer assays were performed according to the protocol of Haley and colleagues with minor modifications, described in Protocol S1 [30]. For Slicer assays with the methylated duplex, Fluc guide strand 5′- UCG AAG UAC UCA GCG UAA GU[mU] and passenger strand 5′- CUU ACG CUG AGU ACU UCG AUU were annealed by incubating 20 µM of each siRNA strand in annealing buffer (100 mM potassium acetate, 30 mM HEPES-KOH at pH 7.4, 2 mM magnesium acetate) for 1 min at 90°C, followed by incubation for 1 hour at 37°C. For guide strand loading of RISC, embryo lysates were incubated with Fluc single-stranded guide strand RNA at a final concentration of 10 µM. Radiolabeled probes and target RNA for gel shift and Slicer assays are described in Protocol S1.
|
10.1371/journal.pcbi.1004983 | Accuracy of Answers to Cell Lineage Questions Depends on Single-Cell Genomics Data Quality and Quantity | Advances in single-cell (SC) genomics enable commensurate improvements in methods for uncovering lineage relations among individual cells, as determined by phylogenetic analysis of the somatic mutations harbored by each cell. Theoretically, complete and accurate knowledge of the genome of each cell of an individual can produce an extremely accurate cell lineage tree of that individual. However, the reality of SC genomics is that such complete and accurate knowledge would be wanting, in quality and in quantity, for the foreseeable future. In this paper we offer a framework for systematically exploring the feasibility of answering cell lineage questions based on SC somatic mutational analysis, as a function of SC genomics data quality and quantity. We take into consideration the current limitations of SC genomics in terms of mutation data quality, most notably amplification bias and allele dropouts (ADO), as well as cost, which puts practical limits on mutation data quantity obtained from each cell as well as on cell sample density. We do so by generating in silico cell lineage trees using a dedicated formal language, eSTG, and show how the ability to answer correctly a cell lineage question depends on the quality and quantity of the SC mutation data. The presented framework can serve as a baseline for the potential of current SC genomics to unravel cell lineage dynamics, as well as the potential contributions of future advancement, both biochemical and computational, for the task.
| A human cell lineage tree describes the entire developmental dynamics of a person starting from the zygote and ending with each and every extant cell. Fundamental open problems in biology and medicine are in fact questions about the human cell lineage tree: its structure and its dynamics in development, growth, renewal, aging, and disease. Consequently, a method to know the human cell lineage tree would allow resolving these problems and enable a leapfrog advance in human knowledge and health. Recent advancements in single-cell genomics have the potential to uncover various properties of the human cell lineage tree and thus promote our understanding of various biological phenomena. In this paper we present a computational framework along with specific results, which enable to understand what can be achieved using the limitations of current technologies and predict future capabilities based on future improvements. This approach can serve as a valuable tool for researchers who plan to perform lineage experiments both in designing and optimizing the actual experimental needs and predicting the costs and limitations of the plan. This work can also help researchers focus on developing what is needed for future advancements.
| Recent advances in SC technologies have generated a unique opportunity to delineate the complex behavior of heterogeneous cell populations and uncover their underlying mechanistic dynamics [1]. The use of SC genomics to reveal cell lineage relationships have been recently demonstrated in various scenarios including diseases such as cancer [2–6] and normal development [7–10]. Lineage analysis of cells sampled from an organism makes use of somatic mutations to discover common history dynamics of the sampled cells. There are several types of somatic mutations that can be used for this task, including Single Nucleotide Variations (SNV) [2, 3, 11–13], Short Tandem Repeats (STR, also called Microsatellites) [6, 8–10, 14–18], Copy Number Variations (CNV) [4, 5, 7], and Transposable Elements (TE) [8] where each type has a different mutational model and different mutation rates. This analysis is mostly effective when analyzing SC since the mixed mutational signal of cell bulks does not allow delineating mutational co-occurrences and cannot distinguish between subpopulations with different mutational patterns. Although published work have shown the great potential of using SC mutational analysis for unraveling cell lineage dynamics, there are still several major limitations, which hamper further generalization of this concept to various biological questions and prevent its use in large scale experiments. These limitations include 1) technical issues related to SC genomics, including the need for DNA amplification that introduces technical noise, 2) lack of high throughput SC isolation techniques, especially if one wants to retain the original 3D structure, or analyze rare cell types that are difficult to isolate, 3) associated costs, such as Whole Genome Amplification (WGA) kits, sequencing costs, and other consumable products (e.g., reagents and microfluidic devices), and 4) lack of computational infrastructure and dedicated algorithms specifically designed for the unique challenges of SC genomics.
The feasibility of using somatic mutations for uncovering cell lineage dynamics is dependent on these issues but also on the specifics of the pursued biological question. Some factors are inherent, such as the mutation rate and number of cell divisions, but others can be overcome by spending more money or by improving biochemical or computational procedures. Using controlled ex-vivo experiments is a close approximation to real biological scenarios; however, it can be very costly and laborious. Furthermore, many scenarios cannot be examined due to technical limitations in trying to mimic real biological dynamics (e.g., cell differentiation leading to changes in cellular dynamics), and also various parameter combinations cannot be studied using an ex-vivo experiment. A computational alternative is to model and simulate various biological scenarios using a range of parameters and conditions. Not only this approach enables to inspect the strengths and weaknesses of existing methods, it can also enable to predict the impact of future improvements.
Until now, there has not been any systematic examination of how much mutational data is required in order to accurately answer questions related to the structure and dynamics of SC lineage trees. In this work we cover few common biological settings, which capture certain tree properties such as depth (corresponds to number of cell divisions) and clustering relationships, in order to systematically evaluate the feasibility of answering cell lineage questions using somatic mutations, and predict future capabilities by extending the range of parameters values to represent future enhancements. Using mutational data from an ex-vivo experiment we estimated and modeled the properties of the mutational signal quality, afflicted mainly by the random noise and ADO caused during the preprocessing and amplification of SC DNA. We then applied this model onto the signal of simulated lineage trees, generated using a dedicated formal language and simulation tool, based on environment-dependent Stochastic Tree Grammars (eSTG) [19], which is capable of generating both the entire modeled cell lineage tree and the corresponding somatic mutations accumulated through cell divisions. We present the results on a variety of parameters values, including different distance relationships (corresponding to different number of cell divisions) between different cell types, different mutation rates and two types of somatic mutations, including STR [20] and SNV. We also take into consideration current estimated costs of biochemical analysis and for each combination of parameters we calculate the cost-optimized number of cell samples and genomic loci that enable to answer the biological question with high confidence. We map the dependency between the quality and quantity of the SC mutational data and the ability to answer cell lineage questions of specific settings, which can be used as a framework for planning cell lineage experiments and predicting the potential of future enhancements, both biochemical and computational.
We have previously presented a formal language, called eSTG, for describing population dynamics [19] and a corresponding programming and simulation environment, called eSTGt (eSTG tool) [21]. The language captures in broad terms the effect of the changing environment while abstracting away details on interaction among individuals. A prominent feature of the tool is that it can stochastically produce lineage trees, each corresponding to a different stochastic program execution. These lineage trees record the entire execution history of the process, including the dynamics that led to existing as well as to extinct individuals. In this paper we simulated cell lineage trees using eSTGt by specifying and executing eSTG programs. The output of each program’s execution is an instance of a stochastic lineage tree, which also includes the corresponding somatic mutations as specified by the eSTG programs. By running multiple executions of the programs we collected sufficient statistics as described below. The program specifications used in this manuscript can be found in S1 File.
As mentioned above there are several types of endogenous somatic mutations, including STR, SNV, CNV and TE. Since CNV and TE have a complex dynamics and are hard to predict and model we decided to focus on STR and SNV, which are the most appropriate candidates for inferring general cell lineages retrospectively. For STR mutations, we used the stepwise mutation model [22], which assigns an equal probability pSTR for either an increase or a decrease of one repeat unit during each cell division (see Methods). Current estimations of the STR mutation rate pSTR range between 10−3–10−5 mutations per locus per cell division depending on various factors such as the STR length, repeat type and the specific cell genotype [20]. The low mutation rate might correspond to short STRs of normal cells whereas the fast mutation rate might correspond to cells harboring Microsatellite Instability, which is common in various types of cancer cells [23]. In order to cover the entire spectrum we chose to simulate three scales of mutation rates, namely, pSTR = 10−3, 10−4, 10−5. SNV mutations were modeled by randomly mutating each base with probability pSNV following each cell division. The mutation rate pSNV is estimated to be between 10−7–10−10 mutations per nucleotide per cell division [24]. Since mutation rate of 10−10 was too low to yield any significant signal we present results only for mutation rates pSNV = 10−7, 10−8, 10−9.
As we mentioned, SC genomics poses many challenges, since the starting material consists of only one copy of each DNA molecule. DNA isolation and amplification introduce technical noise and methods for measuring and reducing it, both biochemically and computationally, are still under extensive research [1]. We chose to model two types of interferences, namely, ADO and random noise. To this end, we used data from an ex-vivo experiment that consisted of clonal expansions from which SCs were sampled and processed. The processing included SC Whole Genome Amplification (WGA) and sequencing of targeted loci. ADO was modeled by taking into consideration both the distribution of samples quality and genomic location, and noise was estimated by comparing the genotype of duplicates, which should be identical (see Methods). After simulating the lineage trees along with their somatic mutations we applied the models of the ADO and noise in order to generate the final mutation table that was used for further analysis. In addition, we also adjusted the parameters of the ADO and noise models in order to predict the performance of future improvements in the processing of SC genomics (see Methods). In the figures below we present results for STR using mutation rate pSTR = 10−4, which may correspond to highly mutable long STR loci of normal cells, for both current and future predicted signals. Results for the other STR mutation rates (10−3, 10−5) and for SNV (with mutation rates 10−7, 10−8, 10−9) for current and future signal quality are presented in the Supplementary Information.
In order to optimize the cost efficiency of a specific analysis, we used a fixed ratio of 1:1000 between the analysis cost of a single cell and the analysis cost of a single STR locus, thus one can tradeoff between the number of cells and the number of loci analyzed, depending on specific constraints such as sample scarcity or sequencing availability. In the examples below we used fixed costs of 10$ for a single cell analysis and 0.01$ for a single STR locus. These costs are based on rough estimations of current processing (e.g., WGA kits and consumables) and sequencing costs (see Methods) and can of course be adjusted as needed.
A triplet tree consists of three leaves sampled from a (full) tree and the subtree they induce on the full tree (Fig 1A). Since there are three possible bifurcation arrangements for the triplet tree, the probability of a random triplet tree reconstruction to correctly reconstruct its topology is 1/3. In order to measure the ability to correctly reconstruct a triplet tree using somatic mutations we simulated such trees with various number of cell divisions along with the corresponding mutational signal, which was distorted with the calibrated ADO and noise. We then measured the percentage of correct reconstructions over 1000 repeated stochastic simulations. Fig 1B shows the percentage of correctly reconstructed triplet trees with various number of cell divisions (X = 2,5,10,20,40, see Fig 1A) as a function of the number of analyzed loci (ranging from 500 to 100,000) using STR mutations with mutation rate 10−4. Fig 1C shows the results that correspond to future signal improvements. It can be seen, for example, that using 5 cell divisions (X = 5) and 25,000 loci the probability of correctly reconstructing a triplet tree is about 50% (compared to 33% for random reconstruction) using the current signal and almost 70% using the predicted future enhancements. Results for the other STR mutation rates and SNV are presented in S2 File.
Many lineage questions are in fact questions about the depth relationship between two cell groups. Examples include questions related both to cancer dynamics and normal development or renewal. For example, is relapse after chemotherapy caused by ordinary tumor cells escaping chemotherapy stochastically, or by a separate population of rarely-dividing cancer-initiating cells that escape chemotherapy due to their slow division rate [6]? If relapse is initiated from slowly dividing cells, these cells would accumulate fewer mutations since they go through fewer cell divisions. By measuring the distance of the cells from the root of the tree (which can be estimated using a combination of unrelated cell bulks) we can compare the depth relationship between different cell groups. Another example question is whether the adult oocyte pool can be renewed during adulthood [10]? Again, by comparing the number of cell divisions between young and adult female, we may know whether oocytes are generated postnatally.
In order to map the feasibility of answering such questions we simulated lineage trees and analyzed two cell groups from different depths in the tree (Fig 2A). For each cell we estimated its relative depth in the tree using its mutational signature and performed a statistical test that compared the relative depth of cells from both groups (see Methods). Fig 2B shows a heatmap that represents the probability of correctly identifying a significant depth difference between the two cell groups, one of depth X and the other of depth X+Y, for X = 40 and Y = 10, as a function of the number of analyzed cells and the number of analyzed genomic loci. It can be seen that in order to obtain a specific success probability one can tradeoff between the number of analyzed samples and the number of analyzed loci (white line in Fig 2B that represents success probability of 95%), however, a minimum cost can be obtained by selecting the combination that corresponds to the minimum of the black line in Fig 2B that shows the corresponding analysis cost. Fig 2C shows a summary of the cost-optimized number of samples and number of loci needed for obtaining success probability of 95% using various combinations of X and Y corresponding to various depths of the two cell groups (as depicted in Fig 2A). Fig 2D and 2E show the performance using enhanced parameters that correspond to future enhancements in SC genomics. Results for the other STR mutation rates and SNV are presented in S2 File.
Identifying the clonal relationship between two cell populations arises in many contexts. For example, do progenitor cells commit to a single cell-type or can they produce multiple types as needed [25]? Does geographic separation imply lineage separation or do cells migrate from one area to another [8]? Are the original tumor and its relapse independent clones [6]? The mutational signature of two cell populations can be used to perform clustering analysis in order to examine whether they are separated or intermixed in the lineage tree.
In order to investigate how well can phylogenetic analysis of somatic mutations be used for answering such questions we simulated lineage trees consisting of two subclones, which have a common ancestor of a specific distance (Fig 3A). We then estimated the distance within and between the two cell groups and performed a statistical test to check whether the two cell groups are separated (see Methods). Fig 3B shows a similar heatmap to Fig 2B but presents the probability of identifying that the two cell groups are independent, using X = 2 and Y = 20 (see Fig 2A). Fig 3C presents the cost-optimized combinations for various values of X and Y. Fig 3D and 3E show the performance using enhanced parameters that correspond to future enhancements. Results for the other STR mutation rates and SNV are presented in S2 File.
During normal mitotic cell division DNA is replicated with very high, but not absolute, precision, which leads to the incorporation of somatic mutations. These somatic mutations accumulated since the zygotic stage, endow each cell in our bodies with a genomic signature that is unique with a very high probability [17]. Sequencing cell bulks for somatic mutations may supply a coarse estimation of the cell population distribution but cannot specify the deterministic position in the lineage tree of each cell and uncover population heterogeneity. Advancements in single cell genomics offer a unique opportunity to detect somatic mutations private to each cell and use them to understand the underlying dynamics of cell lineages with high precision. Unfortunately, sequencing accurately the entire genome of each single cell is still prohibitively expensive and technically challenging. In recent years there have been several attempts to use single cells genomic data in order to uncover various lineage dynamics. These attempts included SC whole genome sequencing [26], exome sequencing [5], and genotyping of targeted loci [6], or combinations of thereof [2]. There is a tradeoff between genomic coverage and sample density and the question of finding their right quantity and balance depends on parameters such as cost, technical constraints and the specifics of the lineage question. In this paper we offer a framework for answering this question by modeling and simulating the entire process of lineage analysis taking into consideration the different aspects of SC genomics analysis, calibrated using real experiments, and possible lineage dynamics. The suggested framework can help researchers in planning and optimizing their lineage experiments and can also point out experimental aspects that should be improved in order to increase the chances for meaningful outcomes. We selected a basic triplet tree structure and two aspects of lineage questions that are widely tackled, namely identifying depth differences and identifying independent clusters, and mapped the feasibility of answering them using a wide variety of parameters, including different mutation types, different mutation rates and various combinations of distances between the cell groups. The results can serve as a guideline for planning a lineage experiment or as a reference point for tailoring a solution for a more specific setting. Future experiments can help in fine-tuning the different modeling aspects, such as ADO, noise and possible lineage scenarios. Furthermore, these aspects can also be updated as new and more advanced biochemical protocols, technological or computational tools are developed.
STR mutations were modeled using the single-step model (SSM) [22]. For each STR loci of length x, its length is updated during each cell division using the following function:
fSTR(x)={x+1withprobabilitypSTR2x−1withprobabilitypSTR2xotherwise
where pSTR is the mutation probability. In this paper we used three mutation scales, namely 10−3, 10−4, 10−5, corresponding to possible STR mutations rates. We note that some STRs can display more complex mutational patterns; however, the SSM is the most common model used and constitutes a good approximation.
SNV mutations were modeled by randomly changing each base with probability pSNV during each cell division, where we used three mutations scales, 10−7, 10−8, 10−9.
We note that most chromosomes, except for the X and Y chromosomes in males, have two copies. This may introduce additional complexity to the analysis of SC genomic loci since a mutation can occur in one copy or the other. However, for MS loci this can be overcome by analyzing only sex chromosomes of males [6, 9, 10] or by analyzing loci with heterogeneous alleles that contain MS with different repeat number [27]. As for SNV analysis, the probability of a double mutation is low enough in order to allow a unique identification of random somatic mutations in each locus. Since the ex-vivo experimental data that we used in order to model the ADO and the noise of the SC genomic signal included mostly data from the X chromosome, we opted to analyze the simplified single allele scenario in this work. However, we are currently working on computational methods for analyzing biallelic signal, which will allow analyzing signal from autosomes and will also enable to extend the results presented here for more complex scenarios.
Since a human cell contains only one copy of a diploid genome there is a big chance that some parts of the DNA will be damaged or lost during the different amplification stages. Because of the stochastic nature of the amplification, one could also expect a relatively large variability in the amplification quality of different samples. In addition, there could be amplification biases where some parts of the genome are better amplified than others, resulting in some loci having a better chance to be detected. In order to simulate the dropout patterning of the experimental data we sought to find a modeling approach that will mimic the real behavior as much as possible. The experimental data evidently show that the allelic dropout is not random but is dependent on both the sample quality and the genomic location. In order to capture the variability of the signal quality in both the individual samples and the different loci we modeled the allelic dropout of single cell DNA samples by assigning distinct dropout probabilities for each sample and for each locus. Given M individual samples and N loci we define the mutation table T = {tij: i = 1.M, j = 1.N} such that tij equals the mutation call of sample i at locus j. In the case of allelic dropout we set tij = ∅. We define the mutation signal table as X = {xij: i = 1.M, j = 1.N}, where
xij={0iftij=∅1otherwise
We define P = (pi: i = 1.M) as the probability of obtaining a signal in each sample and Q = (qj: j = 1.N) as the probability of obtaining a signal in each locus. The probability of obtaining a signal in sample i and locus j thus equals piqj.
In order to estimate these probabilities using the real ex-vivo data, we used a Maximum Likelihood (ML) approach. Given the mutation signal table data X = {xij}, the log likelihood is:
logL(P,Q;X)∝logP(X|P,Q)=∑i=1M∑j=1Nlog(xijpiqj+(1−xij)(1−piqj))
The ML estimator of P and Q is thus:
argmaxP,Q(log(L(P,Q;X)))
We approximated the solution using simulated annealing and validated the results by repeating the procedure with various starting points. For the data X we used an ex-vivo experiment in which 167 single cells were amplified and analyzed for their genomic signal. For prediction of future enhancement we used the calculated probabilities p,q and increased their relative value by 25%.
Noise modeling differs between STR and SNV because STRs are much more prone to errors introduced during the amplification stages. For STR mutations we defined noise as the probability for each locus to randomly shift by one repeat unit compared to its true value. In order to estimate this probability we used the analysis results of duplicate cells from an ex-vivo experiment and measured the rate of inconsistency between supposedly identical genomes.
For SNV mutations we set the probability for noise to be 10−4 as measured using SC calling results of next-generation sequencing data [28].
For prediction of future enhancement we used the noise probability value divided by 2.
We have divided the analysis cost into two parts, namely, the overhead of analyzing a single cell and the analysis cost per single locus. A detailed cost analysis is not presented in this manuscript, however, an approximation for a complete analysis of a single cell is 30$, from which 10$ are considered to be fixed overhead and 20$ are used for analyzing either 2000 STR loci or 20,000 single bases. We thus approximated the analysis cost of a single STR locus to be 20/2000 = 0.01$ and the analysis cost of a single base (SNV) to be 20/20,000 = 0.0001$.
In order to calculate the cost as presented in Figs 2 and 3 we used the following function:
fCost=CostLoc*x+CostSamp*y
where CostLoc = 0.01 for STR and 0.0001 for SNV, CostSamp = 10, x = # of loci, y = # of samples and x,y are constrained to the white line in Figs 2 and 3 (corresponding to success probability of 95%). Minimal cost is obtained by finding the minimum of fCost.
For the triplet trees reconstruction we used the Neighbor-Joining (NJ) algorithm [29] with the absolute distance function:
Given a mutation table
T={Til;i=1..M,l=1..N}
, with M samples and N loci, where
Til
is the genotyping of locus l in sample i, the distance between each two samples is:
D(i,j)=1N∑l=1N|Til−Tjl|
where only loci with signal in both samples are counted. For the three example samples with the following 5 loci genotype:
T1=(10,∅,∅,8,12)
T2=(12,∅,7,8,∅)
T3=(10,∅,7,8,11)
where ∅ means that there is no signal in that locus, the distances are:
D(1,2)=12(|10−11|+|8−8|)=12(1+0)=12
D(1,3)=12(|10−12|+|8−8|)=12(2+0)=1
D(2,3)=13(|12−10|+|7−7|+|8−8|)=13(2+0+0)=23
The result of the NJ tree reconstruction algorithm on these samples is depicted in S1 Fig.
We note that alternatives to distance-based methods for phylogeny estimation exist, which might yield better results or improve the cost efficiency; however, analyzing or developing such methods is not in the scope of this paper and is a subject of an ongoing research in our lab.
Given two groups of cells A = {ai} and B = {bj} we define a binary classifier f that decides whether there is a depth difference between them or not. We define D(x) as the distance between the cell x and the root of the tree where D is calculated using the absolute distance function as defined above. We define the set D(X) = {D(x)}x∈X where X is a group of cells. We define ttest(D(A),D(B)) as the p-value obtained from a t-test between the set of distances D(A) and D(B). The classifier f is defined as follows:
f(A,B)={1ifttest(D(A),D(B))≤0.050otherwise
where f(A,B) = 1 means that there is a significant distance between the cell groups A and B.
In the words of hypothesis testing, if we define the null hypothesis to be that there is no depth difference between A and B then from the definition of f if the depth of the two populations is equally distributed the probability of incorrectly rejecting the null hypothesis, i.e., the type I error, is 5% and the statistical power is depicted in Fig 2B.
Similarly to the case of depth differences, we define D(x,y) as the distance between cell x and cell y, and D(X,Y) = {D(x,y)}x∈X,y∈Y. We define the clustering classifier f to be:
f(A,B)={1ifttest(D(A,A),D(A,B))≤0.050otherwise
i.e., we measure the difference in the average distance of cells within the group A and the distance of cells between group A and group B. Similarly to the case of the depth differences, the type I error is 5% and the statistical power is depicted in Fig 3B.
We note that the measures presented here for identifying significant depth differences and clustering are used for proof of concept and there may be better ones. However, finding better measures is not in the scope of this paper and is a subject of future research.
|
10.1371/journal.pbio.1002402 | MamO Is a Repurposed Serine Protease that Promotes Magnetite Biomineralization through Direct Transition Metal Binding in Magnetotactic Bacteria | Many living organisms transform inorganic atoms into highly ordered crystalline materials. An elegant example of such biomineralization processes is the production of nano-scale magnetic crystals in magnetotactic bacteria. Previous studies implicated the involvement of two putative serine proteases, MamE and MamO, during the early stages of magnetite formation in Magnetospirillum magneticum AMB-1. Here, using genetic analysis and X-ray crystallography, we show that MamO has a degenerate active site, rendering it incapable of protease activity. Instead, MamO promotes magnetosome formation through two genetically distinct, noncatalytic activities: activation of MamE-dependent proteolysis of biomineralization factors and direct binding to transition metal ions. By solving the structure of the protease domain bound to a metal ion, we identify a surface-exposed di-histidine motif in MamO that contributes to metal binding and show that it is required to initiate biomineralization in vivo. Finally, we find that pseudoproteases are widespread in magnetotactic bacteria and that they have evolved independently in three separate taxa. Our results highlight the versatility of protein scaffolds in accommodating new biochemical activities and provide unprecedented insight into the earliest stages of biomineralization.
| Biomineralization is an ancient and ubiquitous process by which organisms assemble crystalline materials for their own benefit. The ability to precisely organize inorganic atoms into crystals with intricate shapes demonstrates a level of control over nanoparticle synthesis that has fascinated biologists for generations. We have been studying how a group of microorganisms, called magnetotactic bacteria, synthesizes iron-based crystals that are used for navigation along magnetic fields. Here, we characterize a protein called MamO that helps to initiate the formation of a magnetic mineral called magnetite in cells of the magnetotactic bacterium Magentospirillum magneticum AMB-1. Although predicted to be a trypsin-like protease, we show that MamO has lost its ancestral catalytic activity and instead gained a new function as a metal-binding scaffold. By solving its structure, we discovered how MamO binds to transition metal atoms and show that this activity is required to crystalize magnetite within cells. Surprisingly, we find that similar repurposed trypsin-like proteases have evolved independently in all three major magnetotactic groups, outlining a fascinating case of convergent evolution. The unique evolutionary history of MamO demonstrates that existing protein scaffolds can be modified to provide new functions and contributes to our understanding of how cells build transition metal-based minerals.
| Biomineralization is the widespread phenomenon by which living organisms transform inorganic atoms into highly ordered, crystalline structures. Controlling the size and shape of such materials requires specialized protein machinery that can define the nano-scale trajectory of crystal growth [1]. Incorporating biochemical principles uncovered from studying biomineralization has the potential to revolutionize the design and synthesis of nanomaterials in vitro [2]. In addition to the well-known examples of tooth, bone, and shell production by multicellular eukaryotes, a number of bacteria have the ability to biomineralize small magnetic crystals within subcellular compartments called magnetosomes [3,4]. These particles allow the cells to passively align in the earth’s magnetic field, facilitating the search for their preferred oxygen environments [5]. Although these magnetotactic bacteria have drawn longstanding interest due to their ability to manipulate transition metals, the biochemical details of how they transform iron into magnetite (Fe3O4) remain poorly understood.
Magnetotactic organisms are phylogentically diverse. Nearly all isolates come from the α-, δ-, or γ- classes of Proteobacteria, but representatives from the Nitrospirae and Omnitrophica phyla have recently been identified [6]. The genes responsible for making magnetosomes are often contained in a genomic region called the magnetosome island (MAI) [7–11]. Comparative genomic and phylogenetic studies have identified a set of core genes that appears to have been assembled a single time and inherited vertically, indicating that magnetosome formation likely predates the divergence of the Proteobacteria [12,13]. The MAI seems to have formed by incorporating elements from other cellular processes, as the majority of the core factors have homology to ancient and widespread protein domains [14,15]. Uncovering the biochemical functions encoded in the MAI in relation to its evolutionary history provides a unique opportunity to understand how new cellular processes evolve.
Due to the availability of genetic systems, α-Proteobacteria such as Magnetospirillum magneticum AMB-1 are used as models for studying the molecular biology of magnetosome formation [16]. AMB-1 contains 15–20 magnetite crystals, each formed within a cytoplasmic membrane invagination and organized in a chain spanning the length of the cell [17,18]. By making deletions within the MAI and characterizing the ultrastructure of the mutant cells, specific genes have been assigned roles in various stages of magnetosome formation [10,11,19,20]. Genes whose deletions produced empty magnetosome compartments or compartments with abnormally small magnetite crystals were termed biomineralization factors. It is the proteins encoded by these genes that hold the secrets of how magnetotactic organisms interface with solid magnetite.
Two genes in the MAI, mamE and mamO, are homologous to the HtrA proteases, a ubiquitous family of trypsin-like enzymes that functions using His-Asp-Ser catalytic triads [21]. An additional pair of genes, called limE and limO, with homology to the protease domains of mamE and mamO, exists in a secondary genomic region termed R9 [10]. Disrupting mamE or mamO causes cells to produce empty magnetosome membranes, but removing R9 has no effect, showing that mamE and mamO are required for biomineralization and that limE and limO are not (Fig 1) [10,22]. Adding variants of mamE or mamO with all three predicted active site residues mutated to alanine could not restore normal biomineralization in the ΔmamOΔR9 or ΔmamEΔR9 strains but complemented single ΔmamO or ΔmamE deletions [23]. These genetic analyses show that mamE and mamO are required for the initiation of magnetite biomineralization. Furthermore, limE and limO are partially redundant in that they can cross-complement the active site-dependent crystal maturation defects of their respective orthologs.
Here, we use a combination of in vivo and in vitro approaches to reveal an unexpected dual role for MamO. It promotes MamE-dependent proteolysis of three biomineralization factors through the use of its C-terminal transporter domain. Separately, the protease domain has lost the ability to carry out catalysis and has instead been repurposed to bind transition metal ions. Two surface-exposed histidine residues that contribute to this metal-binding function are required for initiating magnetite biosynthesis in vivo. Bioinformatic analysis shows that similar pseudoproteases evolved independently in the three major taxa of magnetotactic organisms, highlighting a unique evolutionary mechanism behind microbial nanoparticle synthesis.
Trypsin-like proteases utilize a histidine–aspartate pair to deprotonate the hydroxyl group of a serine residue, providing a nucleophile for catalysis. MamO is unusual in that it contains a threonine instead of serine as the predicted nucleophile. To further clarify the role of MamO’s unusual active site, we focused our initial efforts on assessing the contribution of each putative catalytic residue to biomineralization. We performed these initial studies using the ΔmamOΔR9 strain (referred to as ΔOΔR9) to avoid cross-complementation from limO. Consistent with our previous findings, individual H116A and D149A mutations in MamO severely reduced the cells’ ability to turn in a magnetic field (Fig 2A). Surprisingly, a T225A mutation had no defects in magnetic response (Fig 2A). Using transmission electron microscopy (TEM), we confirmed that mamOH116A and mamOD149A cells have small magnetite crystals while mamOT225A crystals are indistinguishable from wild-type, mirroring the bulk magnetic response measurements (Fig 2B–2D).
Given that the predicted nucleophile is dispensable for magnetosome formation, we found it curious that the other two catalytic triad mutations disrupted crystal maturation. Upon further examination, we found that the phenotypes associated with the mamOH116A and mamOD149A alleles were actually temperature-dependent. Growing wild-type AMB-1 at room temperature instead of the standard 30°C did not dramatically alter the magnetic response. However, both the mamOH116A and mamOD149A alleles displayed improved complementation at the lower temperature. In particular, the mamOD149A mutant restored a nearly wild-type magnetic response to the ΔOΔR9 strain under these conditions (S1 Fig). Our results suggest that protease activity from MamO is not required for biomineralization. None of the three predicted catalytic residues is required for magnetite nucleation, and, although two of the three contribute to crystal maturation, these effects are conditional, suggesting that they are not central to the biomineralization process.
During the course of our experiments, we examined the MamO variants by western blotting. Although no changes in overall protein abundance were present, we noticed that each MamO variant was proteolytically processed, having both a full-length and shorter form (S2 Fig). This finding led us to examine whether other magnetosome proteins are similarly proteolyzed. We found that MamE and another biomineralization factor, MamP, are also proteolytic targets in AMB-1 cells by using antibodies targeted to each protein (Fig 3A). MamP is a c-type cytochrome, and its iron oxidase activity is required for the proper maturation of magnetite crystals [24,25]. Since these three biomineralization factors exist in both full-length and shorter forms, it is likely that proteolytic maturation plays a role in their function.
To examine the potential involvement of mamE and mamO in promoting these proteolytic events, we analyzed processing of MamO, MamE, and MamP in various genetic backgrounds. In both the ΔOΔR9 and ΔEΔlimE strains, we observed only the full-length form of MamP. Similar analyses showed that MamO is required for the proteolysis of MamE and that MamE is required for the proteolysis of MamO. Thus, processing of each target requires both putative proteases (Figs 3A and S2A). Addition of the mamEWT allele restored processing of each target in the ΔEΔlimE strain, but the catalytically inactive form (mamEPD) did not, suggesting that MamE participates in proteolysis directly. Surprisingly, both the mamOWT and mamOH116A alleles restored processing in the ΔOΔR9 strain (Figs 3A and S2). Therefore, the presence of MamO, but not its catalytic triad, appears to be required to promote the activity of MamE. While these results strongly suggest that MamE directly cleaves all three biomineralization factors, we cannot rule out the possibility that its activity contributes to a more complex targeting process.
Given that HtrA proteases are often regulated by the formation of higher order oligomers, an attractive model could be that MamO activates MamE through an interaction involving both protease domains. To test this idea, we exploited the partial redundancy from genes in the R9 region. In addition to its N-terminal protease domain, MamO has a predicted TauE-like transporter domain on its C-terminus. limO, the partial duplication of mamO in R9, is 98% identical to the protease domain of mamO but does not contain a C-terminal TauE domain (S3 Fig). We confirmed that alleles with individual point mutations in the protease domain that reduce bioimineralization in the ΔOΔR9 background do not have defects in the ΔmamO strain, reinforcing that LimO is a functionally redundant copy of the MamO protease domain (S3 Fig). In contrast, processing of MamE and MamP is disrupted in both the ΔmamO and ΔOΔR9 strains, showing that a functional protease domain is insufficient to activate MamE and that activation requires the TauE domain (Fig 3B).
In the trypsin family, loop L1 contains both the nucleophilic serine and oxyanion hole, required for creating the acyl-enzyme intermediate and stabilizing the oxyanion, respectively (S4 Fig) [26]. In addition to the threonine substitution, the entire L1 loop in MamO differs significantly from the trypsin family consensus, once again suggesting that MamO might not be capable of protease activity (S4 Fig). To explore this possibility further, we determined the crystal structure of the protease domain (S1 Table). It crystallizes as a monomer with the chymotrypsin fold and the catalytic residues properly placed (S4 Fig).
Loop L1 of MamO adopts the inactive conformation seen in many HtrA proteases in which the main chain carbonyl of residue 192 (chymotrypsin numbering; W222 in MamO) prevents access to the oxyanion hole (Fig 4A). In other HtrAs, the inactive form is thought to be in equilibrium with an active conformation in which the main chain flips approximately 180°, opening the oxyanion hole (Fig 4A) [27,28]. Switching to the active state forces residue 193 into a configuration that is strongly disfavored for nonglycine residues. Although glycine is highly conserved at position 193 in the trypsin family and is critical for catalysis, MamO contains a glutamate (E223) at this position [29]. Therefore, the active configuration of MamO would contain a strong steric clash between the E223 side chain and the main chain carbonyl of W222.
To illustrate this steric constraint, we examined a set of trypsin-like protease structures and analyzed the configuration of residue 193. We plotted ϕ and Ψ values for this position on a Ramachandran plot along with the favored and allowed geometries for glycine and nonglycine residues (S2 Table) [30]. As expected, the main-chain torsion angles at position 193 form two groups, distinguished by an approximately180° shift in ϕ angle. The groups correspond to active and inactive forms of loop L1. The active configuration is strongly disfavored unless glycine occupies position 193, indicating that the E223 side-chain in MamO likely prevents the formation of an oxyanion hole (Fig 4B).
While refining the MamO structure, we observed a peptide bound near the predicted peptide-binding groove (Fig 5A). We suspect its source is the flexible N-terminal region from a neighboring MamO in the crystal that is not built into the model. However, due to the modest resolution, we cannot confirm the sequence. Despite this uncertainty, the peptide has an interesting mode of binding. Similar to other trypsin-like proteases, the peptide enters the binding cleft parallel to loop L2, splitting the two β-barrels of the chymotrypsin fold. However, the bulky side-chain of W222 in MamO seems to block the exit path between loops LA and LD, forcing the peptide away from the catalytic center (Fig 5B–5D). Overall, the structural features of the L1 loop appear incompatible with protease activity: E223 provides an energetic barrier to catalysis while W222 serves as a physical barrier to productive substrate binding.
Given the results of the genetic studies and structural analysis, we conclude that MamO does not act as a protease during biomineralization. This forced us to consider the possibility that the protease domain promotes magnetite formation using a function not predicted from its primary sequence. While purifying MamO, we serendipitously observed that it consistently bound to immobilized metal affinity columns. Knowing that mutations in MamO cause defects in AMB-1’s ability to transform iron, a metal, into magnetite, we speculated that direct metal binding played a role in biomineralization.
Although the MamO crystals grow in acidic conditions that disfavor metal binding, they could be soaked at pH 8.0 without affecting diffraction. We solved another structure of the protease domain using crystals soaked in NiCl2 at pH 8.0 and identified a metal binding site. Overall, the conformation of the protease domain is highly similar to the original structure (root-mean-square deviation of 0.17 Å over 184 residues), and it contains the unidentified peptide. Additionally, a single Ni2+ ion binds between loop LC and helix 2, with H148 and H263 directly coordinating the metal (Fig 6A–6C). We confirmed the placement of the ligand at this site using single wavelength anomalous dispersion (SAD) data collected at the Ni absorption edge (Fig 6A and 6B). MamO’s di-histidine motif is highly reminiscent of the Zn2+ binding site in another trypsin-like protease, kallikrein-3, in which Zn2+attenuates protease activity by altering the position of catalytic triad residues H57 and D102 (S5 Fig) [31,32].
To confirm the putative metal binding site from the MamO structure, we used transition metal ion Förster Resonance Energy Transfer (tmFRET) to assay binding in solution [33]. This technique measures fluorescence quenching of a cysteine-conjugated fluorophore by a metal bound at a nearby site. Guided by the structure, we targeted Q258 because of its optimal geometry relative to the metal. Adding a number of transition metals, including iron, to a purified Q258C mutant protease domain that was labeled with fluorescein-5-maleimide caused strong fluorescence quenching (S6 and S7 Figs). Although we expect Fe2+ to be the physiological ligand, its propensity to oxidize in ambient atmosphere added significant error to the measurements. Instead, the quenching properties and resistance to oxidation of Ni2+ were most suitable for detailed analysis. MamO bound to Ni2+ with 2.5 μM affinity, compared to 1.1 μM in an H148A/H263A mutant. Additionally, the FRET efficiency was significantly lower in the mutant protein, demonstrating that disrupting these residues changes the metal binding properties of MamO (Figs 6D and S7).
Although the H148A/H263A mutant displays altered behavior in the tmFRET assay, it retains the ability to bind metals (S7 Fig). We could not identify any other metal ions in our Ni2+ soaked crystals, but the two histidines we identified did not appear to be the only metal binding residues in MamO. Consistent with this, both the wild-type and H148A/H263A forms of MamO bound Ni-NTA resin, confirming that, despite its altered binding geometry, the mutant still binds to metal (S7 Fig). Taken together, our biochemical and structural investigations show that MamO binds to transition metal ions using H148 and H263, but that metal binding is not restricted to this motif.
Because disrupting H148 and H263 altered the metal binding behavior in vitro, we predicted that these residues were important for MamO-dependent biomineralization. Indeed, ΔOΔR9 strains with mamOH148A or mamOH263A alleles had no magnetic response and failed to produce electron-dense particles, implying that proper metal binding is required to initiate biomineralization (Fig 6E and 6F). Both mutants displayed normal stability and proteolytic processing, and the biomineralization defects were not conditional, as lowering the growth temperature did not allow for a magnetic response (S1 and S2 Figs). Additionally, both alleles restored a magnetic response in the single ΔmamO background, showing that limO can provide the required metal binding activity independent of cotranslation with the TauE domain (S3 Fig). These are the most disruptive point mutations we have observed in MamO, and they completely recapitulate the phenotype of a mamO deletion. We conclude that H148 and H263 contribute to a metal binding function that is required for magnetite nucleation in vivo.
While the finding that MamO has lost its protease activity to become a metal binding protein in AMB-1 was quite intriguing, we wanted to know whether this mechanism was conserved in other organisms. Due to the fact that magnetotactic Nitrospirae and Omnitrophica have not been isolated in culture, we focused our analysis on the Proteobacteria for which numerous representatives are available in pure-culture. Examining available whole-genome sequences indicated that all magnetotactic strains from the α-, δ- and γ-Proteobacteria contain two predicted trypsin-like proteases in their MAIs. We attempted to understand the evolutionary history of these proteins by including them in a large phylogenetic tree of the bacterial trypsin-2 superfamily (Methods). Within this tree, the MamO sequences from each magnetotactic α-Proteobacterium form a distinct, monophyletic clade (Fig 7A). Each protein has a degenerate catalytic triad, a nonglycine residue at position 193, and a bulky tryptophan in the L1 loop. The metal binding positions in LC and helix 2 also appear conserved. Although one strain has an aspartate at position 263, aspartates are also common metal coordinating residues, and we predict that this H-D motif can also participate in metal binding. We conclude that the MamO family evolved specifically in α-Proteobacterial magnetotactic organisms to be metal-binding pseudoproteases (S8 Fig).
In addition to the MamO proteins, we identified a second magnetosome-specific clade in the trypsin tree (Fig 7A). We named this group the MamE clade because it contains the MamE sequence from the MAI of each representative of the α-Proteobacteria. This group also features two predicted trypsin-like proteases from the MAI of each species of the δ- and γ-Proteobacteria, renamed MamE1 and MamE2 here (S3 Table). The sequence phylogeny indicates that the δ- and γ-Proteobacteria both experienced recent duplications of their MamEs independently. In fact, there appears to have been two rounds of duplication in the δ- family. Strikingly, each duplication event has led to a degenerate catalytic triad in one of the sequences, showing that duplication and loss of function occurred three separate times in the MamE family (Figs 7B and S9). While these inactive MamEs do not have the dihistidine motif identified in MamO, we cannot rule out the possibility that they bind metals by another mechanism. Regardless, our results imply the existence of selective pressure for pairing active and inactive proteases as each major clade of magnetotactic bacteria has evolved this feature independently.
Magnetotactic bacteria control the growth of their associated magnetite crystals with a level of precision that cannot be replicated in vitro. The molecular details of how they perform this task can reveal novel bioinorganic interfaces and be exploited for improved synthesis of nanomaterials. Genetic analysis has shown that magnetite biomineralization is surprisingly complex. It requires over 15 factors in AMB-1, nearly all of which are predicted integral membrane proteins [2,19,34]. A subset of these is required for the initial crystallization of iron within the magnetosome compartment [10]. While the key players for this step are known, their biochemical functions have only been inferred from sequence homology, leaving the mechanism of how magnetite biosynthesis begins a mystery.
Here, we examined two magnetite nucleation factors: the putative HtrA proteases MamE and MamO. We find that the presence of both MamE and MamO is required for the proteolysis of three biomineralization factors, MamE, MamO, and MamP. These events depend on an intact catalytic triad from MamE but not MamO, indicating that MamO activates MamE in a noncatalytic manner. Thus far, we have not been able to detect a physical interaction between MamE and MamO, but we have found that the C-terminal TauE-like transporter domain of MamO is required for activation. The putative ion transport activity of this domain could be the feature that promotes proteolysis. This model is attractive because it does not require a direct interaction between the two proteins. Indirect evidence suggests that TauE family proteins transport sulfite or sulfur containing organic ions, leading us to speculate that the concentrations of specific solutes in the magnetosome might control MamE’s activity [35,36].
Separately, we discovered that a metal-binding function in the protease domain of MamO is required for the initiation of magnetite biomineralization. In our structure, H148 and H263 directly coordinate a single metal ion. Disrupting these residues alters the binding behavior in a modified tmFRET assay, but the effect is unusual in that it lowers the FRET efficiency while slightly increasing the overall affinity for metal. Using the Förster equation and reported radius for Ni2+, we calculated that the fluorophore to metal distance changes from 12.8 Å in wild-type to 15.2 Å in the H148A/H263A mutant, suggesting that the binding geometry changes in a way that allows metal binding in the same vicinity [37]. We favor the explanation that the dihistidine motif identified here is part of a more complex metal coordinating network. Our soaking strategy may have missed additional sites that are inaccessible in our crystal form, and separate attempts at characterizing a fully metal-bound state using co-crystallization have been unsuccessful. Nevertheless, we identified dihistidine motif using the structure that contributes to an unexpected transition metal-binding activity in MamO.
Despite the presence of other binding features, metal binding through H148 and H263 is absolutely required in vivo as disruption of either residue completely abolishes magnetite formation. Our structural analysis shows that MamO has lost its ability to perform proteolysis altogether, supporting the idea that metal binding is now the central function of the protease domain. Consistent with this, T225, the predicted catalytic nucleophile, is completely dispensable for biomineralization. Though disrupting H116 and D149 in the predicted catalytic triad causes conditional crystal maturation defects, magnetite nucleation is not affected. Interestingly, H116 and D149 participate in a hydrogen bond on the opposite face of loop LC from the H148/H263 metal binding motif, suggesting that the conditional phenotypes could be due to temperature-dependent flexibility near the metal binding site (S5 Fig). A potential link between the two motifs is consistent with the reported inhibition of protease activity through a highly analogous metal binding site in the kallikrein family that rearranges the H-D catalytic pair (S5 Fig).
While templating of magnetite growth via an interaction between biomineralization factors and the mineral surface has been proposed, our findings with MamO emphasize that direct interactions with individual solute ions also play a role [34,38,39]. One of the most fascinating aspects of MamO’s metal ion interaction is that the H148A and H263A forms of MamO maintain the ability to bind metals but cannot support any magnetite biosynthesis in vivo. It appears that binding is insufficient and that the precise coordination geometry must be maintained, leading us to speculate that MamO directly promotes nucleation by guiding individual iron atoms into the magnetite lattice. This model is consistent with the phenotypes observed in vivo, the modest binding affinity and the surface exposed nature of the simple dihistidine motif. Additionally, it agrees with topological predictions for MamO placing the protease domain in periplasm, which is continuous with the magnetosome lumen in AMB-1 [18]. More broadly, our results define an unexpected mechanism for MamO in biomineralization. It appears to have lost the ability to perform serine protease activity and instead performs two noncatalytic functions: direct metal binding to promote magnetite nucleation and activation of MamE’s proteolytic activity (Fig 8A).
In addition to the surprising mechanism for MamO in AMB-1, we uncovered a fascinating evolutionary expansion of the trypsin family within magnetotactic bacteria. Our analysis suggests that the ancestral MAI contained a single trypsin-like protease homologous to MamE. The δ- and γ-Proteobacteria experienced independent duplications of this ancestral enzyme, while the α-Proteobacteria appear to have acquired a second, distantly related trypsin-like protease. Despite these different origins, having two redundant proteases seems to have allowed one copy to lose its catalytic ability in all three clades (Fig 8B). In α-Proteobacteria, MamO specialized to promote biomineralization through the two noncatalytic activities identified here. While the pseudoproteases in the other clades remain uninvestigated, the fact that inactive copies are retained strongly suggests that they also play important noncatalytic roles. The pathway that led to convergent evolution of pseudoproteases in magnetotactic organisms highlights the critical role duplication and redundancy play in facilitating diversification of protein function [40].
Perhaps more intriguing is the fact that MamO’s metal binding motif is placed at the same site on the chymotrypsin fold as the highly analogous zinc-binding site seen in the distantly related kallikreins [31,32]. This hints toward the possibility that the ability to bind metals may be a latent biochemical function carried within the fold. Such activities are absent in specific evolutionary states of a protein but can quickly surface under selective pressure [41,42]. Consistent with this, trypsin-like proteases utilize catalytic residues on loops that are well separated from the core, a property termed fold polarity that correlates with the capacity for functional diversification [43,44]. Perhaps neofunctionalization of the trypsin scaffold within magnetotactic organisms is due to an inherent stability and adaptability in the fold that makes it a useful building block for biochemical innovation.
The strains and plasmids used in this study are listed in S4 and S5 Tables, respectively. For general maintenance and genetic manipulation, M. magneticum AMB-1 was grown in MG medium supplemented with ferric malate (30 μM). 0.7% agar was used in plates, and kanamycin was used for antibiotic selection at a concentration of 7 μg/mL (solid) or 10 μg/mL (liquid). For sucrose counterselection, MG plates contained 2% sucrose. Cultures for magnetic response measurements, western blotting, or TEM were grown in 10 mL MG medium containing 25 mM HEPES buffer (pH 7.2) and ferric malate under a 10% oxygen atmosphere. For comparing the temperature dependence of magnetic response, the strains were treated as above, except that they were grown under anaerobic atmosphere. The magnetic response of each culture was assessed using the Coefficient of Magnetism (Cmag), which was measured as described [18].
For complementation of deletion mutants, we used a modified form of pAK253 [10]. This plasmid contains a neutral region of the AMB-1 genome and integrates as a single copy at this site. Each allele is inserted under the control of the mamAB promoter, allowing the constitutive expression of each protein. To create the ΔmamEΔmamOΔR9 (ΔEΔO) strain, plasmid pAK243 (sacB-based counterselection system for deleting mamE) was transformed into strain AK94 (ΔmamOΔR9), and initial integrations were selected using kanamycin. The resulting strains were grown to stationary phase in MG medium without antibiotic selection and plated on sucrose for counterselection. Deletion of mamE was confirmed by antibiotic sensitivity, PCR analysis, and complementation of the magnetic phenotype upon reintroduction of both mamE and mamO.
Cultures of AMB-1 were grown to late-log phase and harvested by centrifugation at 6k x g. The resulting pellets were resuspended in PBS. The cell suspensions were mixed with an equal volume of 4x SDS Loading buffer and heated for 10 min at 75°C. The lysates were separated on SDS-PAGE and transferred to PVDF. The membranes were blotted and visualized using standard western blotting techniques. Polyclonal antibodies to MamE and MamP were raised in rabbits against recombinant forms of the soluble portion of each protein [25]. The monoclonal anti-FLAG antibody was purchased from Sigma.
For TEM, cultures were grown to late-log phase. 1 mL from each culture was pelleted at 16k x g, and the pellet was resuspended in the residual medium. Cell suspensions were spotted on formvar-coated copper grids, rinsed, dabbed dry, and stored at room temperature until imaging. Imaging was performed with a FEI Tecnai 12 TEM at an accelerating voltage of 120 kV. For each strain, 15–20 cells totaling >200 crystals were analyzed.
The protease domain of MmMamO (residues 45–261) was cloned into mcsII of pETDuet to create pAK876 for expression without a tag. BL21 Codon Plus cells transformed with pAK876 were grown at 37°C in 2xYT with carbenicillin (100 μg/mL) and chloramphenicol (25 μg/mL) until the OD600 reached 1.0. The cultures were then equilibrated at 20°C for 30 min, induced with 0.125 mM IPTG, and grown overnight. Cells were harvested by centrifugation, resuspended in Buffer A (25 mM Tris-HCl pH7.4, 400 mM NaCl, 5 mM Imidazole, 10% glycerol) supplemented with 1 μg/mL pepstatin A, 1 μg/mL leupeptin, and 0.5 mM DTT, frozen in liquid nitrogen, and stored at -80°C.
For crystallography, the frozen cell suspension was thawed on ice, lysed by sonication, and clarified at 13,000 x g for 30 min. The supernatant was loaded on a 3 mL Ni-NTA column, which was then washed with five column volumes of Buffer B (25 mM Tris-HCl pH 7.4, 500 mM NaCl, 10 mM Imidazole, 10% glycerol). The protein binds to Ni-NTA with its native metal binding site. After elution with one column volume Buffer C (25 mM Tris-HCl pH 7.4, 250 mM NaCl, 250 mM Imidazole, 10% glycerol), the sample was dialyzed overnight against Buffer D (25 mM Tris-HCl pH 8.0, 50 mM NaCl, 10% glycerol, 0.5 mM EDTA) and loaded onto a 1 mL HiTrap CaptoQImpRes column. The column was developed to Buffer E (25 mM Tris-HCl pH 8.0, 1 M NaCl, 10% glycerol), and the peak fractions containing the MamO protease domain were pooled, exchanged into Storage Buffer (25 mM Tris-HCl pH 7.4, 300 mM NaCl, 10% glycerol), concentrated to approximately 20 mg/mL and frozen in liquid nitrogen for storage at -80°C.
For tmFRET, the expression was performed as above except that the various mutants of the MmMamO protease domains were expressed as fusions to a C-terminal strepII tag. The lysate was prepared as above and loaded onto a 1 mL StrepTrap HP column, which was washed with five column volumes of Buffer F (25 mM Tris-HCl pH 7.4, 250 mM NaCl, 10% glycerol) and eluted with three column volumes of Buffer F supplemented with 2.5 mM desthiobiotin. The eluate was concentrated and loaded onto a 16/60 Superdex 200 column and developed in Storage Buffer. The peak fractions were pooled, concentrated, and used immediately for fluorescent labeling. The purity of all proteins was confirmed by SDS-PAGE.
Frozen aliquots of untagged MamO protease domain were thawed on ice and exchanged to Buffer G (20 mM Tris-HCl pH 7.4, 150 mM NaCl, 5% glycerol, 0.5 mM DTT) while adjusting the protein concentration to 5 mg/mL in a 10 kDa cutoff Amicon ultrafilter. Crystals grew in the hanging drop vapor diffusion format after mixing the protein with an equal volume of well solution (50 mM Na-Acetate pH 4.6, 3.6 M NH4Cl, 5% glycerol) and equilibrating against 1 mL well solution at 18°C. Cubic crystals of MamO appeared in 1–2 d and grew to their full size after 4–6 wk. Each crystal was cryoprotected with a solution of 50 mM Na-Acetate pH 4.6, 3.6 M NH4Cl, and 22% glycerol before plunge-freezing in liquid N2.
For nickel soaking, NiCl2 was added directly to drops containing fully-grown crystals for a 10 mM final concentration. The wells were resealed and the crystals soaked overnight. At harvest, each crystal was passed through three 1 min soaks in a buffer containing 50 mM Tris-HCl pH 8.0, 3.6 M NH4Cl, 22% glycerol, and 10 mM NiCl2 before plunge-freezing.
Diffraction data was collected on beamline 8.3.1 at the Advance Light Source (Lawrence Berkeley National Laboratory, Berkeley, California). Indexing and scaling was performed using HKL2000 [45]. The apo- structure was solved by molecular replacement in Phaser with the protease domain of EcDegP (1KY9) as a search model [46]. The Ni-bound structure was solved using the apo-MamO structure as a molecular replacement search model. Model building and refinement were carried out with alternating cycles of COOT [47] and phenix.refine [48]. Placement of the Ni ion was confirmed with an anomalous map from SAD data collected at the Ni absorption edge.
For fluorescent labeling, variants of the MamO protease domain were diluted to 50 μM in 1 mL of Buffer H (50 mM NaPhosphate pH 7.2, 150 mM NaCl, 10% glycerol). Fluorescein-5-maleimide (dissolved at 50 mM in DMSO) was added to a final concentration of 1 mM, and the reaction was incubated overnight at 4°C. The reaction was quenched with DTT, exchanged into Buffer D using a PD-10 column, and loaded on a 1 mL HiTrap Q FF. The column was developed to Buffer E, and each protein eluted as a single fluorescent peak. The peak fractions were pooled, treated with 1 mM EDTA to remove any trace metal, exchanged extensively into Chelex-treated Storage Buffer, and frozen in small aliquots. Labeling efficiency was measured based on A492/A280 and was between 50% and 55% for all preparations.
For metal binding experiments, all buffers were prepared in acid-washed glassware and treated with chelex resin. Each protein was diluted to 60–80 nM in fluorescence buffer (10 mM Tris-HCl pH 8.0, 150 mM NaCl). Metal solutions were prepared at 10X concentration in chelex-treated H2O. Protein was dispensed into 96-well plates, and the plates were scanned for fluorescence emission in a Tecan Infinite 200 plate reader in top-read mode (Ex: 492 nm; Em: 505–570 nm). Metal solutions were then diluted into each well at the appropriate concentration and the plate was rescanned. Due to the spontaneous oxidation of ferrous iron in ambient atmosphere, all iron binding experiments were performed in the absence of oxygen. For iron binding, all solutions were prepared using anoxic liquids. The samples were prepared under anoxic atmosphere in a clear-bottom 96-well plate. To prevent introduction of oxygen, the wells were sealed by covering the plate with a black adhesive cover. The plate was then removed from the anaerobic chamber and scanned using bottom-read mode.
Each measurement was performed on independently prepared solutions in quadruplicate. The metal-quenched fluorescence spectrum from each well was normalized to the fluorescence before metal addition. Fmetal/F represents the normalized fluorescence averaged from an 11 nm window around the peak. After plotting Fmetal/F as a function of metal concentration, each curve was fit to a two-site model (below), in which Kd2 and E represent the dissociation constant and FRET efficiency for metal binding by MamO, respectively. Kd1 represents a nonspecific, solution-based quenching component [37].
StrepII-tagged forms of the MamO protease domain were purified as described for the tmFRET experiments with the omission of the F-5-M labeling and ion exchange steps. Proteins were diluted to approximately 15 μM in Column Buffer (25 mM Tris-HCl pH 7.4, 300 mM NaCl, 5 mM Imidazole, 10% glycerol) and loaded onto a 0.5 mL Ni-NTA column equilibrated in column buffer. The column was washed with ten column volumes of Column Buffer and eluted with two column volumes Buffer C. Binding was assessed by separating the fractions on 12% SDS-PAGE and staining with Coomassie Blue.
To understand the phylogeny of Mam proteases, we took a broad approach, characterizing their location within the trypsin-like protease family. The Trypsin_2 Pfam (PF13365) was used to generate a hidden Markov model (HMM) that was used to search a database composed of protein sequences from approximately 2,100 bacterial and archaeal genomes along with the Mam proteases [49]. This search was performed using hmmscan (HMMER3.1b1, hmmer.janelia.org), retaining all hits with an e-value for the entire sequence less than 10−5, identifying 6,104 proteins. The hits were clustered using CD-HIT with a sequence similarity cutoff of 0.8, yielding 3,431 sequences [50]. These were aligned using MUSCLE (v3.8.31) with the maxiters parameter set to 2 [51]. The resulting alignment was trimmed using Gblocks, using parameters appropriate for divergent datasets as described by Sassera et al. [52,53]. This alignment was then used to generate a phylogeny using FastTree 2 (v2.1.7) with the default settings [54]. However, this alignment contained only six informative positions. To improve the quality of the alignment, we iteratively removed long branches from the tree and regenerated the alignment. After removing 90 taxa over four iterations, we settled on an alignment with 18 conserved positions.
Two clear branches emerged from this analysis. One branch ('the Deg branch') contained DegP, DegQ, and DegS sequences from the γ-Proteobacteria, in addition to most of the MamE sequences. The bottom branch contained the YdgD sequence from Escherichia coli. We extracted the sequences from the Deg branch (843 sequences) and generated an alignment (46 positions) and tree using the methods described above (Fig 7A). In this tree, the MamE sequences appeared to be closely related, but their phylogenetic relationships were ambiguous. Additionally, we realized that a subset of δ-proteobacterial trypsin-like sequences were falsely excluded from the MamE branch due to substitutions in the catalytic triad that obscured the phylogenetic signal in the short alignment. Additional phylogenetic testing confirmed the relationship of these four sequences to the canonical MamE sequences, so they were merged into the MamE branch.
We extracted the MamE sequences and aligned them using five DegS sequences from the γ-Proteobacteria, which appeared to be closely related based on the phylogeny in Fig 7B. This alignment was much larger (172 positions), and the resulting tree from FastTree 2 gave well-supported interior nodes for the MamE branch. We used two other phylogenetic methods to test this phylogeny. First, we used ProtTest 3.0 to select the best substitution matrix (in this case, the WAG model) and performed 100 independent inferences with 300 bootstraps in RAxML [55,56]. Secondly, we used the nonparametric Monte Carlo Markov Chain algorithm PhyloBayes 3 to generate a tree not based on prior assumptions about the site-specific evolution of the MamE sequences [57]. A summary tree integrating the results from PhyloBayes and RAxML is depicted in Fig 7B. Additionally, we rooted the MamE branch to five closely related sequences from the Clostridiales, and this, too, strongly supported the phylogeny in Fig 7B according to both PhyloBayes and RAxML.
|
10.1371/journal.ppat.1002292 | The Anti-interferon Activity of Conserved Viral dUTPase ORF54 is Essential for an Effective MHV-68 Infection | Gammaherpesviruses such as KSHV and EBV establish lifelong persistent infections through latency in lymphocytes. These viruses have evolved several strategies to counteract the various components of the innate and adaptive immune systems. We conducted an unbiased screen using the genetically and biologically related virus, MHV-68, to find viral ORFs involved in the inhibition of type I interferon signaling and identified a conserved viral dUTPase, ORF54. Here we define the contribution of ORF54 in type I interferon inhibition by ectopic expression and through the use of genetically modified MHV-68. ORF54 and an ORF54 lacking dUTPase enzymatic activity efficiently inhibit type I interferon signaling by inducing the degradation of the type I interferon receptor protein IFNAR1. Subsequently, we show in vitro that the lack of ORF54 causes a reduction in lytic replication in the presence of type I interferon signaling. Investigation of the physiological consequence of IFNAR1 degradation and importance of ORF54 during MHV-68 in vivo infection demonstrates that ORF54 has an even greater impact on persistent infection than on lytic replication. MHV-68 lacking ORF54 expression is unable to efficiently establish latent infection in lymphocytes, although it replicates relatively normally in lung tissues. However, infection of IFNAR−/− mice alleviates this phenotype, emphasizing the specific role of ORF54 in type I interferon inhibition. Infection of mice and cells by a recombinant MHV-68 virus harboring a site specific mutation in ORF54 rendering the dUTPase inactive demonstrates that dUTPase enzymatic activity is not required for anti-interferon function of ORF54. Moreover, we find that dUTPase activity is dispensable at all stages of MHV-68 infection analyzed. Overall, our data suggest that ORF54 has evolved anti-interferon activity in addition to its dUTPase enzymatic activity, and that it is actually the anti-interferon role that renders ORF54 critical for establishing an effective persistent infection of MHV-68.
| Human gammaherpesviruses, Kaposi's sarcoma-associated herpesvirus and Epstein-Barr virus, are the cause of several malignancies, especially in patients immunocompromised due to HIV infection. The study of these human gammaherpesviruses is difficult due to their inability to replicate in cell culture and the lack of a small-animal model. Murine gammaherpesvirus-68 is a genetically and biologically similar virus that is utilized as a mouse model because it offers such advantages as the ability to replicate in cell culture, a manipulatable genome, and infection of mice. In this study, we have identified viral open reading frame 54 (ORF54) as an inhibitor of innate immunity, specifically of the type I interferon response. Although ORF54 is a conserved viral dUTPase, we found that its anti-interferon activity does not require its enzymatic activity. Through infection of cells and mice, we define the critical role of ORF54 in establishing persistent latent infection of MHV-68 by inducing the degradation of the type I interferon receptor. Our studies provide new insights into the far reaching effects of type I interferon signaling and the dual role of ORF54. This work could aid in the development of vaccine strategies to gammaherpesvirus infection.
| Virus infection induces numerous immune responses in the host, the earliest of which is the innate immune response [1], [2]. The innate immune response is comprised of many layers of non-specific defense, including anatomical barriers, such as skin and mucosa, the complement system, inflammation, and various cells, such as natural killer cells, phagocytes, mast cells, macrophages, dendritic cells, neutrophils, and basophils [3]–[5]. The innate immune response plays a crucial role in shaping the ensuing adaptive immune response, in part by the production of cytokines in response to infection [2], [6]. Interferons (IFN) are cytokines secreted upon virus infection that induce the expression of a variety of antiviral gene products, reducing virus replication and further infection [1], [7]–[9].
Interferons are classified as type I and II, as defined by the cell types able to produce them and the receptors they bind to [1]. Unlike the type II IFN-γ that is produced by specific cells of the immune system, IFN-α and IFN-β are type I IFNs that can be produced in most cell types [10]. Mammals encode a single IFN-β and several IFN-α species. All type I IFN species bind to the same ubiquitously expressed receptor, called the type I interferon receptor, or IFNAR [11]. This receptor is a heterodimer comprised of IFNAR1 and IFNAR2 [12]. Although normally unassociated, IFNAR1 and IFNAR2 dimerize upon the binding of IFN-α or IFN-β first to IFNAR2, and then to both receptors in the dimer [13]. IFNAR1 and IFNAR2 are each pre-associated with one of the members of the Janus protein tyrosine kinase family, where TYK2 is associated with IFNAR1 and JAK1 with IFNAR2. IFN binding and formation of the complete IFNAR dimer leads to cross-phosphorylation of TYK2 and JAK1, and the phosphorylation of the IFNAR chains they are permanently associated with. These phosphorylation events set up a platform for the recruitment of STAT1 and STAT2, which in turn are also phosphorylated. The phosphorylated STAT proteins dimerize prior to joining with IFN-regulatory factor 9 (IRF9) to form the Interferon-Stimulated Gene Factor 3 (ISGF3γ) transcription factor, which translocates to the nucleus where it induces the expression of hundreds of interferon stimulatory genes (ISG) (reviewed in [1]).
Herpesviruses are large, double-stranded DNA viruses defined by their ability to persist for the lifetime of the host by establishing latent infections and by evading the host immune response [14]. Both lytic and latent infections of herpesviruses are able to directly cause disease [15]–[23]. Gaining understanding of the mechanisms by which herpesviruses maintain persistent infections and evade immune surveillance is a key step in controlling the diseases they are associated with. Herpesviruses are subdivided into alpha, beta, and gamma herpesviruses [14]. The human gammaherpesviruses are Epstein-Barr virus (EBV) and Kaposi's sarcoma-associated herpesvirus (KSHV). Both EBV and KSHV are associated with malignancies; EBV with Burkitt's lymphoma, nasopharyngeal carcinoma, and oral hairy leukoplakia and KSHV with Kaposi's sarcoma, primary effusion lymphoma, and multicentric Castleman's disease [23]–[28]. We use the biologically and genetically related virus, murine gammaherpesvirus-68 (MHV-68) as a model to study the human gammaherpesviruses [29]–[32]. Like KSHV, MHV-68 is a gamma-2-herpesvirus. MHV-68 establishes lytic and latent infections in mice [33], replicates readily in in vitro cell culture systems, and has a genome that can be genetically manipulated by utilizing a bacterial artificial chromosome (BAC) system [34].
In KSHV and MHV-68 several studies have identified viral proteins involved in the inhibition of the host innate and adaptive immune responses. In particular, KSHV open reading frame (ORF) 10 binds to JAK and STAT proteins to block IFN mediated signaling [35]. KSHV and MHV-68 ORF36 bind to phosphorylated IRF3, thus inhibiting the production of IFN-β [36]. KSHV ORF45 interacts with and inhibits IRF7 from entering the cell nucleus [37]. The KSHV immediate early transcription factor and E3 ligase, RTA, targets IRF7 for protein degradation [38]. Other viral ORFs also contribute to immune evasion by inducing the downregulation of surface molecules critical for immune activation. K5 of KSHV directs the downregulation of Tetherin/BST2, ICAM, and MHC class I [39]–[42]. Furthermore, K3 of both KSHV and MHV-68 inhibit the surface expression of MHC class I [39], [43].
2′-deoxyuridine 5′-triphosphate pyrophosphatase (dUTPase) reduces the misincorporation of uracil in the DNA genome by controlling the level of dUTP through conversion of dUTP into dUMP, ultimately leading to an increased amount of dTTP and a lower dUTP∶dTTP ratio [44]. This enzyme can be found in all classes of organisms and in many RNA and DNA viruses as well [45], [46]. In this study, we define the role of viral ORF54, a functional dUTPase, in evading the host innate immune response to virus infection. In an effort to understand these two separate functions of ORF54, we analyzed the signaling pathways altered by wild-type ORF54 and a dUTPase-null mutant. Through in vitro and in vivo infection, in wild-type and transgenic mice, with recombinant MHV-68 harboring mutations in ORF54, we found that ORF54 interferes with-type I interferon signaling, which further affects persistent infection and the establishment of latency.
To systematically identify MHV-68 viral ORFs that inhibit type I IFN signaling, we conducted a screen where 293T cells were transiently transfected with a reporter construct containing firefly luciferase driven by the interferon-stimulated response element (ISRE_firefly-luciferase). Cells were also co-transfected with a reporter construct containing renilla luciferase driven by the constitutively active PGK promoter and each MHV-68 viral ORF or a vector control. Transfected cells were treated with human IFN-α and the induction of the ISRE reporter was measured by dual luciferase assay. Since this screen examines cellular responses after IFN-α treatment, the viral proteins previously identified to inhibit IRF3 or IRF7 signaling, thus preventing the induction of type I IFN production, would not necessarily be identified. Of all the MHV-68 viral ORFs that were screened, we found 8 ORFs that were potentially able to inhibit type I IFN signaling to a level that is 50% of the activation seen with vector control. Two of them are M2 and M8, ORFs specific to MHV-68. M2 has been previously shown to inhibit IFN signaling [47], thus validating our screen. Among the other 6 ORFs that also have homologues in KSHV and EBV, ORFs 10, 11, and 54 are particularly interesting because a previous sequence analysis study identified a shared dUTPase-related domain, although only ORF54 contains catalytic active sites [48]. Since ORF54 is a viral dUTPase and is one of the top three strongest inhibitors, our following study focuses on its potential anti-IFN function and the biological significance of this function during viral infection. Cells transfected with ORF54 demonstrated a diminished activation of ISRE following treatment with IFN-α (Figure 1), with only 20% of the activation seen in control transfections. KSHV ORF54 also demonstrated diminished activation of ISRE, at 23% of control (Figure 1), suggesting that the ability to inhibit type I IFN signaling is a gammaherpesvirus conserved function of ORF54.
To test whether the dUTPase function is required for the anti-IFN activity of ORF54, we constructed a catalytic domain mutant of MHV-68 ORF54 by replacing the amino acid histidine at position 80 with alanine and the amino acid aspartic acid at position 85 with asparagine (ORF54 H80A/D85N). These two amino acids were chosen for mutagenesis due to their proximal location in the putative active site of ORF54 dUTPase and their predicted importance for enzymatic reaction. Viral and non-viral dUTPases typically share five highly conserved motifs, although the arrangement found in herpesviruses is different compared to human dUTPase [44], [48]. Several studies have identified the presence of dUTPase motif III, which is critical for catalytic activity, in the herpesvirus dUTPases [44]. Aspartic acid at position 85, located in motif III, was altered because aspartic acids at positions 84 and 86 in human endogenous retrovirus (HERV-K) were found to be critical for catalytic activity, but not for dUTP binding [44]. Histidine at position 80 was chosen because it is conserved in gammaherpesviruses, and in EBV a histidine at position 71 contains a necessary imidazole group [49]. The ORF54 H80A/D85N mutant demonstrates a complete loss in enzymatic activity, although the protein expression level remained the same (Figure 2). When co-transfected with ISRE_firefly-luciferase, ORF54 H80A/D85N maintains the ability to diminish IFN-α induced activation of the ISRE to 32% of the activation seen in control samples (Figure 1). This result suggests ORF54 inhibition of the type I IFN signaling cascade is independent of its dUTPase enzymatic activity. To further clarify if the dUTPase function was sufficient to inhibit type I IFN signaling, we tested the ability of murine cellular dUTPase to inhibit IFN-α induced activation of ISRE in our reporter assay. The cellular dUTPase was unable to block activation of ISRE_firefly-luciferase (Figure 1), further indicating that dUTPase enzymatic activity does not necessarily correlate with anti-IFN activity.
The type I IFN signaling pathway begins with the binding of IFN-α to the surface IFNAR and results in the production of ISGs [7], [8]. As each step in the JAK/STAT pathway that ensues is well defined [1], we aimed to identify the step where ORF54 exerts its function. We first assayed a central event in type I IFN signaling, the phosphorylation of STAT1 protein. 293T cells ectopically overexpressing MHV-68 ORF54 or ORF54 H80A/D85N both demonstrated a reduced level of phosphorylation of STAT1 following treatment by IFN-α in comparison to cells transfected with vector control or an unrelated MHV-68 ORF (Figure 3A). By assay of steps upstream of the phosphorylation of STAT1, we show that cells expressing MHV-68 ORF54 or ORF54 H80A/D85N both demonstrated a reduced level of total IFNAR1 (Figure 3B). As a control for specificity, we also found that ORF54 and ORF54 H80A/D85N do not alter levels of the surface protein type I insulin-like growth factor receptor-β (IGF1β) or IFNAR2 (Figure 3B). These results suggest that ORF54 induces the degradation of IFNAR1 independently of its dUTPase enzyme activity, and that this degradation results in a reduction of the type I IFN response, including the phosphorylation of STAT1. The transcript level of IFNAR1 remains comparable between the vector control and viral ORF transfected cells (Figure 3C), suggesting that the ORF54 induced reduction of IFNAR1 is at the protein level.
We generated three recombinant MHV-68 to study the importance of ORF54 (Figure 4A). Because ORF54 is not required for the virus to replicate in cultured cells [50], the first mutant is an ORF54-null virus that has triple translational stop codons inserted near the N-terminus of MHV-68 ORF54 (54Stop). The second virus is a revertant for 54Stop, where the translational stop codons have been removed and reverted back to wild type (54R). This virus ensures that any phenotype demonstrated with 54Stop is due to the ORF54-null mutation and not any additional recombination or unintentional mutagenesis of the MHV-68 viral genome. The third virus (54DM) has the same two amino acid mutations that abolish dUTPase activity as in the ORF54 H80A/D85N expression construct (Figure 4A).
NIH3T3 cells were infected with an MOI of 2 to initiate single-step growth kinetics. 24 hours post infection, cells were either treated for 15 minutes with 500 units/mL of mouse IFN-α or left untreated. Uninfected cells were also treated as a control. Immunoblots probing for the phosphorylation of STAT1 protein demonstrate that while wild-type (WT) MHV-68, 54DM, and 54R are able to block STAT1 phosphorylation, higher levels of phosphorylated STAT1 are found in cells infected with 54Stop (Figure 4B). Total STAT1 is comparable in all samples and all cells are infected to a similar level, as demonstrated by equal production of the MHV-68 late protein ORF65. Therefore, this result supports the role of ORF54 in the inhibition of type I IFN signaling during viral infection, as lacking ORF54 partially alleviates the block. However, we also noted that the phosphorylated STAT1 seen in cells infected with 54Stop is still not at the level observed in uninfected and treated cells. This is perhaps due to the other viral ORFs that are present and still able to inhibit the type I IFN signaling pathway, as evidenced by the multiple ORFs identified in our original screen. As an additional control to demonstrate that infected cells are still capable of responding to stimuli, NIH3T3 cells were identically infected with an MOI of 2 and treated at 24 hours post infection with 160 µM prostratin (12-deoxyphorbol 13-acetate) for 30 minutes (Figure 4B). Prostratin initiates a signaling cascade that induces the reactivation of latent viruses [51]. In this control assay, immunoblots probing for the phosphorylation of p44/p42 (Erk1/2) demonstrate that infected cells are viable enough to respond to stimuli even 24 hours post infection. These studies demonstrate that although at this MOI the 54Stop virus can establish a robust viral infection, the lack of functional ORF54 makes it unable to effectively block the resulting induction of the type I IFN signaling cascade.
Since we showed the degradation of IFNAR1 in the presence of overexpressed ORF54 (Figure 3B), we assayed this phenotype during virus infection to determine the biological relevance. By infecting cells with WT MHV-68 and 54R we found that IFNAR1 is degraded (Figure 4C). This degradation is reduced in cells infected with 54Stop virus, suggesting that ORF54 is required for the virus to induce degradation of IFNAR1. Furthermore, because cells infected with 54DM virus show a similar level of IFNAR1 reduction as with WT or 54R infection, this indicates that ORF54 dUTPase enzymatic function is not required for the degradation of IFNAR1 and the inhibition of the type I IFN signaling cascade. Interestingly, unlike with STAT1 phosphorylation, we observed a comparable level of IFNAR1 between 54Stop-infected cells and uninfected cells, indicating that ORF54 is the sole viral protein responsible for IFNAR1 degradation during viral infection. As a control for specificity, we also found that virus infection does not alter the levels of IFNAR2 or the surface protein IGF1-β (Figure 4C). Therefore, ORF54 mediated degradation is not a general phenomenon for all surface proteins and IFNAR1 is a specific target for such ORF54 function. The transcript level of IFNAR1 remains comparable amongst cells infected with WT, 54Stop, 54DM, and 54R viruses (Figure S1), suggesting that ORF54 does not alter the transcription of IFNAR1 and the ORF54 induced degradation of IFNAR1 is at the protein level.
Because we found 54Stop defective in inhibiting type I IFN signaling due to its inability to induce IFNAR1 degradation, we assayed the downstream induction of ISGs following infection of bone marrow derived macrophages. Macrophages were chosen for infection due to their high endogenous induction of anti-viral genes following virus infection. 24 hours post infection at an MOI of 2, cells were harvested for immunoblot against the ISG IFIT2. All infected cells show induced expression of IFIT2 compared to uninfected cells. IFIT2 protein expression was highest in 54Stop infected cells compared to in cells infected with WT, 54DM, and 54R (Figure 5A), suggesting a virus lacking functional ORF54 is not as effective in blocking ISG induction as WT MHV-68. Additionally, total RNA was harvested from infected macrophages to measure the transcript level of several ISGs, including MX1, IFIT1, and IFIT3. In all cases, the ISG induction in cells infected with 54Stop is higher than with infection by WT, 54DM, and 54R, suggesting ORF54 has a role in inhibition of type I IFN responses (Figure 5B).
Although ORF54 is not essential for MHV-68 replication, the 54Stop virus exhibits a defect in STAT1 activation and downregulating IFNAR1 expression. Hence, we speculated that lack of ORF54 might have some effect on viral replication in cells that are capable of producing and responding to type I IFN, such as NIH3T3 cells. Indeed, as shown in Figure 6A, we found that the 54Stop virus has moderately attenuated multiple-step growth in NIH3T3 cells (approximately 3.8-fold on day 4 and 4.1-fold attenuation on day 5 post infection compared to WT). Moreover, this attenuation of the 54Stop virus was not observed in Vero cells (Figure 6A), which are unable to produce type I IFN in response to infection [52]. To ensure this difference in phenotype of the 54Stop virus is due to its inability to block IFN signaling, we infected bone marrow derived macrophages from wild-type and IFN α/β receptor knockout (IFNAR-/-) mice at an MOI of 4. By immunoblot of infected wild-type macrophages, we see a lower production of the MHV-68 capsid protein, ORF65, with 54Stop infection compared to WT, 54DM, and 54R infection. However, the ORF65 production with 54Stop infection is rescued in the IFNAR−/− macrophages (Figure 6B). Similarly, the infectious titer produced from wild-type macrophages infected with 54Stop is between 6.3- to 7.5-fold lower than that of WT, 54DM, and 54R, while the infectious titers produced from IFNAR−/− macrophages are similar amongst all four virus types (Figure 6C). To further demonstrate the role of ORF54 in antagonizing IFN signaling during viral replication, virus production in NIH3T3 cells with or without the treatment of 100 units/mL of IFN-α was compared (Figure 6D). The peak viral titers of WT, 54DM, and 54R viruses were relatively unaffected at this low dose, while the average drop for 54Stop was 3-fold, p-value = 0.022. The growth defects observed with 54Stop in the presence of type I IFN response demonstrate the role of ORF54 in antagonizing the signaling pathway. The consistent, but modest, defects seen with 54Stop infection are also meaningful as it is known the virus has multiple anti-IFN genes [34]–[38], [53], [54]. Additionally, all of above in vitro viral replication results demonstrate that 54DM behaved similarly to WT and 54R viruses but not to 54Stop, indicating that the anti-IFN function of ORF54 is independent of its dUTPase enzymatic function.
We examined the role of ORF54 in vivo by infecting Balb/C mice with WT MHV-68, 54Stop, 54DM, or 54R and measuring lytic replication in the lungs on 5 and 7 days post infection (dpi) and the establishment of latency at its peak on 14 dpi. Lytic infection does not critically require ORF54. With intranasal infection we found that lytic replication in the lungs at 5 dpi appears to be only slightly affected by the lack of ORF54. The average infectious titer was approximately 3.5- to 3.8-fold lower with 54Stop infection than with WT MHV-68, 54DM, or 54R, with p-values 0.047, 0.002, and 0.018, respectively (Figure 7A). At 7 dpi the effects from the lack of ORF54 on the lung viral titers were even smaller (1.2- to 1.8-fold lower than others), and possibly insignificant with p-values of 0.048 for WT, 0.396 for 54DM, and 0.022 for 54R (Figure 7B). Similar results and trends were obtained with analysis of the viral genome copy number in the lung lysates (Figure S2). Infection of 54DM was comparable to WT and 54R viruses, indicating that function(s) other than dUTPase activity of ORF54 play a role, however minor, during productive infection in the lung.
At 14 dpi, we analyzed the establishment of splenic latency by performing an infectious center assay and quantifying viral genome copy numbers of the spleens of mice infected with WT MHV-68, 54Stop, 54DM, or 54R. The infectious center assay quantifies the amount of latent virus that reactivates from a population of B-cells upon ex vivo culturing. We found that the 54Stop virus has approximately 53- and 45-fold less reactivated virus per 107 lymphocytes compared to WT and 54R viruses, respectively (Figure 7C). Quantitation of the viral genome copies is an unbiased measurement that does not rely on virus activity in the assay itself, and it showed an even more pronounced reduction. The 54Stop genome is almost undetectable, with 140- to 164-fold lower genome copies than WT MHV-68, 54DM, and 54R (Figure 7D). Lower levels of viral genome copies and infectious centers suggest the defect of the 54Stop virus is in the establishment of latency and not in the ability to reactivate. If this drastic reduction in 54Stop latency is due to the inability of the virus to block type I IFN signaling, we would expect to see latency of 54Stop rescued in IFNAR−/− mice, which lack the type I IFN receptor. Indeed, in IFNAR−/− mice splenic latency of 54Stop at 14 dpi is greatly increased to a level similar to 54R virus; the number of infectious centers with 54Stop infection is 2.8-fold lower than with 54R, but with a statistically insignificant p-value of 0.248 (Figure 7E). Viral genome copies of 54Stop and 54R viruses from IFNAR−/− mice splenocytes are nearly identical (Figure 7F). Our results indicate that ORF54 is required for establishing a latent infection, and this requirement is based on its anti-IFN activity as the deficiency of 54Stop is rescued in mice unresponsive to type I interferon. This result concludes that the major role of ORF54 for establishing and/or maintaining latency is to inhibit type I interferon responses.
Viral pathogens have adapted several avenues of immune evasion, including inhibition of the innate immune response. The importance of the interferon system is highlighted by the multiple evasion strategies employed by viruses, such as herpesviruses [14], [55]. Here we present a report on ORF54, a conserved viral dUTPase, identified by our unbiased screen designed to isolate gammaherpesvirus ORFs involved in the inhibition of the type I IFN induced signaling pathway. Interestingly, ORF54 enzymatic activity proves dispensable for its anti-IFN function. Using KSHV ORF54 we were further able to demonstrate the anti-IFN role of ORF54 is conserved amongst KSHV and MHV-68. We have found that when high levels for ORF54 are present, either by ectopic expression or by infection with MHV-68, the total amount of cellular type I interferon receptor 1 is reduced, causing depression of the type I IFN response. Finally, through manipulation of the viral genome and murine host, we uncovered the biologically relevant function that ORF54 plays in not only blocking a critical component of the innate immune response, but also in persistent infection of gammaherpesviruses.
The most surprising and significant finding of our study is that the primary role of ORF54 during MHV-68 infection of cells and mice appears to be its anti-IFN function rather than its dUTPase activity. In three separate cell culture systems where type I IFN signaling is functioning, we demonstrate attenuation of the 54Stop virus, but not 54DM, a dUTPase-null virus (Figure 6). None of the defects seen with the 54Stop virus are observed with 54DM infection, indicating that the dUTPase activity is dispensable for the role of ORF54 during MHV-68 replication in cultured cells. In mice, while the 54Stop virus has a relatively normal productive infection in the lung, it shows a very strong deficit in spleen latency that is not found in the dUTPase-null virus (Figure 7C, 7D). However, this deficit is largely rescued in IFNAR−/− mice (Figure 7E), supporting the conclusion that it is the anti-IFN but not the dUTPase activity of ORF54 that plays a major role for MHV-68 establishment of latent infection of mice.
Across eukaryotes, prokaryotes, and viruses, dUTPases are highly conserved proteins, especially at the structural level [56]. Most eukaryotic dUTPases have 5 conserved motifs ordered 1–5 from the N- to C- terminal [45], [48], [49]. Functional dUTPases are formed as homotrimers containing three complete active sites formed from motifs 1–5 of each protein [45], [49]. However, the herpesvirus monomer is an active dUTPase with one active site formed by motifs 1–5 [48], [57]. The herpesvirus dUTPase protein sequence is twice as large as the cellular, with the C-terminal half containing domains 1, 2, 4, and 5 [48] and a herpesvirus unique domain called motif 6 between motifs 2 and 4. Towards the center of the protein, there is an actual motif 3, that was maintained after the loss of motifs 1, 2, 4, and 5 from the N-terminal portion [45], [48]. By examining the ability of murine cellular dUTPase to inhibit type I IFN signaling in our reporter assay, we found that dUTPase activity alone does not diminish IFN-α induced activation of ISRE, while an ORF54 dUTPase-null mutant still does (Figure 1). Therefore, it is possible that the N-terminal half of ORF54 may have evolved the additional function of type I IFN inhibition. For the gammaherpesviruses MHV-68, KSHV, and EBV, the ORF54 motif 3 starts at amino acid 82, 78, and 73, respectively. Since the motifs required for active site formation are found by the C-terminal halves of each protein, the N-terminal portion upstream of motif 3 in herpesviral dUTPases contain no recognizable or conserved sequences from cellular dUTPases. The sequences in the N-terminal half of ORF54 demonstrate some conservation between the gammaherpesviruses examined, but further studies designed to functionally map ORF54 anti-IFN activity are required to identify such a domain. ORF54 is an early protein [58] that, like other viral dUTPases, is expressed in both the cytoplasm and nucleus of infected cells, while cellular dUTPases are only found in the nucleus and mitochondria of cells [59]. Using a recombinant MHV-68 with a FLAG epitope tag on ORF54 to study its expression kinetics, we found that during early stages of infection ORF54 appears evenly distributed, but has preference for the nucleus at late stages of virus replication when cytopathic effect is obvious (Figure S3). As the MHV-68 genome is replicated in the nucleus, perhaps ORF54 functions as a dUTPase in the nucleus to help maintain genomic integrity and as an inhibitor of type I IFN signaling in the cytoplasm. It would be informative to identify ORF54 protein interaction partners in both compartments and at the various stages of virus infection.
Interestingly, beta- and gammaherpesviruses contain several ORFs with dUTPase-related domains that demonstrate strong divergence, where besides their catalytic motifs they exhibit different functions [48], some involving innate immune responses. Betaherpesviruses are unique in that many do not have a single ORF with dUTPase catalytic activity [60], [61]. However, several betaherpesvirus ORFs exist with dUTPase-related domains, such as UL72, UL82, UL83, UL84, and UL31 [48]. Human cytomegalovirus (HCMV) encodes pp65, also called UL83, that contains the dUTPase-related domain motif 6 [48]. HCMV pp65 expression led to the inhibition of phosphorylation of IRF3 and to its sequestration in the cytoplasm following induction of the IFN pathway [62], [63]. In gammaherpesviruses, although ORF54 is the only functional dUTPase, ORFs 10 and 11 both contain the dUTPase-related domain motif 6 [45], [48]. Here we showed in a transient transfection reporter assay that KSHV ORF54 is capable of reducing IFN-α responses (Figure 1). However, EBV ORF54 was found to induce expression of several pro- and anti-inflammatory cytokines when cells were treated with purified EBV ORF54 [64]–[67]. The biological significance of immune modulation by EBV ORF54 in the context of virus infection remains to be determined. KSHV ORF10 is a viral lytic protein that blocks IFN signaling by forming inhibitory complexes with JAK1 and TYK2 [35]. EBV ORF11 (LF2) was found to bind to IRF7 to inhibit dimerization and IRF7-mediated activation of type I IFN production, in a manner unrelated to its dUTPase domain [54]. Previously, we also found that MHV-68 ORF11 has a similar effect to ORF54 in inhibiting ISRE reporter induction by IFN-α [34]. Therefore, it is possible for large DNA viruses, such as herpesviruses, to utilize multiple proteins to inhibit type I interferon responses. This is further supported by the observation that STAT1 activation upon IFN-α treatment is not fully recovered in cells infected with the 54Stop virus (Figure 4B), indicating the presence of other viral proteins with overlapping anti-IFN functions.
Our results have revealed a previously unrecognized and critical anti-IFN function of ORF54, but they also raise a question about the biological role of its dUTPase activity. Numerous RNA and DNA viruses, such as retroviruses, poxviruses, and herpesviruses, encode a functional dUTPase in their genomes, suggesting its importance for the viral life cycle [45], [68], [69]. However, viral dUTPases are generally dispensable for virus replication [50], [70], [71], likely because cellular dUTPase is readily available and active in most dividing cells. Many cellular enzymes involved in DNA replication and nucleotide metabolism are strongly cell-cycle dependent. Therefore, it is presumably advantageous for the virus to encode its own enzymes for genome replication in terminally differentiated and non-dividing cells where the cellular counterparts may not be available. For example, herpesviruses encode several other enzymes in addition to dUTPase, such as thymidine kinase (TK), exonuclease, Uracil-DNA glycosylase (UNG), and ribonucleotide reductase (RR). For MHV-68, these viral enzymes are not required for in vitro virus replication in actively dividing cultured cells, but disruption of TK or RR leads to severe attenuation in acute productive infection in the lung as well as in splenic latency [50], [72], [73]. However, because site-specific mutations targeting the enzymatic activity of these proteins were not used in these studies, it cannot be completely excluded that other possible mechanisms unrelated to their known TK or RR functions account for the observed severe attenuation.
In this study, we constructed a dUTPase-null virus, 54DM, to study the specific contribution of the enzyme activity of ORF54. Interestingly, unlike with TK and RR, lack of dUTPase activity or even the entire ORF54 protein does not detrimentally retard virus replication in the lungs of infected mice. Previous studies in our lab have found that ORF54-null viruses constructed with a large transposon insertion cause a more significant defect on lytic replication in the lungs of infected mice [50]. This discrepancy may be due to the large disturbance each transposon causes in the highly compact MHV-68 viral genome, where most promoter regions overlap coding regions. The ORF54-null virus, 54Stop, shows a very strong defect in the infectious center assay that is not found with the dUTPase-null virus, implying that the lack of ORF54 enzymatic activity does not dramatically hinder the establishment of the MHV-68 latent load, while lack of the entire protein does. However, mice infected with 54DM do have a slight, but statistically insignificant, reduction in the amount of reactivated virus from infected splenocytes, at 2.8- and 2.4-fold defect compared to WT and 54R viruses, respectively (Figure 7C). This marginal defect of 54DM was not observed when the viral genome copy number was analyzed (Figure 7D), thus it is possible that the phenotype seen in the infectious center assay is due to the lack of dUTPase during reactivation. The 54Stop virus lacks any expression of ORF54, including its dUTPase activity, and the 2.8-fold reduced level of infectious centers compared to 54R in IFNAR−/− mice mirrors that seen with 54DM virus during reactivation in Balb/C mice. Indeed, as with 54DM in wild-type mice, 54Stop has nearly identical viral genome copies as 54R in the splenocytes of IFNAR−/− mice, again suggesting that the minor defect in the infectious center assay is due to reactivation. Although it was expected that dUTPase-null viruses grow normally in actively dividing cell culture, it is surprising that the only phenotype we observed in mouse infection is a moderate defect in reactivation from latency. However, our observation is not unique. Following in vivo infection, HSV-1 dUTPase mutants replicate like wild type in the footpad, sciatic nerve, and dorsal root ganglia; a defect of about 10- to 100-fold is only visible when the virus moves to the CNS spinal cord. These dUTPase mutants were able to establish latent infections but demonstrated a defect in reactivation from latency [74]. Taken together, our results indicate that the dUTPase activity of ORF54 does not have a significant role during MHV-68 productive infection in the lung or latency in the spleen. However, our interpretation for the maintenance of dUTPase in the viral genome is that dUTPase activity is likely required for genome stability over many generations. Furthermore, reactivation is considered much less effective than lytic replication in permissive cells, thus a less optimal replication environment may have a more profound effect on the reactivation process over time.
Gammaherpesviruses are characterized by their ability to establish latent infection in lymphocytes. Establishment of viral latency first requires the efficient infection of lymphocytes. By affecting persistent infection of gammaherpesviruses, the IFN response has far reaching effects on the viral life cycle and the establishment of latency, instead of only acting primarily on initial infection. Furthermore, the type I IFN response is critical in shaping the adaptive immune response [1], [2], [6], [9], [36]. Altering the balance between the host immune response and viral immune evasion genes could drastically affect the overall outcome of viral infection. The anti-IFN function of ORF54 appears to be relatively dispensable for lytic replication in vitro and in vivo, but absolutely required for the establishment of latency. As a whole, our data suggests that evading the type I IFN response is critical for the establishment of latent infection in lymphocytes and that ORF54 plays a major role in this evasion. The reduced latency observed in the infection of 54Stop may be due to the inability of the virus to replicate well in a particular cell type, such as lymphocytes, that is more sensitive or responsive to the anti-viral effects of IFN. Although several genes are likely required and may have some overlapping functions, lack of one of these genes, such as ORF54, retards establishment of latent infection. Our results in IFNAR−/− mice not only emphasize the importance of ORF54 in the type I IFN pathway by demonstrating a rescue in an infectious center assay, but also hint at the necessity of this host pathway in blocking the establishment of latent infection in lymphocytes. However, the mechanisms by which the virus spreads from the inoculation and lytic replication sites to the latency compartment remain largely elusive. Thus, although our data suggests its importance, we are currently unable to isolate at which stage during this viral spread the anti-IFN function of ORF54 is required. Analysis with a detailed time course and different cell types is required to further understand the interaction between type I interferon responses and latency establishment.
Defining immune evasion strategies employed by MHV-68 will allow better design of a live attenuated vaccine to human gammaherpesviruses KSHV and EBV. One strategy to increase the success of a vaccine is to limit the establishment of latency and increase its immunogenicity by removal of viral immune evasion genes [75]. Because ORF54 is not required for virus replication, plays a role in type I IFN inhibition, and is necessary for the efficient establishment of latency, a virus lacking ORF54, as well as other immune evasion genes and genes required for latency, is a promising vaccination strategy.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the Public Health Service (National Institutes of Health). The protocol was approved by the Institutional Review Board and Animal Research Committee of the University of California Los Angeles (assurance of compliance number: A3196-01). All procedures were performed under ketamine and xylazine anesthesia and all efforts were made to minimize suffering.
293T and Vero cells were cultured in complete Dulbecco's modified Eagle medium (DMEM) containing 10% FBS. NIH3T3 cells were cultured in DMEM containing 10% BCS. Bone-marrow derived macrophages (BMDM) from both wild-type and IFNAR−/− mice were immortalized by v-raf and c-myc oncogenes, cultured in RPMI containing 10 mM HEPES, 10% FBS, and MCSF, and were a kind gift from Shankar Iyer (Genhong Cheng Lab, UCLA). All culture medium was also supplemented with penicillin and streptomycin and all cells were cultured at 37°C with 5% CO2.
dUTPase assay was adapted and modified from previous methods used to measure enzymatic activity [44], [59], [76]. Briefly, each potential FLAG-tagged dUTPase construct was transfected into 293T cells using Lipofectamine 2000 (Invitrogen). 48 hours post transfection, cells were lysed with buffer containing 100 mM Tris pH 7.5, 50 mM NaCl, and 1% NP-40, supplemented with 1 mM PMSF, 1 µg/mL Aprotinin, 1 µg/mL Leupeptin, 1 µg/mL Pepstatin A, 1 mM Na3VO4, and 1 mM NaF. Lysates were kept on ice for 10 minutes and clarified with a high speed spin at 4°C for 15 minutes. FLAG-tagged proteins were immunoprecipitated by incubating lysates with Protein G Sepharose (GE Healthcare 17061801) and 4 µg of anti-FLAG antibody (Sigma F3165), followed by elution from the antibody by incubation with 20 µg of 3×FLAG peptide (Sigma F4799). 5 ul of purified FLAG-tagged proteins were incubated at 37°C with 5 ul of 5 mM dUTP (Promega U1191) for 0, 2, 5, 10, 15, 20, and 25 hours in 10 ul of 2× reaction buffer comprised of 100 mM Tris pH 7.5, 20 mM MgCl2, 20 mM DTT, and 0.2 mg/mL BSA. Reactions were terminated by freezing. Potentially digested dUTP are at a maximum final concentration of 1.25 mM, lower if digested. PCR was conducted using MHV-68 BAC DNA as a template, with primers to amplify a region in ORF57 (forward primer: 5′-ACTGAAACCTCGCAGAGGTCC-3′ and reverse primer 5′-GCACGGTGCAATGTGTCACAG-3′) using the potentially digested dUTP alongside dATP, dCTP, and dGTP. Cycle conditions were 95°C 5 min; 95°C 30 seconds, 60°C 30 seconds, 72°C 30 seconds for 35 cycles; 72°C 10 minutes; hold at 4°C. PCR products were run on a 3% agarose gel.
The recA+ Escherichia coli (E. coli) strain GS500 harboring a BAC containing the wild-type MHV-68 genome was used to construct recombinant MHV-68 by allelic exchange with the conjugation competent E. coli strain GS111 containing a suicide shuttle plasmid pGS284, as previously described [34], [77], [78]. For each recombinant MHV-68, overlap extension PCR was used to construct a unique shuttle plasmid pGS284 harboring the desired mutation with a 500 base pair flanking region. To screen for the correct mutation, restriction enzyme digests were performed on PCR products amplified with the outer primer sets on the BAC MHV-68 clones. WT MHV-68 and 54R contain a BamHI restriction enzyme site that is altered to a HindIII site in 54Stop. The base pair changes in 54DM allow the introduction of an AseI site that is absent in WT MHV-68. Following selection of the desired recombinant clone, the MHV-68 BAC DNA was purified and transiently transfected with Lipofectamine 2000 into 293T cells with an equal amount of a plasmid expressing Cre recombinase to remove the BAC sequence. Three days post transfection, a single viral clone was isolated by limiting dilution and propagated for future studies. Produced viruses were quantified by plaque assay and limiting dilution. Genomic integrity of the final recombinant viruses was analyzed by SmaI and EcoRI restriction enzyme digestion and Southern blot with DIG-labeled DNA probes to the whole genome or 6 kb of the left end (Roche 11585614910). Primers used for the construction of each shuttle plasmid are listed 5′ to 3′ as follows: 54R and 54Stop, outer primers: CAGGACAGATCTCACTAGACACTGTGACTATAGAC and CTGTCCCCCGGGTAGCAGACACAGGTCCTCAG; 54Stop, inner primers: CACATAACTGAGTAAGCTTACCAGACTAAAATTTCAAAACATCAC and TGGTAAGCTTACTCAGTTATGTGCCTTGTGTAGTTAGGGCC; 54DM, outer primers: GAAGAGATCTCACTCCCCCACTGAGGGACGTATGTGTCAGC and GTTGGCTAGCCAATTTCAGACTTGTCTGCAGCTTCGTGCGGAACCCTAATAAAC; 54DM, inner primers: ATTAATCAGTCCAGTTGCGCCGGTGACGAGACTGTT and GCAACTGGACTGATTAATCCCGGCTACCGGGGTGAAAT.
MHV-68 concentrated virus was titered using plaque assay and limiting dilution. For plaque assay, 10-fold serial dilutions of each virus were incubated on a monolayer of Vero cells for 1 hour. The infected cells were overlaid with 5% methylcellulose DMEM. At 6 days post infection, cells were fixed with 2% crystal violet in 20% ethanol. Plaques were counted at the optimum dilution to calculate virus titer. To determine the 50% Tissue Culture Infectious Dose (TCID50), 12 wells seeded with Vero cells in a 96-well plate were infected by 100 ul of each 10-fold serial dilution. 7 to 10 days post infection, every well was scored for cytopathic effect (CPE). TCID50 was determined by the calculation TCID50/mL = 10×10∧[(highest dilution with 100% CPE)−[0.5+((# of wells with CPE/total wells) in next highest dilution)+(each following dilution's CPE fraction)]]. Viral titers of multiple step growth curves were quantified by plaque assay.
Wild-type MHV-68 BAC was constructed as previously described [34]. The firefly luciferase reporter driven by the interferon-stimulated response element (ISRE_firefly-luciferase) and the renilla luciferase reporter driven by the promoter of housekeeping gene phosphoglycerate kinase (PGK_renilla-luciferase) were kind gifts from Dr. Genhong Cheng and Dr. Lily Wu, respectively (UCLA). All MHV-68 open reading frames were cloned into pENTR (Invitrogen) by PCR amplification from MHV-68 DNA, then transferred to a modified destination vector resulting in a 3×FLAG epitope tag at the N-terminus. Specifically, for MHV-68 ORF54, primers used are 5′-CCGAGCGAATTCAATGAAAGTGGAATACTCCTTTGTG-3′ and 5′-GCTCGGGGTACCTTAATTCACCCCACTTGACCCAAAC-3′. To construct ORF54 MHV-68 H80A/D85N by overlap extension PCR, the same inner primers as for 54DM virus were used, the outer primers used are 5′-CCGAGCGAATTCAATGAAAGTGGAATACTCCTTTGTG-3′ and 5′- GCTCGGGGTACCATTCACCCCACTTGACCCAAAC -3′. Wild-type ORF54 and ORF54 H80A/D85N have slightly different molecular weights due to the cloning process of FLAG-ORF54 H80A/D85N. A modified Gateway cloning system (Invitrogen) was utilized to generate entry clones of the target protein, then recombined by an LR reaction into the same destination vector as FLAG-ORF54. This process inserted 6 additional amino acids to the N-terminus of the protein (from the multiple cloning site) after the 3×FLAG tag and before the ORF54 H80A/D85N ATG start site. Also, the H80A/D85N clone utilizes a STOP codon in the destination vector while the FLAG-ORF54 clone includes the endogenous STOP codon. Therefore, 10 additional amino acids (from the multiple cloning sites) are added to the C-terminus of FLAG-ORF54 H80A/D85N. In total, the FLAG-ORF54 H80A/D85N protein has an additional 16 amino acids, or approximately 1.7 kDa additional molecular weight, compared to the wild type FLAG-ORF54.These changes demonstrated no effect on ORF54 or ORF54 H80A/D85N function. KSHV ORF54 was cloned from the a KSHV BAC construct, a kind gift from Dr. Jae Jung (University of Southern California). The primers used are 5′-TAATGGATCCATGAACAACCGCCGAGGCTC-3′ and 5′-TAATGTCGACCTAAAACCCAGACGACCCCAG-3′.
293T cells were seeded at 5×104 cells per well in a 48-well plate 16 hours prior to transfection. Cells in each well were transfected using Lipofectamine 2000 (Invitrogen), with 20 ng of ISRE_firefly-luciferase, 5 ng of PGK_renilla-luciferase, 200 ng of viral ORF or vector control, and 175 ng of filler DNA. 24 hours post transfection, identically transfected wells were left untreated or were treated with 3×104 units of human interferon-α (both human and mouse IFN-α are purchased from pbl Interferon Source). 24 hours post treatment, both firefly and renilla luciferase activity was measured (Promega Dual-Luciferase Assay Kit). Fold activation was calculated by first normalizing all values to their internal renilla luciferase control, and then by dividing luciferase activity in the treated samples by that of the untreated samples. Expression of viral ORFs was determined by immunoblot against the FLAG epitope expressed at the N-terminus of each of the ORFs in our MHV-68 expression library.
NIH3T3 cells were seeded in 24-well plates and infected with FLAG-54 MHV-68 at an MOI of 1. Cells were fixed and permeabilized with treatment for 30 minutes at room temperature with 100% methanol. Cells were then washed in three changes of PBS and stored in PBS until staining. Mouse anti-FLAG M2 (Sigma F3165) was incubated with cells overnight for 16 hours at a dilution of 1∶750, and Alexa Fluor 594 goat-anti mouse IgG (Invitrogen A11005) at a dilution of 1∶1000 was incubated with cells for 1 hour. Hoechst dye was added for 10 minutes prior to analysis.
Cells were lysed for 10 minutes on ice in lysis buffer (50 mM Tris pH 7.5, 1% NP-40, 0.25% sodium deoxycholate, 150 mM NaCl, 1 mM EDTA) supplemented with 1 mM PMSF, 1 mM Na3VO4, and 1 mM NaF. Lysates were then combined with 4× protein sample buffer (0.25 M Tris pH 6.8, 8% SDS, 40% glycerin, 20% β-mercaptoethanol, 0.008% Bromophenol blue), sonicated, and boiled for 10 minutes prior to loading on a 10% polyacrylamide gel. Membranes were stripped with Multi-Western Stripping buffer (Bioland Scientific), prior to re-probing with each subsequent antibody. The antibodies used in this study were rabbit anti-human phosphoSTAT1 (Cell Signaling 9167), rabbit anti-mouse phosphoSTAT1 (Millipore 07307), rabbit anti-total STAT1 (Cell Signaling, 9175S), mouse anti-FLAG M2 (Sigma F3165), rabbit anti-human IFNAR1 (Abcam, ab45172), mouse anti-β-actin (Sigma A5316), rabbit anti-IGF-1 Receptor β (Cell Signaling 3027S), rabbit anti-IFNAR2 (Novus Biologicals 31665), rabbit anti-phosphoErk1/2 (Cell Signaling 4376), and rabbit anti-IFIT2 (Abcam 55837). Rabbit serum against viral lytic protein ORF65 was derived in our lab. Secondary antibodies conjugated to HRP were donkey anti-rabbit IgG (GE Healthcare NA934V) and sheep anti-mouse IgG (GE Healthcare NXA931).
Total RNA was extracted from cells by RNeasy Mini Kit (Qiagen) and reverse transcribed into cDNA by qScript cDNA Synthesis Kit (Quantas). Primers used in RT-PCR to quantify cellular transcripts are as follows: actin: 5′-GTATCCTGACCCTGAAGTACC-3′ and 5′-TGAAGGTCTCAAACATGATCT-3′; human IFNAR1: 5′- AACAGGAGCGATGAGTCTGTC-3′ and 5′- TGCGAAATGGTGTAAATGAGTCA-3′; murine IFNAR1: 5′- AGACGAGGCGAAGTGGTTAAA-3′ and 5′- GCTCTGACACGAAACTGTGTTTT-3′; murine MX1: 5′-GAATAATCTGTGCAGGCACTATGA-3′ and 5′-CTCTCCACTCCTCTCCTTCTTTC-3′; murine IFIT1: 5′-TGCTTTGCGAAGGCTCTGAAA-3′ and 5′-TTCTGGATTTAACCGGACAGC-3′; murine IFIT3: 5′-AGTGAGGTCAACCGGGAATCT-3′ and 5′-TCTAGGTGCTTTATGTAGGCCA-3′.
All in vivo procedures were performed according to protocols approved by the University of California, Los Angeles, the Animal Research Committee, and the Institutional Review Board. Balb/C mice were purchased from Charles River Laboratories. IFNAR−/− mice were a kind gift from Dr. Genhong Cheng at UCLA. Mice were intranasally infected with 500 pfu under sedation by I.P. injection of 2 mg ketamine and 0.04 mg xylazine. At 5 and 7 days post infection, mice were sacrificed and lung tissue was harvested in 1 mL of complete DMEM. Lung tissue was homogenized to measure the viral titer by plaque assay. At 14 days post infection, mice were sacrificed for infectious center assay of splenocytes. Briefly, a single cell suspension was isolated from the spleen of each infected animal. The splenocytes were co-cultured for one day with a monolayer of Vero cells, then overlaid with 5% methylcellulose DMEM for 6 additional days prior to fixing cells with 2% crystal violet in 20% ethanol. Each viral plaque reflects MHV-68 reactivated from splenocytes. Plaques were counted at the optimum dilution and number of infectious centers calculated per 1×107 splenocytes. To determine viral genome copies, total genomic DNA for quantitative-PCR was harvested from lung lysates and splenocytes using QiaAmp DNA Mini Kit (Qiagen). The PCR reaction was comprised of 100 ng of total genomic DNA as a template and the primers used were 5′-ACCTTGAAACCCGTGAAGG-3′ and 5′-CATCTGCCACGACCTGAGAT-3′.
Swiss-Prot accession numbers for the proteins/genes used in this study are as follows: MHV-68 ORF54, P88991; KSHV ORF54, P88942; MHV-68 ORF48, P88986; and murine cellular dUTPase (DUT), Q9CQ43. The MHV-68 viral genome GenBank accession number is U97553.2.
|
10.1371/journal.pbio.1002248 | Fitness Benefits of Mate Choice for Compatibility in a Socially Monogamous Species | Research on mate choice has primarily focused on preferences for quality indicators, assuming that all individuals show consensus about who is the most attractive. However, in some species, mating preferences seem largely individual-specific, suggesting that they might target genetic or behavioral compatibility. Few studies have quantified the fitness consequences of allowing versus preventing such idiosyncratic mate choice. Here, we report on an experiment that controls for variation in overall partner quality and show that zebra finch (Taeniopygia guttata) pairs that resulted from free mate choice achieved a 37% higher reproductive success than pairs that were forced to mate. Cross-fostering of freshly laid eggs showed that embryo mortality (before hatching) primarily depended on the identity of the genetic parents, whereas offspring mortality during the rearing period depended on foster-parent identity. Therefore, preventing mate choice should lead to an increase in embryo mortality if mate choice targets genetic compatibility (for embryo viability), and to an increase in offspring mortality if mate choice targets behavioral compatibility (for better rearing). We found that pairs from both treatments showed equal rates of embryo mortality, but chosen pairs were better at raising offspring. These results thus support the behavioral, but not the genetic, compatibility hypothesis. Further exploratory analyses reveal several differences in behavior and fitness components between “free-choice” and “forced” pairs.
| The last half century has seen a tremendous interest in the study of mate choice and the evolution of traits that make individuals attractive to others. In some species, however, individuals can differ substantially in who they find attractive, and this variation has typically been interpreted as “mate choice for compatibility.” Here, we quantify the benefits of such mate choice in a socially monogamous passerine bird, the zebra finch. We found that pairs that resulted from free mate choice achieved a 37% higher reproductive success than pairs that were forced to mate with a randomly assigned individual. Forced pairs suffered from increased failure to fertilize eggs and from increased mortality of hatched offspring. In females, we observed a reduced readiness to copulate with the assigned partner, while males that were force‐paired showed reduced parental care and increased activity in courting extra‐pair females. These findings support the hypothesis that zebra finches choose mates on the basis of behavioral compatibility. In contrast, it appears that zebra finches have not evolved a mechanism that would allow them to select a partner with whom they could minimize the rate of embryo mortality. This argues against mate choice for genetic compatibility.
| The evolution of mate choice has been the focus of much research, and many studies have attempted, with a variety of experimental approaches, to measure the fitness benefits gained by choosy individuals (e.g., [1,2–7]). Those benefits can be either direct, if offspring quality or quantity is increased due to the partner’s behavior (including reproductive investment), or indirect, if offspring quality is improved by the genetic contribution of the partner. To date, the central debate has been about (i) the relative importance of direct versus indirect fitness benefits arising from the overall quality of the chosen partner (i.e., good parent versus good genes; Fig 1, vertical black arrow) [8,9], or about (ii) the relative importance of the two types of indirect benefits (i.e., good genes versus compatible genes; Fig 1, horizontal black arrow) [10–13].
Several studies on mate choice have shown that, in some species, mating preferences can be largely specific to the individual [14–19]. Such mate preferences may function to maximize offspring viability by bringing together compatible combinations of genes (top right in Fig 1). However, the alternative hypothesis that mate choice could lead to direct benefits arising from the phenotypic (e.g., behavioral) compatibility of the two partners (bottom right in Fig 1) has received only little attention [20–26], despite suggestive evidence that the combination of both parents’ behaviors or other phenotypes can affect breeding success. Compatible partners could, for instance, be better at coordinating tasks, at sharing them or at complementing each other’s performance on various tasks [23–25,27–31], or they might simply be more effective at stimulating one another’s reproductive investment [32–34]. Mate choice for such behavioral compatibility might be especially important in species with intense bi-parental brood care and with long-lasting, monogamous pair bonds, like humans or many bird species.
Previous experiments that aimed to quantify the fitness benefits of mate choice arising from partner compatibility typically compared two categories of individuals: those paired up with their preferred partner versus those that were given a non-preferred partner [35–43], or a random partner from the population [44,45]. The problem with this approach is that the effects of individual quality and pair compatibility are confounded, because only the force-paired group includes individuals that might never have been chosen (i.e., low-quality individuals). Some studies addressed this issue by presenting evidence that the rejection of a particular mate depended on the choosing individual’s identity [46], or that non-preferred and preferred individuals did not differ in morphological traits [43]. Other studies have compared the reproductive success of a choosing individual (paired with its preferred partner) with the reproductive success of a naïve individual paired with that same (or another) preferred individual [2,47–50]. In the latter design, chosen and assigned partners are on average of equal quality. However, choosing individuals were often discarded if they did not meet a certain criterion regarding their strength of preference ([2,48,49], but see [47,50]). If choosiness is associated with an individual’s quality [17,51], the selected subset of choosing individuals might differ in quality from the random pool of naïve individuals to which they are compared. However, in two experimental studies on invertebrates, none of the above issues apply; these studies found no fitness benefit of mate choice for compatibility [47,50].
Here, we employ an experimental design somewhat similar to [50], to eliminate the effect of mate quality: we compare the fitness of individuals that bred with their preferred partner with those that obtained, after having expressed their preference, the preferred partner of another individual. The main aim of this study is thus to quantify the benefits of mate choice that arise from partner compatibility, while circumventing confounding effects of variation in partner quality. The second aim of our study is to tease apart indirect compatibility advantages (compatibility of parental genes expressed in the offspring) from direct ones (parental phenotypic compatibility), using a model species in which these benefits of mate choice can be disentangled.
The zebra finch (Taeniopygia guttata) is a socially monogamous species with biparental care, in which partners mate for life [52]. In this species, female mate preferences are predominantly individual-specific (i.e., females show little consensus regarding which male is the most attractive) [15,53–57], suggesting that they may target genetic or behavioral compatibility. In captive and wild populations, high rates of embryo and offspring mortality are found, even in the absence of inbreeding [25,52,58–60]. Cross-fostering of freshly laid eggs (see [61,62] and S1 Text) showed that most of the variance in embryo mortality (before hatching) is explained by the identity of the genetic parents rather than the foster‐parents (based on n = 1,529 fertilized eggs, S1 Table), whereas most of the variance in offspring mortality (after hatching) is explained by foster-parent rather than genetic parent identity (n = 1106 offspring, S1 Table). Based on these results, we assume that, in zebra finches, embryo mortality primarily reflects genetic incompatibility (as in other species [63,64]), while offspring mortality primarily results from the behavior of the caring parents (here broadly referred to as “behavioral incompatibility”). Experimentally preventing mate choice should thus lead to an increase in embryo mortality if mate choice is targeting genetic compatibility, and to an increase in offspring mortality if mate choice is targeting behavioral compatibility. Alternatively, if individual-specific mate preferences only reflect indecision by the animal or measurement error [15], preventing mate choice would have no fitness consequences.
We studied 160 bachelor birds from a recently wild-derived population of zebra finches. Each individual could freely choose a partner from a group of 20 individuals of the opposite sex during a long, nonbreeding season. This setup reflects the natural situation in the sense that zebra finches are opportunistic breeders and do not reproduce if the environment is not suitable, but they still form life-long pair bonds irrespective of breeding opportunities. Furthermore, the species is gregarious, such that individuals have many potential partners to choose from. Pairs were identified by the occurrence of allopreening because we found that this best reflects mutual preferences rather than being the outcome of intra-sexual competition (see S2 Text). We hereafter focus on female preferences; however, because observed allopreening preferences were mainly reciprocal between females and males (see also S2 Text), any observed effects of the experimental treatment described below could be due to females, males, or both sexes not being able to breed with their (most) preferred partner. Females from these pairs were alternately assigned to one of two treatments: half of them were allowed to stay with their chosen partner, while the other half were force-paired with the chosen partner of another female from the same aviary. This ensured that, on average, individuals of both treatments were of the same quality, even if assortative pairing for quality had happened due to intra-sexual competition. All pairs were then placed in individual cages for a few months to enforce pair-bonding in the non-chosen pairs (force-pairing is effective in this species if assigned mates are co-housed in a cage for long enough, see “Methods”). After this period in separate cages, pairs were given the opportunity to breed for about five consecutive months (allowing about three successful broods) in communal aviaries, each containing three pairs from each treatment group. This entire procedure was repeated a second time with the same birds (i.e., free choice during a nonbreeding period, force pairing in cages, and breeding in communal aviaries). This was planned a priori to obtain repeated measurements on individuals under different pairing conditions with a large enough sample size to allow the detection of weak effects. For the second breeding period, two-thirds of the pairs from the first breeding period were broken up; individuals chose a new partner and were either assigned to the same or the other treatment. The other third of the pairs were allowed to keep their partner (chosen or non-chosen) from the first breeding period. This allowed us to better control for any effects of pair-bond duration in statistical models comparing chosen and non-chosen pairs, given that pair-bond formation in chosen pairs systematically started earlier (during the free choice period) than in non-chosen pairs (in cage). In total, we monitored behavior and reproductive success of 46 chosen pairs (C) and 38 non-chosen pairs (NC). Measures of reproductive success were based on paternity analyses that included dead embryos, dead chicks, and surviving offspring. Behavior was scored based on direct observations (285 h) and video recordings (1,424 h).
When released into communal breeding aviaries, each of which contained three chosen and three arranged pairs, the proportion of pairs that stayed together differed between treatment groups (C: 46 out of 50 pairs, NC: 38 out of 50 pairs; Fisher’s exact test p = 0.05). This suggests that birds that were force-paired with a partner they did not choose were more reluctant to breed together. However, this differential rate of divorce between the treatment groups is unlikely to induce a bias in our experimental results for the following reasons: (1) Individuals that divorced during one of the breeding periods did not appear to differ in intrinsic quality from individuals that stayed together, as judged from a comparison of reproductive success obtained in the other season when they did not divorce (relative fitness, mean ± standard error [SE]; divorced: 1.06 ± 0.13, n = 23; not divorced: 0.98 ± 0.08, n = 59; general mixed effect model accounting for treatment, birds matched for year, p = 0.58). (2) To induce a bias that is large enough to explain our results, the difference in intrinsic quality between divorcees and the remaining population would need to be unrealistically large (for further details, see simulation in S3 Text).
Only those pairs that remained together were considered for further analyses. Parameter estimates of traits for the two treatments (C versus NC) are given for each general and generalized linear mixed-effect model. Other relevant statistics, as well as the structure of the models, are provided in Table 1 (referred to as “T1-test #”).
We calculated relative fitness of individuals as the total number of genetic offspring produced in a given breeding period that reached independence (35 d old), relative to the number produced in the same period by the other individuals in the same aviary. Males of chosen pairs had a 45% higher relative fitness than males of non-chosen pairs (C = 1.16, NC = 0.80, p = 0.03, n = 84 male breeding periods, see T1-1 for model details, Fig 2). Females of chosen pairs had a 30% higher relative fitness than those of non-chosen pairs, but the difference was not significant (C = 1.09, NC = 0.84, p = 0.12, n = 84, T1-2, Fig 2). The difference between the sexes was not significant (interaction between treatment and sex: p = 0.36) and resulted from extra-pair paternity (see below). Thus, on average, individuals from the chosen pairs had a 37% higher fitness. This difference in fitness was not due to differences in pair bond duration between the treatments groups, as this covariate did not correlate with fitness (non-significant trends against the expectation, Table 2: T2-1 and T2-2) and was therefore removed from the models T1-1 and T1-2.
The overall fitness difference observed was not due to differential investment in egg production by the females of the two treatment groups (total number of eggs laid: C = 13.5, NC = 14.4, p = 0.56, n = 84, T1-3). However, non-chosen pairs tended to have a higher proportion of disappeared or buried eggs (C = 12%, NC = 19%, p = 0.07, n = 1172 eggs laid, T1-4), and had significantly more clutches that contained infertile eggs (C = 8%, NC = 23%, p = 0.01, n = 216 clutches, T1-8).
To test the genetic incompatibility hypothesis, we compared the proportion of dead embryos between treatment groups, considering all fertilized and incubated eggs. We only included the genetic eggs of each pair, that is, we excluded all extra-pair young (9% of the eggs), but included eggs that were dumped into the nest of other pairs (13% of the genotyped eggs). Note that removing dumped eggs (potentially suffering higher rate of embryo mortality [65]) from the analysis did not change the conclusions. Furthermore, we only included eggs that were incubated without interruption, excluding those that were buried in the nest material before incubation was completed (based on daily nest checks). The rate of embryo mortality did not differ between chosen and non-chosen pairs (C = 20%, NC = 22%, p = 0.68, n = 707 fertilized eggs, T1-5, Fig 3A).
To test the behavioral compatibility hypothesis, we compared the proportion of dead offspring between treatment groups, considering all hatched eggs in a pair’s nest (including extra-pair offspring and hatchlings from dumped eggs). Offspring mortality was significantly higher when chicks were reared by non-chosen pairs (C = 32%, NC = 52%, p = 0.03, n = 594 hatched eggs, T1-6, Fig 3B). Pair bond duration did not influence this result (T2-6). The probability of survival may also decrease if the offspring is unrelated to one or both of the parents. To check this, we added the status of the offspring (within-pair versus extra-pair young, offspring from dumped versus not dumped egg) into model T1-6. We found that the treatment effect was still significant (p = 0.045), but offspring status was not (mortality of within-pair young = 38%, extra-pair young = 55%, p = 0.15; dumped = 40%, non-dumped = 39%, p = 0.91; underlying data can be found in S1 Data).
Many studies have attempted to quantify the benefits of mate choice [2–7,35–46,48–50,66,67], but only a few have quantified the fitness benefits of mate choice for compatibility while excluding quality benefits (see e.g. [46] and [50]) (Fig 1). Our experimental design allowed us to circumvent the potentially confounding effect of mate quality by comparing pairs of individuals that chose each other with pairs that were composed of random individuals who did not choose each other, but had both been chosen by another individual. Pairs that formed through free mate choice had a 37% higher fitness than pairs that were “forced” experimentally (Fig 2). This suggests that it is unlikely that the between-individual disagreement about mate attractiveness simply reflects indecision or measurement error. Our results suggest instead that individual-specific mate preferences lead to significant fitness consequences.
Our study system, furthermore, allowed us to disentangle direct (behavioral) benefits of mate choice from indirect (genetic) benefits (Fig 1). Chosen pairs, compared to arranged ones, had a 38% lower rate of offspring mortality (Fig 3B). Under the assumption that offspring mortality systematically depends on parental behavior, this result supports the hypothesis of mate choice for behavioral compatibility. Ideally, our experiment should be repeated while cross-fostering eggs to exclude confounding factors. Indeed, our conclusions depend on the generalizability of the results from our previous study (S1 Text). The finding that offspring mortality after hatching primarily depends on the rearing parents and not on the genetic parents (S1 Table) can likely be generalized from our previous cross-fostering experiment to this study; in both studies, many offspring apparently died from starvation, and an offspring that is not fed will die irrespective of its genetic quality. Finally, chosen and arranged pairs had an equal rate of embryo mortality (Fig 3A). Given that embryo mortality primarily depends on the genetic parents and less on the incubating parents (S1 Table), this result argues against the hypothesis of mate choice for genetic compatibility. At least, our results suggest that individuals did not select a partner with whom they would have minimized the rate of embryo mortality.
Several earlier experimental studies favored the genetic compatibility hypothesis based on the observation that offspring from “free-choice” pairs had a higher viability than those from “forced” pairs [35–37,40,43,46,66]. However, in these experiments females were forced to mate with random males from the population or with non-preferred males, some of which may have been of lower absolute quality (but see [46]). Hence, the previously observed effects on offspring viability may be explained by differences in both genetic quality and compatibility. In general, mate choice for genetic compatibility may not easily evolve, because it requires that the incompatibility-causing loci are tightly linked (e.g., via pleiotropy) to a detectable phenotype and to a mechanism ensuring the appropriate assortative or disassortative preference [68]. At least in zebra finches, such a complex adaptation that would allow them to minimize embryo mortality by choosing a genetically compatible partner, does not seem to exist (this study). Similarly, inbreeding avoidance is absent in this species when birds can only judge genetic similarity per se [61] (although it does take place when siblings are familiar with each other [69]).
Our results are consistent with the hypothesis that behavioral compatibility between the pair members leads to benefits of mate choice. This could come about through different mechanisms: the emerging behaviors of a pair in terms of coordination or complementarity [23,24,27–29], and/or the individual-specific stimulation of a partner’s sensory system leading to a greater investment in reproduction [32–34]. Currently it is unclear which of these factors leads to the observed variation in parental care compatibility, and it is also unclear to what extent there is a genetic basis for this variation in compatibility. In the following, we discuss our exploratory analyses on fitness components and behaviors of “free-choice” and “forced” pairs, to provide testable ideas about how such behavioral compatibility benefits could arise.
We found that non-chosen pairs (1) more often had clutches with infertile eggs, (2) had more offspring dying at an early stage (presumably from starvation), and (3) tended to have more eggs that disappeared (presumably due to poorer care and nest defense). These effects on components of fitness may be due to differences in the behavior of chosen and non-chosen pairs.
The most prominent behavioral differences were that (a) females with assigned partners responded less positively to within-pair courtship and they tended to copulate less frequently with their partner, and (b) males with assigned partners showed poorer nest attendance during the egg hatching period. The females’ reduced tendency to participate in within-pair courtship and copulation when in a “forced” pair may explain the higher incidence of infertile eggs. Indeed, in a previous experiment in which continuous video recording allowed us to witness about 80% of all copulations over a 4-mo period (partly reported in [53]) we found that the probability of laying an infertile egg declined significantly with the number of copulations witnessed during the 10 d prior to egg laying (p = 0.04, n = 376 eggs laid by 31 females, estimates: 27% infertile at 0 copulations versus 15% infertile at the median of 5 copulations; underlying data can be found in S1 Data). Alternatively, apparently infertile eggs may in fact represent cases of very early embryo mortality. This seems unlikely because egg fertility scores in zebra finches were tightly linked to the number of sperm that reached the egg [70]. Likewise, the lower nest attendance during hatching by males in non-chosen pairs could indicate a reduced motivation to care for the young or defend the nest when in a forced partnership, leading to greater offspring mortality and egg loss. Consequently, the results of these exploratory analyses further support the behavioral compatibility hypothesis.
If males and females in “forced” pairs indeed invest less in reproduction (copulation or care), as our results suggest, the question remains why. Reduced investment by members of “forced” pairs could be a long-term effect of a single stressful event (trauma), namely the loss of the chosen partner (an event that could also happen in the wild due to predation [71]). This explanation seems unlikely, however, because fitness was affected by the treatment per se and not by the number of partner losses experienced by an individual (see scheme in “Methods”) when both factors were fitted within one model (males: treatment p = 0.02, number of mate losses p = 0.63; females: treatment p = 0.06, number of mate losses p = 0.35). Alternatively, being forced to breed with a non-preferred partner (unlikely to occur in the wild) might cause chronic stress. Being chronically stressed when paired to a specific partner A but not when paired to partner B would be part of the “phenotypic incompatibility” phenomenon.
Our score of “pair harmony,” which was based on affiliative and sexual behaviors, as well as behavioral synchrony and the tendency to reunite, did not significantly correlate with pair fitness. A study on zebra finches in the wild reported that behavioral synchrony was associated with brood size [25], but further experimental work suggested that variation in synchrony might have been the consequence and not the cause of variation in reproductive success [72].
Evidence supporting the idea that pair coordination is important mainly comes from studies showing an increase in breeding success with pair bond duration ([27–29,73,74] but see [75]). We specifically designed our experiment to create variation in pair-bond duration (pairs stayed together for one or two breeding periods). However, this covariate did not have an effect on any of the fitness components (mostly showing non-significant trends opposite to expectation, Table 2) and was therefore removed from most final models. This suggests that behavioral compatibility (with synergistic effects on fitness) did not increase with pair bond duration.
The only traits that were affected by pair-bond duration were those related to extra-pair behavior (Table 2): females responded less positively to extra-pair courtships and received fewer extra-pair courtships with increasing pair-bond duration. In contrast, male courtship rate towards extra-pair females increased with pair-bond duration. In other words, it seems that females decreased and males increased their promiscuous behavior.
It has been suggested that individuals choose each other based on their respective personality, which would determine their behavioral compatibility [22]. Individuals that show similar behavioral types, or similar plasticity (and therefore predictability), could be better at negotiating or coordinating their actions, and could therefore have reduced conflicts over parental care and higher reproductive success ([26,76,77] but see [78]). So far, besides observational studies [76,79,80], only two experiments (both conducted on zebra finches) aimed at testing this hypothesis, and none of them found consistent evidence for pair combination effects on rearing success, based on any of the personality traits measured [26,78]. We did not measure any personality traits of individuals prior to the experiment, because we did not have clear a priori predictions about the advantages of being behaviorally similar. Instead, we scored the synchrony of activities during breeding, but this did not differ between treatment groups (see S4 Text).
Although an effect of lack of coordination between pair members cannot be excluded, our exploratory analyses suggest a reduced investment or commitment in individuals of “forced” pairs (lower female within-pair responsiveness, higher male extra-pair courtship rate, lower male nest attendance). Previous experimental work on zebra finches shows that the amount of male singing activity can affect egg quality [33]. More generally, courtship and other affiliative behaviors, which may occur more frequently in chosen pairs, may affect the level of reciprocal stimulation [32,81,82]. Earlier studies that favored the genetic compatibility hypothesis cannot rule out that the treatment (chosen versus non-chosen pairs) affected maternal investment (e.g., egg quality) with potential effects on offspring viability [35–37,43,46]. Artificial insemination would be needed to experimentally demonstrate that higher offspring viability arises from genetic compatibility and not from maternal (e.g., egg nutrients) or paternal effects (e.g., sperm allocation) following greater stimulation by a preferred partner (see e.g., [5]).
If forced pairs reduced their investment in breeding together, as our analyses suggest, the question remains whether this behavior is adaptive. Reduced investment in current reproduction could be adaptive, if it saves resources for future reproduction with a better (preferred) partner. However, this explanation seems unlikely for a species such as the zebra finch, because life-long monogamy largely precludes breeding with a different partner in the future [52]. Moreover, in a follow-up experiment consisting of a third breeding season where all individuals could freely choose their mate, individuals could not compensate for the lower fitness previously obtained with a non-chosen partner (see S6 Text). Therefore, the reduced investment in breeding by members of non-chosen pairs could be maladaptive, either because this never occurs under natural conditions (because individuals are never forced to mate or breed with a particular partner), or because some constraints limit the adaptive behavioral flexibility of the animals.
To conclude, chosen pairs had significantly higher fitness than forced pairs, apparently due to behavioral rather than genetic compatibility effects. The mechanisms behind such behavioral compatibility, in terms of willingness or ability to cooperate with certain individuals and in terms of coordination between partners need further study, in particular in the context of offspring provisioning. In humans, some studies suggest that individuals are more satisfied, more committed, and less likely to engage in domestic violence, when involved in a love-based rather than an arranged marriage ([83,84], but see [85]). The challenge there is also to find out whether stable and happy marriages result from motivation to cooperate (and to identify what stimulates such feelings, see [86–89]), or from congruence in terms of partners’ intrinsic behavioral types [90].
The study was approved by the Animal Care and Ethics Committee of the Max Planck Institute for Ornithology.
A scheme of the design with its timeline is depicted in Fig 4. All experimental birds hatched in the summer of 2011 in large semi-outdoor aviaries. The origin of the birds (population #4 in [91]), and rearing and housing conditions have been described in detail elsewhere [60]. This population has been derived from the wild only about ten generations ago [91]. Shortly after independence (age 45 d), individuals were put into eight mixed-sex peer-groups of ten males and ten females. When birds reached sexual maturity (100 d old) they were color-banded, and peer-groups were joined two by two (yielding four groups, each allowing 20 possible pairs to form). Sixty-six pairs were identified during ad libitum observations in the winter of 2011–2012. Mid-April 2012, half of the identified pairs were randomly assigned to the treatment group NC (in which all birds are assigned the partner of someone else: “non-chosen”), while the remaining pairs went to the treatment group C (in which birds are allowed to stay with their chosen partner). In order to induce pair formation in the randomly created pairs of the NC group, these pairs were put into individual cages for a period of two months and were allowed to lay one clutch. Pairs from the C group also went to such cages and were allowed to lay one clutch in order to standardize all experiences apart from the re-pairing. On the 21st of May, three pairs of each treatment group (chosen randomly but excluding the initial chosen partners of individuals of non-chosen pairs) were put into a breeding aviary (10 replicates, 60 pairs in total). Both members of each pair had been previously color-banded on both legs with one random color out of six (dark blue, light blue, black, yellow, orange, white), so that a pair would be unmistakably identifiable in its aviary. Forty-five pairs (26 C, 19 NC) did not divorce and were considered for the analyses. After one week of intensive focal pair observations, we introduced nest material, and checked nests daily until 21 August, when the experiment was stopped and newly laid eggs were replaced by dummy eggs, but pairs were still allowed to raise all offspring from eggs laid before that.
In October 2012, once all offspring had reached independence, we assigned treatment groups for the second breeding season. First, we randomly selected eight pairs from each treatment group (among the 26 C and 19 NC that were available) that were allowed to stay with their partner throughout the second season. In this way, we could better separate the effects of choice treatment from the potential effects of pair-bond duration. These 16 pairs and all other adults (remaining C, NC, and previously divorced birds) were put into one of two big flocks, to allow a second round of choosing a partner. Each group contained 20 widowed males (i.e., their former breeding partner was in the other group), 20 widowed females, and eight established pairs. As a result, each widowed female could choose a new partner among a set of 20 new males, which never included her previous breeding partner (but for half of the females from non-chosen pairs it did include again the initially chosen mate because this could not be avoided). In December 2012, after pair identification and random assignment to treatment group (without regard to previous treatment), pairs were put into cages for six months and allowed to lay two clutches. The longer period of force-pairing in cages resulted in a lower divorce rate compared to the previous season (only one pair of each treatment divorced). On 21 May 2013, pairs were put again into breeding aviaries and allowed to breed as in the previous year. Of 52 pairs identified in the winter groups, 42 (21 of each treatment group, across seven aviaries) contributed to the second breeding period (12 birds died accidentally because food dispensers were blocked for 2 d in early March 2013). The design itself and the accidental food shortage may have led to selection of the highest quality individuals. Although it did not induce bias (selection was independent of treatment group), it can result in an underestimation of the real fitness benefits of mate choice. Furthermore, one member of a pair died for unknown reasons (and its partner was removed) within the first week of each breeding period (1 C in 2012 and 1 NC in 2013), and these two pairs were excluded from the analyses.
Each aviary contained 7 nest boxes. Every morning, all nests were checked, the individual(s) attending the nest identified, and the fate of each egg and each offspring noted. Unhatched eggs were opened when neglected by the parents (for instance, after offspring had fledged) and embryos were collected for parentage analysis (using 11 microsatellite markers, following [92] and [93]). For the same purpose, small (~10 μl) blood samples were taken from 8–10 d old offspring, or tissue samples if they died earlier. Of 1,434 eggs laid by all birds including divorcees, 28% (n = 402) could not be assigned through parentage analysis, and were assigned to the social pair that attended the nest. These eggs included apparently infertile eggs (5.6%, n = 80), and eggs that were buried in the nest and did not develop (typically after a nest take-over by another pair) or disappeared (presumably they broke and were eaten by the birds) (21.6%, n = 310), as well as eight dead embryos and four dead hatchlings that yielded bad DNA samples. Relative fitness of an individual was calculated as the total number of genetic offspring produced in a given breeding period that reached independence (age 35 d) divided by the average number of offspring produced by all same-sex individuals of the same aviary that did not divorce.
Each aviary was equipped with a dome camera set to record different aviary positions during each day of the week. During 3 d, we filmed an artificial tree, on which 69% of all courtships took place (calculated from direct observations, described below). For one day, we recorded each of the two sets of nest boxes, and for 2 d, a set of perches on which individuals often allopreen. We analyzed the first hour of each day, when copulations are most frequent [53]. In all pairs considered for the analyses (those that did not divorce), we recorded 1,942 within-pair (WP) and 2,999 extra-pair (EP) courtships (in the latter, a divorced female or male may have been the extra-pair partner). For each courtship, we scored female responsiveness as follows: threat or aggression toward the male (−1), flying away (−0.5), mixed or ambiguous signs (0), courtship hopping and beak wiping (+0.5), and copulation solicitation (+1), and noted whether it led to a successful copulation.
We also conducted direct observations, following a protocol inspired by studies on cockatiels, Nymphicus hollandicus [21,23,94]. Observations were carried out both in the pre-breeding period (first week after release into aviaries before nesting material was added) and during the entire breeding period, to test whether pairs with greater behavioral compatibility before breeding (as in [23]), or during breeding activities, would have greater reproductive success. The observer stood behind a one-way glass window (built into each aviary door) and carried out focal-pair watches by monitoring a pair for 3 min. During these watches we observed 613 WP and 800 EP courtships. We noted their location and whether they led to a successful copulation. For a subset of 561 WP and 782 EP courtships, we also scored female responsiveness, as described above. During focal-pair watches, we also recorded whether within-pair allopreening or aggression occurred during the 3 min period (“yes” or “no”). Every 30 s, we recorded the distance between the partners and their activity. Distance was averaged for each 3 min watch. Activities were split into nine categories: feeding, cleaning, nesting or parental behavior (nest building or attendance, and feeding of fledglings), sleeping, sitting, involved in aggression, involved in courtship, flying, and “other.” We defined pair synchrony as the sum of the observations in which both partners engaged in the same activity (range 0–6). For each pair member, we also recorded all occurrences of an individual flying away from or back towards (<50 cm) its partner (e.g., female flying away: Faway, male flying back: Mback). From those counts, we calculated the tendency of the pair to reunite: (Σ Fback + Σ Mback) / (Σ Faway + Σ Maway), and a mate guarding index: (Σ Faway − Σ Fback) − (Σ Maway − Σ Mback). The latter is positive in case of male mate guarding, and negative for female mate guarding. The six pairs in an aviary were watched successively in a randomized predetermined order, and the time of observation of each aviary was randomized over the course of each day (i.e., from sunrise to sunset). In 2012, pairs were watched 9–13 times (median = 11) in the pre-breeding period, and 37–39 times (median = 38) during the breeding period. In 2013, 16–21 focal watches (median = 21) per pair were performed during pre-breeding, and 68–70 (median = 69) during breeding. For each pair, all measures were averaged for all focal watches separately for the pre-breeding and breeding period, because these periods were analyzed separately as planned a priori (see Results and S4 Text). Male courtship rates (WP and EP courtships per hour) and best linear unbiased predictors (BLUPs, i.e., random effect estimates) of female responsiveness (to WP and EP courtships) were also calculated (see S5 Text) for both periods and included in a principal component analysis (PCA). All observations were done blind to the treatment of the birds.
All statistical tests were conducted in R [95]. General and generalized mixed-effect models were performed with the “lmer” and “glmer” function of the lme4 package [96] and the PCAs with the “principal” function of the psych package [97]. All fixed effects were chosen a priori by considering (a) their biological relevance (e.g., hatching order when looking at offspring mortality), (b) their mathematical relevance (e.g., clutch size when looking at the presence of infertile eggs in a clutch), (c) the experimental design (e.g., year), and (d) consistency with previously published models (e.g., how to model the fertile period when looking at female responsiveness). P-values for general mixed effect models (lmer) were obtained from model comparison (with and without the explanatory variable) with the function anova in R; p-values for generalized mixed effect models (glmer) were taken from the model output (calculated from z-values).
|
10.1371/journal.pcbi.1004406 | A Stochastic Multiscale Model That Explains the Segregation of Axonal Microtubules and Neurofilaments in Neurological Diseases | The organization of the axonal cytoskeleton is a key determinant of the normal function of an axon, which is a long thin projection of a neuron. Under normal conditions two axonal cytoskeletal polymers, microtubules and neurofilaments, align longitudinally in axons and are interspersed in axonal cross-sections. However, in many neurotoxic and neurodegenerative disorders, microtubules and neurofilaments segregate apart from each other, with microtubules and membranous organelles clustered centrally and neurofilaments displaced to the periphery. This striking segregation precedes the abnormal and excessive neurofilament accumulation in these diseases, which in turn leads to focal axonal swellings. While neurofilament accumulation suggests an impairment of neurofilament transport along axons, the underlying mechanism of their segregation from microtubules remains poorly understood for over 30 years. To address this question, we developed a stochastic multiscale model for the cross-sectional distribution of microtubules and neurofilaments in axons. The model describes microtubules, neurofilaments and organelles as interacting particles in a 2D cross-section, and is built upon molecular processes that occur on a time scale of seconds or shorter. It incorporates the longitudinal transport of neurofilaments and organelles through this domain by allowing stochastic arrival and departure of these cargoes, and integrates the dynamic interactions of these cargoes with microtubules mediated by molecular motors. Simulations of the model demonstrate that organelles can pull nearby microtubules together, and in the absence of neurofilament transport, this mechanism gradually segregates microtubules from neurofilaments on a time scale of hours, similar to that observed in toxic neuropathies. This suggests that the microtubule-neurofilament segregation can be a consequence of the selective impairment of neurofilament transport. The model generates the experimentally testable prediction that the rate and extent of segregation will be dependent on the sizes of the moving organelles as well as the density of their traffic.
| The shape and function of axons is dependent on a dynamic system of microscopic intracellular protein polymers (microtubules, neurofilaments and microfilaments) that comprise the axonal cytoskeleton. Neurofilaments are cargoes of intracellular transport that move along microtubule tracks, and they accumulate abnormally in axons in many neurotoxic and neurodegenerative disorders. Intriguingly, it has been reported that neurofilaments and microtubules, which are normally interspersed in axonal cross-sections, often segregate apart from each other in these disorders, which is something that is never observed in healthy axons. Here we describe a stochastic multiscale computational model that explains the mechanism of this striking segregation and offers insights into the mechanism of neurofilament accumulation in disease.
| Axons are long slender projections of nerve cells that permit fast and specific electrical communication with other cells over long distances. The ability of nerve cells to extend and maintain these processes is critically dependent on the cytoskeleton, which is a dynamic scaffold of microscopic protein polymers found in the cytoplasm of all eukaryotic cells. The axonal cytoskeleton comprises microtubules, intermediate filaments called neurofilaments, and microfilaments. Microtubules and neurofilaments are both long polymers that align in parallel along the long axis of the axon, forming a continuous overlapping array that extends from the cell body to the axon tip [1, 2]. Microtubules are stiff hollow cylindrical structures about 25 nm in diameter that serve as tracks for the long-range bidirectional movement of intracellular membranous organelles and macromolecular cargo complexes. In axons, this movement is known as axonal transport [3]. The cargoes of axonal transport are conveyed by microtubule motor proteins: kinesins in the anterograde direction (towards the axon tip), and dyneins in the retrograde direction (towards the cell body) [4]. Neurofilaments, which are the intermediate filaments of nerve cells, are flexible rope-like polymers that measure about 10 nm in diameter [5]. These polymers function as space-filling structures that expand axonal cross-sectional area, thereby maximizing the rate of propagation of the nerve impulse [6, 7]. In large axons, neurofilaments are the single most abundant structure and occupy most of the axonal volume [8]. Mutant animals that lack neurofilaments develop smaller caliber axons and exhibit delayed conduction velocities [9–11].
In addition to their structural role in axons, neurofilaments are also cargoes of axonal transport, moving along microtubule tracks powered by kinesin and dynein motors [12–16]. The filaments move at rates similar to membranous organelles but the movements are less frequent, resulting in a “stop and go” motile behavior characterized by short bouts of movement interrupted by prolonged pauses on a time scale of seconds or shorter [17, 18]. The net result is an average rate of transport that is much slower than that for many other cargoes.
Neurofilaments have been observed to accumulate abnormally in axons in many neurodegenerative diseases including amyotrophic lateral sclerosis, hereditary spastic paraplegia, giant axonal neuropathy and Charcot-Marie-Tooth disease (also known as hereditary distal motor and sensory neuropathy) [5, 19–23], and also in many toxic neuropathies [24–28]. In extreme cases, these accumulations can lead to giant balloon-like axonal swellings [29–34]. These accumulations are thought to be caused by alterations in neurofilament transport, but the mechanism is not understood [35–40].
In healthy axons, microtubules and neurofilaments align along the length of an axon and are interspersed in axonal cross-sections [1, 41–43], with microtubules often forming small clusters in the vicinity of membranous organelles [8, 44, 45]. However, in many toxic and neurodegenerative disorders these polymers segregate, with microtubules and membranous organelles typically clustered in the center of the axon, and neurofilaments displaced to the periphery (Fig 1). This striking cytoskeletal reorganization, which is never observed in healthy axons, has been reported in neurodegenerative disorders as diverse as giant axonal neuropathy [46–48] and Charcot-Marie-Tooth disease [34, 49], as well as in neurotoxic neuropathies induced by exposure to agents as diverse as 2,5-hexanedione and 3,3’-iminodiproprionitrile (IDPN)[24, 50–57], aluminum [58], carbon disulfide [59, 60], estramustine phosphate [61], 1,2-diacetylbenzene [62] and 1,2,4-triethylbenzene [63], and in a transgenic mouse expressing a mutant neurofilament protein [64]. However, the mechanism of this segregation and its relationship to the neurofilament accumulation that also occurs in these different conditions is not known.
Microtubule-neurofilament segregation has been studied most extensively for IDPN and 2,5-hexanedione. IDPN is a compound closely related to the naturally occurring food poison 3-aminopropionitrile which causes the neurological disorder lathyrism [65–68], and 2,5-hexanedione is a metabolite of the industrial solvent hexane. The mechanism of toxicity is not known, but it is thought to involve chemical modification of neurofilaments, which presumably disrupts their normal interactions with microtubules in some way [25, 28, 69–74]. Systemic administration of IDPN or 2,5-hexanedione to rats by intraperitoneal injection or by addition to the drinking water causes selective impairment of neurofilament transport [75–79], focal accumulations of axonal neurofilaments leading to axon enlargement, and neurological defects similar to amyotrophic lateral sclerosis (ALS) in humans [80–83]. Sub-perineurial injection of IDPN or 2,5-hexanedione into peripheral nerves results in local microtubule-neurofilament segregation within just a few hours, preceding the accumulation of neurofilaments by hours or days [50–52, 56, 84]. This segregation does not appear to affect the axonal transport of membranous organelles, which continue to interact with and move along these tracks in spite of their clustering. Moreover, in the case of IDPN the segregation has been shown to be reversible [24, 50], as has the impairment of neurofilament transport [85]. In [24], a single injection of IDPN into rat sciatic nerves resulted in segregation in axons at the injection site within a few hours, but the segregation disappeared in about a day. In [50], a single injection of IDPN into the body cavity of rats resulted in segregation within the axons of the sciatic nerve after 4 days, and this disappeared after six weeks. Thus the microtubule-neurofilament segregation caused by IDPN and 2,5-hexanedione is fast, local and reversible.
Though the segregation of microtubules and neurofilaments in axons was first described more than 30 years ago, the underlying mechanisms are still poorly understood. Given that neurofilaments move along microtubule tracks and that microtubule-neurofilament segregation precedes neurofilament accumulation and axonal enlargement in rodent models, it is attractive to speculate that the segregation reflects an uncoupling of neurofilaments from their transport machinery [75]. However, the mechanism by which such an uncoupling at the molecular level might generate polymer segregation at the population level remains unclear.
To address these questions, we have developed a stochastic multiscale model for the cross-sectional organization of microtubules and neurofilaments in axons. The model describes microtubules, neurofilaments, and organelles as interacting particles that move in a 2D domain representing a cross-section of an axon, and incorporates axonal transports of neurofilaments and organelles, as well as volume exclusion and Brownian motion of all the particles. Neurofilaments and organelles dynamically bind to and unbind from nearby microtubules through molecular motors, and the motor cross-bridges are modeled as elastic springs. The longitudinal movement of neurofilaments and organelles along axons is modeled by stochastic addition and removal of these cargoes. The multiscale nature of the model lies in that it is built upon molecular processes that occur on a time scale of seconds or fractions of a second, and addresses segregation phenomena of two populations of polymers that occur on a time scale of hours to days.
Simulations of the model demonstrate that if we block neurofilament transport by preventing neurofilament from binding to microtubules, then organelles pull nearby microtubules together and gradually segregate them from neurofilaments on the same time course as observed in toxic neuropathies; while if we restore neurofilament transport, then microtubules and neurofilaments start to remix until their spatial distribution returns to normal. This suggests that the microtubule-neurofilament segregation observed in disease can be a consequence of the impairment of neurofilament transport. The model further predicts that (1) during the segregation process, microtubules first form small clusters, small clusters merge into bigger clusters, and eventually a single cluster forms close to the center of the domain; (2) in the absence of neurofilament transport, larger organelles are more effective in causing complete cytoskeletal segregation than small organelles with the same density. Further experimentation will be required to verify the insights and predictions of the model.
In our model, microtubules, neurofilaments and organelles are described as individual particles that move in a circular domain D with fixed radius R0, representing a cross-section of an axon. Microtubules and neurofilaments are rod-like polymers that align along the length of axons, thus they are treated as nondeformable disks in D (Fig 2A), with center positions denoted by xik = (xik, yik) and radii by r i k. Here k = M or N is the index for particle type: M for microtubule, N for neurofilament; and i with 0 ≤ i ≤ nk is the index for the k-type particle where nk is the total number of k-type particles. The radii of microtubules and neurofilaments are constant, with r i M = 12 . 5 nm and r i N = 5 nm. Organelles in axons have different sizes and shapes, and their cross-sectional geometry depends on their position relative to the cross-section (Fig 2B). In this model, we took organelles as spindle-shaped objects and, for simplicity, we did not consider possible shape changes (Fig 2C). Therefore the organelles exist as non-deformable disks in D, and as an organelle crosses D, its cross-sectional radius, riO, varies according to its position, z i O, relative to D,
r i O = b ( 1 - ( z i O ) 2 a 2 ) , - a ≤ z i O ≤ a . (1)
Here a is half of the organelle length, b is its maximum cross-sectional radius, z i O is the distance of its center to D, and the index “O” stands for organelle. By varying the parameters a and b, we can vary the overall dimension of the organelles (Fig 2C).
We examined three key molecular mechanisms that contribute to the cross-sectional distribution of microtubules and neurofilaments: slow axonal transport of neurofilaments, fast axonal transport of organelles, and volume exclusion of all the particles. In the following sections we describe in detail how these mechanisms were incorporated into our model. We denote the unit vector pointing from xik to xj l by eij kl, and the surface distance between the i-th particle of k-type and j-th particle of l-type by dij kl, given as,
dijkl=|xik − xjl|−rik−rjl. (2)
The parameters used in our model are physical, and thus they are all measurable. Most of them have already been measured [107–115], or there exist experiments that can be used to estimate them. Table 1 summarizes all the parameter values, and the detailed methods to obtain these parameters are given in the S1 Text. The units of these parameters reflect the time scales for the molecular processes integrated into the model, which are seconds or fractions of a second.
To solve the model numerically, we treated the binding and unbinding, arrival and departure of cargoes explicitly at discrete time steps, and integrated the model system Eq (8) using the explicit Euler’s method. Because σk, k = M, N, C are constant in time, the numerical integrator has strong order 1.0 [116]. We chose a time step h much smaller than all the time scales involved in Mechanisms 1–3. For the simulations of segregation and remixing over hours to a day, we used h = 1/50 sec if there was no organelle in D, and h = 1/1600 sec otherwise in order to deal with the stiffness of the equations introduced by the pushing of organelles when they move into D. The detailed simulation algorithm is included in the S1 Text. The computational tool is written in C++.
Morphometric studies suggest that neurofilaments are spaced randomly in axonal cross-sections when packed at low densities, but as the density increases they start to experience the volume-exclusionary repulsive forces of their neighbors and assume a less random distribution characterized by a more even neurofilament spacing [41, 43, 102]. In this section we demonstrate that the neurofilament distribution generated using our model agrees well with these experimental data.
Different methods have been used to characterize neurofilament distribution in axonal cross-sections. Kumar et al [102] used the radial distribution function (RDF) (also known as the pairwise correlation function). The RDF, denoted as g(r) describes how density varies as a function of distance from a reference particle. For particles that move randomly and completely independently, g(r) is a constant value of 1; while for crystalline structures g(r) forms peaks at precisely defined intervals. For neurofilaments in axons the shape of g(r) typically lies between these two extremes, increasing sharply from 0 and forming a peak around 30 − 50 nm [102]. Another method used often is to calculate the occupancy probability distribution (OPD), which is the distribution for the number of particles within an observation window of a specified shape and size [41, 43, 102]. For neurofilaments, the OPD can be approximated by Guassian [43, 102].
In previous experimental studies, the RDF and OPD of neurofilaments were calculated in selected regions of axonal cross-sections with almost no microtubules and organelles. To mimic such conditions, we performed simulations with exclusively neurofilaments, i.e., nM = nO = 0, and thus the only acting mechanisms are the pairwise repulsions and the Brownian motion of neurofilaments. We used a square domain with side length 1μm, and to minimize the effect of the boundary we used periodic boundary conditions. Under such conditions, the system Eq(8) reduces to
dxiN=∑j, j≠iRi,jNN/μNdt+σNdWiN, 1≤i≤nN.
We initially put neurofilaments on a hexagon lattice inside the domain, and then “randomized” the distribution by simulating the model for sufficient time to observe no further change in the OPD or RDF. To solve the model, we used the explicit Euler’s method with a time step h = 1/200 sec.
We first investigated how the neurofilament distribution depends on its density. We took εNN = 0.5 pN and used increasing neurofilament densities of 200 and 400 per μm2 (Fig 3, Rows A and B). For each case, the left panel is a plot of the coordinates of the neurofilaments after randomizing for 25 sec; the middle panel is a plot of the RDF which represent averages over 50 time frames between 25 sec and 30 sec; and the right panel is a plot of the averaged OPD and its Gaussian fit. The methods that we used to calculate the RDF and OPD are the same as in [102] and described in the supporting information (S1 Text). These plots show that as the neurofilament density becomes higher, the separation of the peaks of the RDF becomes smaller, and the average and variance of the OPD becomes larger as the neurofilament density becomes larger. General features of these plots are in tight agreement with experimental data presented in [41, 43, 102].
The magnitude of the repulsion between two neurofilaments depends on the charges of their sidearms. As the phosphorylation level of their sidearms becomes higher, their mutual repulsion becomes larger. We next investigated how the neurofilament distribution depends on the effect of sidearm phosphorylation by fixing the neurofilament density and varying εNN. We took the neurofilament density to be 400 per μm2, and εNN to be 0.25 pN and 0.5 pN. Fig 3B and 3C shows that as εNN becomes larger, the locations of neurofilaments become more regular, the peaks of the RDF are better defined, and the variance of the OPD becomes smaller.
To investigate the mechanism of microtubule-neurofilament segregation in axons, we compared our simulations to experimental data obtained for IDPN in laboratory animals. We focused on IDPN because there is published data on both the rate and reversibility of the segregation. When IDPN is administered transiently by local injection into peripheral nerves, segregation appears within 2–6 hours and then disappears within 24 hours [24, 50, 57].
Since neurofilament accumulation and axonal swelling occur on a much slower time course than the segregation, they can be ignored for the purposes of our current analysis. Therefore, for simplicity, we took D to be a disk with fixed radius R0 = 1 μm, and set the total number of neurofilaments nN to be constant. Specifically, if a neurofilament that was engaged with a microtubule left D, then it was replaced by a new neurofilament that entered D by association with a new randomly chosen microtubule. The total number of microtubules and neurofilaments in the domain were determined based on the experimentally determined densities of 18/μm2 and 115/μm2, respectively [50]. We thus calculated nM by the formula n M = floor ( 18 π R 0 2 ) = 56 and similarly we obtained nN = 356. Here the function floor(u) is the largest integer that is smaller than u. We considered organelles with b = 140 nm and a/b = 10 (Fig 2C) based on experimental data. All the parameter values are summarized in Table 1 and the estimation methods are given in S1 Text.
We started the simulations by including axonal transport of both neurofilaments and organelles, mimicking the conditions of normal axons. To distribute the neurofilaments and microtubules randomly without overlap, we first placed them on a hexagon lattice in D with no organelles, and then introduced volume exclusion and Brownian motion for enough time to randomize their positions. Starting from this initial condition, we then turned on the movement of both neurofilaments and organelles. Fig 4A is a snapshot of the simulated distribution of microtubules, neurofilaments, and organelles in a normal axon. The small grey dots are neurofilaments that are not engaged with microtubules, the small purple dots are neurofilaments that are engaged with microtubules, the large black dots are microtubules, and the large cyan circle is an organelle pushing into the cross-sectional domain. Note that a small fraction of the neurofilaments are bound to microtubules and moving along microtubules, that one microtubule can transport multiple cargoes (neurofilaments or organelles), and that one organelle can engage with multiple microtubules simultaneously.
We then blocked neurofilament transport selectively by resetting the binding rate of neurofilaments to microtubules, k o n N, to be 0 at t = 1 h. This disengaged neurofilaments from their microtubule tracks and thus blocked their movement so that none could enter or leave D. Meanwhile, the transport of organelles was not affected: they continued to grab microtubules stochastically, pulling them together. This “zippering” effect caused the microtubules to gradually cluster (Fig 4B). By 6 hours, almost all the microtubules had migrated to the center of D and formed a single island surrounded by neurofilaments (Fig 4C and 4D). The central microtubule cluster contained organelles but relatively few neurofilaments whereas the peripheral zone of neurofilaments contained relatively few microtubules or organelles. This segregation pattern is strikingly similar to that observed in experiments and in disease, and the rate of segregation is comparable to that observed experimentally for local injection of IDPN into peripheral nerves of laboratory animals [24, 50].
After observing segregation, we restored neurofilament transport by resetting k o n N to its original value at t = 13 h. This immediately allowed neurofilaments on the periphery of the microtubule core within a distance Rb of a microtubule to bind to that microtubule stochastically and then either unbind or exit D after a short while, as dictated by their stop-and-go transport behavior. As explained in the Methods, each neurofilament that exited D was replaced with a new neurofilament seeded adjacent to a randomly selected microtubule, but only if that microtubule was within a distance of Rb from another neurofilament already in that plane. Over time this resulted in a gradual infiltration of neurofilaments into the microtubule cluster in a centripetal manner (i.e. from the outside edges progressing inward), leading to a gradual dispersal of the microtubules (Fig 4E) and a return to their normal interspersed organization (Fig 4F). These results agree tightly with previous experimental findings.
To characterize the reversible segregation of microtubules and neurofilaments, we plotted the distribution and the mean of the pairwise distance between two microtubules (PDMT) as a function of time. Fig 5A and 5B are plots calculated from the simulation shown in Fig 4, which demonstrate a significant progressive decrease of the PDMT upon elimination of neurofilament transport (t = 1h) and subsequent increase upon restoration of neurofilament transport (t = 13 h). Fig 5C and 5D are plots for a normal axon for comparison. We see that under normal conditions, because microtubules and neurofilaments are interspersed, the distribution of PDMT is broad and the mean of it is about 0.8R0; and as microtubules and neurofilaments segregate from each other, the distribution becomes more compact and the mean of the PDMT decreases by almost 40%.
Another way to incorporate blockage of neurofilament transport is to increase the off-rate of neurofilaments k o f f N. We performed simulations with k o f f N 100 times larger, and obtained similar results as in Fig 4.
In the above section we have shown that in the absence of neurofilament transport, organelle transport leads to microtubule-neurofilament segregation. As we noted earlier, organelles can interact with multiple microtubules simultaneously and thus pull or zip nearby microtubules closer together. We next investigated the importance of this zippering mechanism for the segregation of microtubules and neurofilaments. To do this, we introduced a maximum number of microtubules that a single organelle can interact with simultaneously, denoted by mmax, and investigated how the PDMT depends on mmax in the absence of neurofilament transport.
Fig 6 plots the mean of PDMT as a function of time given different values of mmax. Each curve is averaged over five realizations with unpredictable seeds, and the error bars indicate the standard deviations over the realizations. If each organelle is only allowed to bind to one or two microtubules, i.e., mmax = 1 or 2, then microtubules and neurofilaments remain mixed over time and segregation does not occur at all (blue and green). Indeed, for mmax = 1, the mean of PDMT is slightly larger than that for a normal axon shown in Fig 5D. This is because organelles stir microtubules and neurofilaments and separate microtubules apart. As mmax increases, the PDMT curve decreases faster and the time needed to reach complete segregation decreases. Scatter plots of microtubules and neurofilaments (not shown here) show that for mmax = 4, partial but significant segregation was observed by 18 hours in all five realizations; for mmax = 6, complete segregation was observed by 18 hours in four out of five realizations; for mmax = 8 or 16, complete segregation was observed in all realizations within 10 hours. These results suggest that microtubule zippering by moving organelles is the causal mechanism for the segregation of microtubules and neurofilaments in the absence of neurofilament transport.
We next investigated the dependence of the segregation on the size and the flux rate of the organelles by simulating the model with different sizes of organelles, b = 140 nm or b = 70 nm, and different flux rates k i n O. Fig 7 plots the mean PDMT over time for four situations: b = 140 nm and k i n O = 0 . 105 / s (shown in blue), which is the same as in Fig 4; b = 140 nm and k i n O = 0 . 1575 / s (shown in green); b = 70 nm and k i n O = 0 . 105 / s (shown in red); and b = 70 nm and k i n O = 0 . 21 / s (shown in cyan). These results suggest that (1) for organelles of the same size, the more frequently they move through D, the faster the segregation occurs; (2) given the same flux rate across D, larger organelles are more capable of clustering microtubules and segregating them from neurofilaments than small organelles, and this is because on average larger organelles can interact with more microtubules simultaneously.
Interestingly, simulations of the model demonstrate that during the segregation process microtubules frequently form smaller clusters first, then these small clusters gradually merge with each other to finally form a single large cluster near the center of the domain. These intermediate states were more apparent in simulations with small organelles, presumably because the rate at which the smaller clusters merge is slower under this condition. Fig 8A–8C are snapshots of these intermediate states captured in a single realization (corresponds to the cyan curve in Fig 7). A similar pattern of isolated clusters of microtubules has also been reported by Zhu et al [73] (Fig 8D) (see Discussion).
We finally investigated the cross-sectional distribution of microtubules and neurofilaments when neurofilament transport is partially blocked. In the case of segregation induced by IDPN, this might be considered equivalent to lowering the IDPN concentration. To do this we reduced k o n N by different extents at t = 0h. Fig 9 plots the mean of PDMT over time for k o n N equals 0.5, 0.2 and 0 times of its original value. Each curve was obtained by averaging over 5 realizations with unpredictable seeds, and the error bars indicate the standard deviations about the mean. The data indicate that when k o n N is small enough, there is insufficient neurofilament transport to counteract the organelle-dependent microtubule clustering, and segregation is observed. However, as k o n N becomes larger, the rate of microtubule clustering becomes slower and the resulting clusters become less compact, reflecting less efficient segregation. Increasing k o f f N has a similar effect: as k o f f N becomes larger, the rate of microtubule clustering becomes faster and the clusters become more compact (not shown). Thus the rate and extent of microtubule-neurofilament segregation is dependent on the extent of inhibition of neurofilament transport.
We developed a novel stochastic multiscale model for the cross-sectional distribution of microtubules and neurofilaments in axons. The model describes microtubules, neurofilaments, and membranous organelles as interacting particles in an axonal cross-section. It incorporates detailed descriptions of key molecular processes that occur within seconds, including the axonal transport of neurofilaments and membranous organelles through this plane, as well as volume exclusion and Brownian motion of all the particles, and addresses the segregation phenomena that occur on a time scale of hours to days. The positions of the particles in the plane are governed by a system of stochastic differential equations.
Mathematical models of the axonal transport of neurofilaments and organelles have been developed previously to describe the longitudinal distribution of cargoes along axons [117–123]. However, those models were in 1D and did not consider the spatial arrangement and mechanical interactions of the cargoes and tracks in the radial dimension which are essential in understanding the segregation of microtubules and neurofilaments as well as the subsequent axonal swelling in neurological diseases. In our model, we describe in detail the dynamic interactions of neurofilaments, organelles, and nearby microtubules through molecular motors and volume exclusion in cross-section. Simulations of the model are in tight agreement with experimental data and generated a number of predictions that can be tested experimentally.
Simulations of the model demonstrate that if we block neurofilament transport selectively by preventing neurofilament binding to microtubules, while allowing organelle movement to continue, then the moving organelles tend to zipper nearby microtubules together so that they gradually segregate from the neurofilaments. The microtubule zippering action of the membranous organelles arises because we allow multiple motors to engage with a single organelle, which is consistent with experimental data and theoretical considerations [89, 90, 113, 124]. Restoration of neurofilament transport in the model allows the neurofilaments and microtubules to remix until their spatial distribution returns to normal. This suggests that neurofilament transport and organelle transport are competing processes in determining the cross-sectional distribution of microtubules: neurofilament transport can insert neurofilaments between adjacent microtubules, pushing those microtubules apart, while organelle transport can pull microtubules together when they move along multiple microtubules simultaneously, similar to a zipper. In normal axons, a dynamic balance between these two processes leads to the interspersed distribution of microtubules and neurofilaments, while in the absence of neurofilament transport, the microtubule zippering effect of organelle transport causes microtubules and neurofilaments to segregate. Thus our model predicts that the microtubule-neurofilament segregation that is observed in axons in neurotoxic and neurodegenerative diseases is a simple emergent property of the motile properties of membranous organelles that is triggered by selective impairment of neurofilament transport. An important and experimentally testable prediction of this study is that segregation is dependent on organelle movement. Further experimentation will be required to verify whether or not this prediction is correct.
An intriguing feature of microtubule-neurofilament segregation, which is consistent across all published reports, is that the microtubules generally cluster in the center of the axon, surrounded by a peripheral band of neurofilaments (see Introduction and Fig 1). It is interesting to note that this was usually the case in our simulations also. According to our model, the segregation generated by microtubule clustering is caused by an exclusion of neurofilaments from the microtubule domain due to their failure to interact. The central location of the microtubule bundle is essentially a boundary effect which arises because microtubules at the periphery of the axon can only be pulled towards microtubules that are located more centrally whereas microtubules in the center can be pulled towards microtubules on all sides. The net result is that microtubule zippering by moving organelles tends to pull these polymers towards the axon center, displacing the neurofilaments to the periphery. The organelles co-segregate with the microtubules because they must follow the available tracks.
An interesting observation in our simulations is that microtubule-neurofilament segregation tends to proceed initially via the formation of small microtubule clusters that subsequently merge together. This was more apparent in simulations with smaller organelles, which are less efficient at zippering microtubules together (see discussion below). Multiple small microtubule clusters have been reported in some studies on microtubule-neurofilament segregation induced by IDPN [54, 73] (see Fig 8D), but there is no published time course of segregation so it remains to be proven that these clusters are indeed intermediate states. Interestingly, microtubule zippering in our simulations also gives rise to the formation of small microtubule clusters in healthy axons. However, with ongoing neurofilament transport these clusters are transient and rarely merge to form larger ones. This is consistent with reports that small clusters of microtubules, often adjacent to one or more membranous organelles, are commonly observed in electron micrographs of axons [8, 44, 45, 91].
Our analysis gives us some insights into the factors that influence the rate of microtubule-neurofilament segregation. First, given the same number density, larger organelles are more effective at causing segregation, because they can interact with more microtubules simultaneously and they can pull together microtubules that are farther apart. Second, segregation occurs faster if the flux rate of the organelles is larger. Third, segregation occurs faster if the degree of neurofilament transport impairment is larger. These predictions are experimentally testable. It is also clear that there must be some dependence on the density of motors on the organelle surface, as well as the neurofilament:microtubule ratio. We are currently performing an extensive investigation on how the segregation phenomena depend on combinations of the model parameters using model simplification, nondimensionalization and mathematical analysis. These efforts will provide further insight of the biological problem and will be published elsewhere in the future.
The best experimental data on the kinetics of microtubule-neurofilament segregation is for animals treated with the neurotoxin IDPN. However, the rate of segregation in animals treated with IDPN depends on the mode of administration. When applied systemically to rats using a single intraperitoneal injection, segregation was first noted after 4 days, and after 4 such injections at 3 day intervals, the resulting segregation persisted for 6–16 weeks [50]. In contrast, when applied locally at high concentration by sub-perineurial injection into peripheral nerve, microtubule-neurofilament segregation was evident after 2 hours, with the microtubule clusters becoming increasingly compact over the next 4–10 hours [24, 52]. Nagele et al. [57] analyzed the pairwise distance between microtubules (PDMT) and observed full compaction by 8 hours after injection. Sixteen hours later, segregation was no longer seen in most axons, indicating an almost complete reversal [52]. In our simulations, we observed segregation within 4–12 hours of a complete cessation of neurofilament transport, and remixing within 2–8 hours after a complete resumption of neurofilament transport. This rate of segregation is comparable to the kinetics observed experimentally for injections of IDPN into nerves, and suggests that this delivery method results in a transient but acute inhibition of neurofilament transport. We predict that the slower time course of segregation that is observed when IDPN is administered systemically is due to the lower effective dose experienced by the axons in those studies. The rate of remixing was a bit shorter in our simulations than in the experimental reports, which may be because we assumed an instantaneous recovery of neurofilament transport rather than a gradual one, which is more likely.
It is important to note that the impairment of neurofilament transport that leads to microtubule-neurofilament segregation in toxic neuropathies and neurodegenerative diseases also leads eventually to focal neurofilament accumulations and axonal swellings (see Introduction). Since microtubules are the tracks along which neurofilaments move, and since microtubule-neurofilament segregation appears early and precedes neurofilament accumulation and axonal swelling, it has been hypothesized that the segregation reflects the uncoupling of neurofilaments from their transport machinery [24, 28]. Our modeling supports this hypothesis, but the molecular mechanism is unclear. Many of the neurotoxic agents that cause microtubule-neurofilament segregation and impair neurofilament transport (e.g. hexanedione, IDPN, carbon disulfide) are reactive molecules that could, or are known to, modify neurofilaments chemically [28]. It is thought that these compounds react with specific amino acid residues to form protein adducts which may then modify protein interactions, and that such chemical modifications target neurofilaments preferentially or that they somehow render these polymers more susceptible than other cargoes to transport impairments. This selectivity could arise, for example, due to the unique structure or unusual amino acid composition of neurofilament proteins. The mechanism of impairment could be by interfering with their interaction with molecular motors or with the interaction of these motors with the microtubule tracks. Future experimental studies will be required to resolve such questions.
The mechanism by which neurofilament accumulations arise is also of great interest given that this occurs in so many neurodegenerative diseases. Since local accumulations can only form if more neurofilaments move into a segment of axon than move out, the appearance of local swellings along axons implies some longitudinal instabilities in the transport of these cargoes. Therefore we propose that neurofilament segregation is an early event in neurofilament transport impairment but that longitudinal instabilities or non-uniformities in the transport impairment must arise to give rise to local accumulations and axonal swellings. We plan to address this in future studies. Due the complex spatial and temporal nature of this problem, which entails the interactions of multiple dynamic components, we believe that a full understanding can only be achieved by a combination of experimental and modeling approaches. Our present study is an important first step.
|
10.1371/journal.pntd.0006384 | Trypanosoma cruzi 80 kDa prolyl oligopeptidase (Tc80) as a novel immunogen for Chagas disease vaccine | Chagas disease, also known as American Trypanosomiasis, is a chronic parasitic disease caused by the flagellated protozoan Trypanosoma cruzi that affects about 8 million people around the world where more than 25 million are at risk of contracting the infection. Despite of being endemic on 21 Latin-American countries, Chagas disease has become a global concern due to migratory movements. Unfortunately, available drugs for the treatment have several limitations and they are generally administered during the chronic phase of the infection, when its efficacy is considered controversial. Thus, prophylactic and/or therapeutic vaccines are emerging as interesting control alternatives. In this work, we proposed Trypanosoma cruzi 80 kDa prolyl oligopeptidase (Tc80) as a new antigen for vaccine development against Chagas disease.
In a murine model, we analyzed the immune response triggered by different immunization protocols based on Tc80 and evaluated their ability to confer protection against a challenge with the parasite. Immunized mice developed Tc80-specific antibodies which were able to carry out different functions such as: enzymatic inhibition, neutralization of parasite infection and complement-mediated lysis of trypomastigotes. Furthermore, vaccinated mice elicited strong cell-mediated immunity. Spleen cells from immunized mice proliferated and secreted Th1 cytokines (IL-2, IFN-γ and TNF-α) upon re-stimulation with rTc80. Moreover, we found Tc80-specific polyfunctional CD4 T cells, and cytotoxic T lymphocyte activity against one Tc80 MHC-I peptide.
Immunization protocols conferred protection against a T. cruzi lethal challenge. Immunized groups showed a decreased parasitemia and higher survival rate compared with non-immunized control mice. Moreover, during the chronic phase of the infection, immunized mice presented: lower levels of myopathy-linked enzymes, parasite burden, electrocardiographic disorders and inflammatory cells.
Considering that an early control of parasite burden and tissue damage might contribute to avoid the progression towards symptomatic forms of chronic Chagas disease, the efficacy of Tc80-based vaccines make this molecule a promising immunogen for a mono or multicomponent vaccine against T. cruzi infection.
| Chagas disease is a neglected tropical disease caused by the parasite Trypanosoma cruzi that affects more than 8 million people around the world. Unfortunately, the diagnosis is generally performed too late, when anti-parasitic drugs are no longer effective. About 30–40% of infected individuals progress toward a symptomatic stage with cardiomyopathy as the main manifestation, leading annually to approximately 12,000 deaths. Therefore, new strategies for the control of the parasite must be explored. In this context, prophylactic or therapeutic vaccination has arisen as an interesting alternative to control the disease. In the present work, we present a parasite virulence factor (T. cruzi 80 kDa prolyl oligopeptidase, Tc80) as a new antigen for vaccine development against Chagas disease. Tc80-immunized mice developed a strong specific humoral and cellular immune response and most importantly were protected against Trypanosoma cruzi infection. Upon T. cruzi challenge, vaccinated mice presented a reduced parasite burden and less indicators of tissue damage. These results make Tc80 an interesting candidate to develop a mono or multicomponent vaccine against Chagas disease.
| Chagas disease, also known as American Trypanosomiasis, is a chronic parasitic disease caused by the flagellated protozoan Trypanosoma cruzi. It is estimated that 8 million people over the world are infected by this parasite and more than 25 million are at risk of contracting the infection. It is a vector-borne disease transmitted by insects from Reduviidae family (colloquially known as ‘kissing bugs’) but the parasite can also be transmitted by congenital route, blood transfusions, organs transplantation or by ingesting food contaminated with the infective stages of the parasite [1]. Despite of being endemic on 21 Latin-American countries, Chagas disease has become a global concern due to migratory movements [2].
T. cruzi infection presents two distinguishable stages. An usually asymptomatic acute phase, that lasts about 2 months and is characterized by a high level of parasites in blood. Unfortunately, the available parasiticide drugs for the treatment are only effective in this phase. On the other hand, the chronic phase begins when the parasitemia decreases. It can remain asymptomatic lifelong, however, about 30–40% of infected individuals develop heart or digestive manifestations. T. cruzi infection represents the main cause of infectious cardiomyopathy [3]. Different clinical trials were conducted in order to assess the etiological treatment on chronic phase and the outcomes were somehow discouraging. While the parasite burden was reduced, no improvement in heart manifestations was observed [4,5].
In this scenario, prophylactic vaccines have emerged as an interesting modality of disease control. In the current work, we propose the T. cruzi 80 kDa prolyl oligopeptidase (Tc80) as a promising vaccine candidate. Tc80 is an enzyme expressed in the extracellular blood trypomastigote and the replicative intracellular amastigote [6]. Tc80 is secreted by the parasite and it is able to degrade the major extracellular matrix components such as collagen and fibronectin, contributing to the invasion of the parasite to the mammalian cells [7]. Furthermore, it was demonstrated that selective inhibitors for Tc80 were able to block the parasite infection in vitro [6,8]. Considering these features, we evaluated Tc80 immunogenicity and its ability to confer protection against T. cruzi challenge in a murine model.
Three prophylactic vaccination strategies were carried out: 1) A sub-unit vaccine formulated with recombinant Tc80 (rTc80) plus oligodeoxynucleotides CpG (ODN-CpG) as adjuvant; 2) An attenuated bacterial vector delivering the Tc80 gene; and 3) A priming with Tc80 DNA delivered by a bacterial vector followed by a boosting with rTc80 + ODN-CpG (heterologous prime-boost strategy).
Here we demonstrate that Tc80-specific immune response elicited by the vaccination was able to confer protection during the acute and chronic phase of T. cruzi infection. These findings make Tc80 a promissory immunogen for a mono or multicomponent vaccine to prevent Chagas disease.
In this work, different groups of inbred mice from the BALB/c and C3H/HeN strains were used for immunization protocols. Mice breeding was carried out in the animal facilities of the “Instituto de Microbiología y Parasitología Médica” (IMPaM, UBA-CONICET). Experiments with animals were approved by the Review Board of Ethics of the School of Medicine (UBA, Argentina) and conducted in accordance with the guidelines established by the National Research Council [9]. Animal sample size was estimated by a power-based method [10].
T. cruzi bloodstream trypomastigotes (from RA, K98 or β-galactosidase- expressing Tulahuen strains) were obtained from blood of CF1 infected mice. Culture-derived epimastigotes were used to obtain the soluble fraction of a parasite lysate (F105) as described [11].
BHK-21 (ATCC CCL-10) cells were used to induce rTc80 expression on eukaryotic cells and Vero cells (ATCC CCL-81) were used for T. cruzi in vitro infection. Both cell lines were maintained in RPMI 1640 medium (Gibco) and routinely tested for Mycoplasma spp. infection with DAPI stain.
Tc80 gene was amplified by PCR from genomic DNA of T. cruzi strain RA as described [8]. For expression in bacteria, the forward primer (5’- GAGCCATATGCGCAGCGTTTACCC-3’) included NdeI restriction site (underlined) and the translation start codon (ATG, bold). Reverse primer (5’-GATACTCGAGTCAGTGATGGTGATGGTGATGCTCTTTCCACGAAGCATTGA-3’) included XhoI restriction site (underlined), translation stop codon (bold) and 6 his tag coding sequence (bold and underlined). Tc80 amplicon was cloned in the expression vector pET23a+ and the construct was used for transforming chemically competent Escherichia coli BL21 (DE3).
Recombinant Tc80 (rTc80) was induced for 18 h with 0.5 mM isopropyl-β-D-thiogalactopyranoside at 20°C and then purified in native conditions with a Ni2+-NTA-agarose resin. Purified rTc80 was dialyzed against phosphate-buffered saline (PBS) with 10% glycerol and LPS was removed with polimixin B-agarose (Sigma). Residual LPS was determined using HEK-Blue LPS Detection Kit (Invivogen) and it presented a concentration below the quantification limit: 0.3 ng/ml.
Additionally, for DNA vaccination, Tc80 gene was cloned in the eukaryotic plasmid pcDNA3.1+. In this case, the amplification was performed with the followings primers: forward primer: 5’-GAATAAGCTTGCCACCATGCGCAGCGTTTACCC-3’, including HindIII restriction site (underlined), the kozak sequence (double underlined) and the translation start codon (ATG, bold). The reverse primer: 5’-GCTCCTCGAGTCACTCTTTCCACGAAGCATTGA-3’ contained a XhoI restriction site (underlined) and the translation stop codon (bold). Tc80-pcDNA3.1 plasmid was used for transforming electrocompetent bacteria Salmonella enterica serovar Typhimurium aroA SL7207 which were used as DNA-delivery system.
PCRs were performed with ZymoTaq DNA polymerase (Zymo Research) with an initial denaturation 95°C for 5 min, 30 cycles of denaturation (95°C, 1 min)–annealing (59.3°C, 1 min)–extension (72°C, 2 min) and a final extension 72°C, 10 min.
Four groups of 7-week-old C3H/HeN (H2-Kk haplotype) mice (n = 5–6 animals per group) were immunized with four doses separated by ten days. Group I (GI): rTc80im, was immunized intramuscularly in the quadriceps muscles with 10 μg of rTc80 adjuvanted with 10 μg of ODN-CpG 1826 (Invivogen). GII: STc80, received 4 doses of attenuated Salmonella carrying a Tc80-coding eukaryotic plasmid via oral (1x109 CFU/mouse). GIII: Prime boost group (Pboost) was a combination of the two previous groups. Animals received two doses of Tc80 DNA delivered by the attenuated Salmonella (1x109 CFU/mouse) followed by two doses of 10 μg rTc80 + 10 μg CpG-ODN by intramuscular route. As control group (SaroA), mice were intramuscularly injected twice with PBS + 10 μg CpG-ODN and then two doses of attenuated Salmonella carrying an empty plasmid pcDNA3.1 by oral route [12].
Tc80-specific antibody titers (IgG, IgG1 and IgG2a) were determined as previously described [13]. Briefly, 96-well polyvinyl chloride plates (Nunc, Thermo Scientific) were coated with 0.2 μg rTc80/well for 1 h at 37°C. Then, non-specific binding sites were blocked with 3% BSA in PBS for 2 h at 37°C. Plates were washed three times with 0.05% Tween-20 in PBS and two-fold serial dilutions (in 1% BSA-PBS) of immunized or T. cruzi infected mice sera were added and incubated for overnight at 4°C. Next day, 3 washes were performed and anti-IgG-HRP (Sigma B6398) antibody (1/10000 dilution), or biotin-conjugated anti-IgG1 or anti-IgG2a antibody (Pharmingen Becton Dickinson) diluted 1/2000, were added as secondary antibodies and incubated for 1 h at 37°C. When biotinylated antibodies were used, an additional incubation with streptavidin-HRP (BD Biosciences) was carried out at 37°C for 30 min. Reactions were revealed with TMB (tetramethylbenzidine, BD OptEIA) and stopped with 4 N H2SO4. Absorbance at 450 nm was determined on an ELISA plate reader (Labsystems Multiscan EX). Antibody end-point titer was calculated as the reciprocal of the dilution with a DO450nm = 0.5
The ability of rTc80-specific antibodies to block parasite infection was assessed as previously described [14]. Briefly, T. cruzi blood trypomastigotes (Tulahuen strain) expressing E. coli β-galactosidase were pre-incubated with the sera of immunized mice (1/10 dilution) and subsequently used to infect Vero cells (5x104 parasites/5x103 cells). After an overnight incubation at 37°C, 5% CO2, cells were washed with PBS to remove non-infecting trypomastigotes and the culture was maintained for 4 more days. Cells were lysed with Nonidet P40 (1% v/v) and β-galactosidase activity was spetrophotometrically measured against the chromogenic substrate chlorophenol red β-D-galactopyranoside (CPRG). Inhibition percentage was calculated as follows: {1 - [(ABS595nm of cells infected with parasites in presence of immunized mice serum)/(ABS595nm of cells infected with parasites in presence of non-immunized mice serum)]} x 100.
Recombinant Tc80 (4 nM) was pre-incubated with 1/10 dilutions of sera from immunized mice. Then, residual prolyl oligopeptidase activity was determined in the presence of the dipeptide Z-Gly-Pro-AMC (Bachem), the fluorogenic substrate. The formation of the 7-amino-4-methyl coumarin (AMC) product was monitored by fluorometry (λexcitation = 355 nm and λemission = 460 nm). Reactions were carried out at 37°C in a final volume of 100 μl reaction buffer (25 mM Tris, 250 mM NaCl, 2.5 mM DTT, pH 7.5) in a 96-well black plate (Costar Corning). Fluorescence measurements were made on a PerkinElmer Victor3 fluorimeter.
The AMC formation was recorded over time in relative fluorescence units (RFU) and the slope (ΔRFU/Δtime) from the linear region of the curve was used for the calculation of initial reaction velocity (Vi).
Splenocytes from immunized mice (1x106/200 μL) were incubated in the presence or absence of rTc80 (10 μg/ml) with complete RPMI medium in 96-well U-bottom plates. After 4 h of incubation, brefeldin A was added at 10 μg/ml and 12 h later, surface and intracellular cytokine staining was carried out.
Splenocytes were incubated with an APC-labeled anti-mouse CD4 antibody (eBioscience Inc.) for 45 min at 4°C, fixed with 2% PFA for 15 min at RT and incubated for 45 min at 4°C in permeabilization buffer with PE-labeled anti-IFN-γ and PE-Cy7-labeled anti-TNF-α antibodies (eBioscience Inc).
To perform suitable data acquisition and analysis, autofluorescence, single-stained and FMO (fluorescence minus one) controls were included. Stained cells were passed through the BD FACSaria II flow cytometer.
Cytotoxic ability of splenocytes from immunized mice (effector cells) to induce death on cells loaded with a Tc80 peptide (target cells) was studied. For target cells preparation, spleen cells from non-immunized C3H/HeN mice were incubated in the presence of 10 μM Tc80 peptide SEAELRKKI (H-2Kk) or a non-related peptide to Tc80 sequence (AEEAFRLSV) [15] for 4 h at 4°C. These cells were washed with PBS and stained with 5 μM CFSE (CellTrace CFSE, Invitrogen) as described by manufacturer. Effector cells (1x106) obtained from immunized mice were co-incubated with 1x105 target cells for 4 h at 37°C and 5% CO2. Finally, the cells were washed with PBS and stained with 5 μg/ml propidium iodide (PI) to assess target cell death by flow cytometry (CFSE+PI + cells).
The percentage of specific death was calculated with the following formula [16]:
%Specificdeath=[(%CFSE+PI+%CFSE+)Tc80peptideloaded(%CFSE+PI+%CFSE+)non–loaded−1]×100
Spleen cells from non-immunized C3H/HeN mice were incubated for 4 h at 4°C with Tc80 peptide (10 μM) and stained with 8 μM CFSE (CFSEhigh population). Another group of splenocytes was loaded with a non-related peptide (AEEAFRLSV) and was stained with 3 μM CFSE (CFSElow population). Equal amounts of both populations were mixed and 2x107 cells were intravenously injected into immunized C3H/HeN mice in a volume of 100 μl. After 18 h, the splenectomy was performed, the spleens were disrupted and the splenocytes suspension was analyzed by flow cytometry (BD FACScanto).
The calculation of the specific lysis was done with the following formula[17]:
%Specificlysis=[1−(CFSEhighCFSElow)immunizedmice(CFSEhighCFSElow)controlmice(SaroA)]×100
Two weeks after last immunization, immunized female C3H/HeN mice were challenged intraperitoneally with 200 blood trypomastigotes of T. cruzi strain RA. Parasitemia was monitored by counting blood parasites every 2–3 days in a Neubauer chamber. For this purpose, a 1/5 dilution of blood in lysis buffer (0.75% NH4Cl, 0.2% Tris, pH 7.2) was made. Mice deaths were daily recorded.
Two weeks after the last dose of vaccination protocols, immunized male C3H/HeN mice were challenged intraperitoneally with 2.5x105 blood trypomastigotes of the sub-lethal T. cruzi strain K98. Parasitemias were recorded weekly during the acute phase. At 100 days post-infection (dpi), different parameters of disease outcome were assessed.
Target tissue damage was assessed at 100 dpi by determining serum activity of the creatine kinase (CK) and its cardiac isoform (CK-MB), glutamate oxalacetate transaminase (GOT), and lactate dehydrogenase (LDH). Enzyme activity determinations were performed by spectrophotometry at 340 nm with commercial kits following the instructions given by the manufacturer (Wiener Lab). Results were expressed as absorbance variation per minute (ΔABS/min).
Immunized and subsequently infected mice were anesthetized with ketamine-xylazine (100 mg ketamine and 16 mg xylazine/kg mouse) at 100 dpi and heart electrical activity was recorded with a Temis TM-300-V electrocardiograph. Corrected QT interval was calculated by the Bazett formula adapted for mouse [18].
At 100 dpi, skeletal (quadriceps) and heart muscles from immunized and subsequently infected mice were dissected and fixed with 4% formalin in PBS. Then the material was embedded in paraffin, sectioned and stained with hematoxylin and eosin. Finally, 10 microscopic fields (100x magnification) were analyzed in 10 sections of each sample. Inflammation was qualitatively evaluated according to the number and spreading of inflammatory foci. Samples were classified with the following score: (+) isolated foci; (++) multiple non-confluent foci; (+++) multiple confluent foci; and (++++) multiple diffuse foci [19,20].
Parasite burden in skeletal (quadriceps) and heart muscle at 100 dpi was determined as described by Cummings et al. [21]. Briefly, total DNA was extracted from about 50–100 mg of muscle using a phenol-chloroform-isoamyl alcohol mixture (25:24:1 v/v, QuickDNA, Kalium Technologies). DNA concentration was adjusted to 25 ng/μL and it was used as template for DNA amplification with T. cruzi-specific primers (Pf: 5’- GGCGGATCGTTTTCGAG -3’, Pr: 5’- AAGCGGATAGTTCAGGG -3’). Samples were also amplified with mouse TNF-α-specific primers as normalizer gene (Pf: 5’- TCCCTCTCATCAGTTCTATGGCCCA -3’, Pr: 5’- CAGCAAGCATCTATGCACTTAGACCCC -3’). PCR reaction was performed using HOT FIREPol EvaGreen qPCR Mix Plus (Solis Biodyne).
To make the standard for parasite burden quantification, about 500 mg of muscle of non-infected mouse were mixed with 1x108 T. cruzi epimastigotes. Total DNA was extracted and the concentration was also adjusted to 25 ng/μL. To construct standard curve, ten-fold serial dilution of the standard were made using non-infected mice muscle DNA (25 ng/μL) as diluent. Parasite burden was expressed as parasite equivalent/50 ng of total DNA.
Statistical analysis was performed using 1-way ANOVA along with the post-tests indicated in each trial. The homogeneity of variances was validated using the Levene test. Normality was verified using the Shapiro-Wilks test. The log-rank test (Mantel-Cox) was used to analyze the survival curves using the Prism 6.0 program (GraphPad, San Diego, CA). The statistical analyses were referred to the control group of each experiment, except when indicated. Values of p < 0.05 were considered significant.
Tc80 gene was cloned in pET23a+ plasmid and the construct was used to transform E. coli BL21 (DE). Bacterial rTc80 expression was induced O.N. and the protein was purified with a Ni2+-NTA agarose resin (Fig 1A). As shown in Fig 1B, the recombinant protein showed prolyl oligopeptidase activity. The identity of the protein was confirmed by immunoblot, as rTc80-specific polyclonal antibody recognized the native protein in a parasite lysate (Fig 1C).
To obtain a DNA vaccine for expression in mammals, the Tc80 gene was also cloned in the eukaryotic expression vector pcDNA3.1+. After transfection of BHK21 cells with this construction, the recombinant protein was evidenced in an indirect immunofluorescence assay (Fig 1D).
To analyze the humoral immunity triggered by the immunization protocols, Tc80-specific antibody titers in serum were determined by ELISA at day 15 after last immunization dose. We observed that mice immunized at least twice with the recombinant protein (rTc80im and Pboost group) elicited antibody titers considerably higher than control group (SaroA) (p<0.001). By contrast, STc80 group which was immunized only with Tc80 DNA carried by Salmonella, did not elicit significant specific antibody titer comparing to SaroA (Fig 2A). Besides, antibodies isotypes reflected a Th1-biased response since IgG2a levels were higher than IgG1 (Fig 2B). Interestingly, Pboost group showed an IgG2a/IgG1 ratio about 25-fold higher than rTc80im group, indicating that STc80 priming accentuated the bias towards a Th1 response.
Antibodies against rTc80 in STc80 group were only differentiated from control in an ELISA using a biotin-streptavidin signal amplification (Fig 2C). This observation is concordant with our previous experience in DNA immunization with attenuated Salmonella as carrier, that showed detectable, functional, but modest titers of specific antibodies [22,23].
As the antigen-specific B-cell response plays a role in protection [24,25], we assessed not only the magnitude but also the functionality of antibodies from immunized animals.
Neutralization of trypomastigotes cell infection was detected in sera from all vaccinated animals (Fig 2D). However, higher neutralization capacity was observed in rTc80im and Pboost groups that showed greater antibody titers. These groups were able to block nearly 50% of T. cruzi infection (Fig 2D). Moreover, anti-Tc80 antibodies from immunized animals were able to inhibit almost a 100% of the prolyl-oligopeptidase enzyme activity (Fig 2E), a fact that also correlated with the level of Tc80-specific antibodies.
Furthermore, antibodies from all the immunized mice were able to trigger trypomastigotes lysis in a complement-dependent cytotoxicity assay. Nearly 40% of parasite lysis was detected upon addition of a complement source independently of the vaccination protocol (Fig 2F).
To assess in vivo cell-mediated immunity, a delayed-type hypersensitivity (DTH) reaction was carried out on mice footpad. We found that all immunized groups developed a strong antigen-specific cellular response compared with SaroA control group. Among them, STc80 group showed the highest DTH reaction (Fig 3A). This result correlated with the proliferation of spleen cells that was evaluated ex vivo upon antigen re-stimulation where a similar pattern was obtained with STc80 group showing the highest response (Fig 3B).
Furthermore, splenocytes from all immunized mice were able to secrete IL-2 and IFN-γ upon antigen recall, being these differences significant in STc80 and Pboost compared with SaroA (Fig 3C-I and II).
By flow cytometric analysis we found that all immunized groups presented a higher percentage of IFN-γ or TNF-α producing CD4+ T cells compared with SaroA group, though it was only significantly higher only in STc80 (Fig 3D). Moreover, the frequency of antigen-specific polyfunctional CD4+ cells that simultaneously produced both IFN-γ and TNF-α was significantly increased in STc80 group (Fig 3E-I), representing more than 60% percent of the cytokine-producing cells within this group (Fig 3E-II). These polyfunctional populations were also associated with a high extent of cytokine production compared with single cytokine-producing cells (Fig 3F). In that way, mean fluorescent intensity (MFI) of IFN-γ channel indicates that double positive CD4+ lymphocytes produced 2 times more IFN-γ compared to that of single positive cells. Similarly, a 300-fold increase was detected in the TNF-α production of these cells.
In order to assess the immune response mediated by cytotoxic T lymphocytes (CTL), we searched for Tc80 H2-Kk epitopes with different algorithms: SYFPEITHI [26], RANKPEP [27–29] and IEDB [30]. From peptide prediction results (S1 Dataset) we selected and synthesized the one with the highest MHC-I binding score: SEAELRKKI.
CTL activity against the predicted H-2Kk nonapeptide was then evaluated ex vivo and in vivo. In the ex vivo approach, splenocytes from immunized mice induced death of target cells loaded with Tc80 peptide. We observed that the cytotoxic activity of STc80 group was significantly higher than the control group (Fig 4A). This result was further confirmed with an in vivo citotoxicity assay. Nonapeptide-pulsed and CFSE-labelled target cells were transferred to immunized mice and 16 h later, specific cell lysis was evaluated in the spleen by flow cytometry. Vaccinated mice presented higher in vivo CTL activity compared to control group since immunized animals showed a reduced frequency of peptide-pulsed cell population (CFSEhigh) with respect to the non-specific lysis subset (CFSElow) (Fig 4B-I). Similar to what we observed for the CD4+ population, STc80 group presented the highest level of CTL functionality between immunized animals (Fig 4B-II). Taken together, these results show that immunization with the Tc80 antigen was able to elicited specific CTL response that may contribute to the elimination of T. cruzi infected cells.
To evaluate if the immunization protocols were able to confer protection, vaccinated mice were challenged with the highly virulent T. cruzi RA strain. All immunized mice presented significantly reduced parasitemias during the acute phase of infection compared to the control (Fig 5A) and this was reflected in a 3-fold reduction of the areas under the curve (Fig 5B). We observed that on the early acute phase (9 dpi) rTc80im group achieved the highest control of parasitemia. By contrast, at 23 dpi when most of the control animals died, parasitemias were significantly lower in STc80 and Pboost compared with rTc80im (Fig 5C).
More important, all vaccinated mice had an increased in the survival rate compared with SaroA group. Specifically, those mice immunized with at least 2 doses of Salmonella sp. carrying the Tc80 gene presented a higher survival rate at 23 dpi: 80% and 67% for STc80 and Pboost, respectively (Fig 5D).
As T. cruzi is able to persist, even in vaccinated animals, the ability of the Tc80-based vaccine to prevent chronic phase-associated disorders, was evaluated after a challenge with the sub-lethal T. cruzi K98 strain (K98). In accordance with the observations after the lethal challenge, immunized mice presented lower parasitemias with respect to SaroA group during the acute phase of infection (Fig 6A). Since rTc80im group was not significantly protected against T. cruzi lethal challenge, this group was not subjected to a K98 infection.
At the chronic stage (100 dpi), immunized mice showed significantly lower serum levels of tissue damage-associated enzymes with respect to control group, more importantly these values were similar to those observed in non-infected mice (Fig 6B). Notably, immunization with Tc80 was able to avoid heart electrophysiological disorders such as corrected QT and PR intervals prolongation (Fig 6C). Thus, STc80-immunized and subsequently infected mice presented a QT interval significantly less prolonged than SaroA group and more interestingly they were similar to those registered in non-infected control mice (Fig 6C-I). Similar results were observed for the PR interval which tended to be less prolonged compared to control group and similar to non-infected mice (Fig 6C-II).
When qPCR was conducted on tissue samples, both immunized groups showed a 4-fold reduced parasite burden in skeletal muscle compared with SaroA group (Fig 6D). Additionally, by hematoxilin-eosine stain, we observed that immunized and infected mice also presented reduced extent of inflammatory infiltrates (Fig 6E). While SaroA group showed multiple diffuse inflammatory foci (++++), STc80 and Pboost groups showed non-confluent foci (++). On the contrary, in heart muscle, no significant difference was observed between the different groups. Interestingly, we found that parasite burden was relatively low (Fig 6D) which correlated with the low levels of isolated inflammatory foci observed in the histopathologic studies. Amastigotes nest in immunized and control groups were not detected neither in heart nor skeletal muscle.
Tc80 is a T. cruzi virulence factor involved in extracellular matrix degradation favoring tissue and cell invasion by the parasite. In addition, it has been shown that the inhibition of Tc80 enzymatic activity blocks non-phagocytic cell invasion [6,8]. In this scenario, we hypothesized that Tc80 would be an interesting target for the design of novel vaccines against Chagas disease.
To test Tc80 immunogenicity and its protective efficacy, we cloned and expressed it as a recombinant protein (rTc80). Tc80 cloned from T. cruzi RA strain presented 98–99% sequence identity with T. cruzi strains from different DTUs, including CL Brener, Tulahuen, SylvioX10/1 and Dm28c (GenBank Accession Numbers: XP_820337.1, AAQ04681.1, EKG04331, respectively). This high homology among different strains point out Tc80 as a good universal candidate for a vaccine against T. cruzi infection.
We were able to express soluble Tc80 with a high yield, and more importantly, the protein was enzymatically active against its specific substrate. This fact is relevant because it clearly indicates that Tc80 is adequately folded and therefore conformational epitopes would be available for antibody generation in a Tc80 based immunization.
Throughout this work we have used different vaccination protocols against T. cruzi infection in order to achieve protection. Thus, we explored a subunit vaccine constituted by rTc80, adjuvanted by ODN-CpG and administered intramuscularly. This adjuvant is a good inducer of humoral and cellular immune response with a Th1 profile [31] and have been tested with other T. cruzi antigens [13,16,32]. Considering the relevance of the cellular immune response to the control of T. cruzi infection, we also evaluated a vaccine based on Tc80 DNA [33]. To immunize with Tc80-coding DNA, we used an attenuated bacterial vector (Salmonella enterica serovar Typhimurium aroA SL7207) which was administered by oral route. This approach, among other advantages, increases transgene transfection efficiency compared with naked plasmid, has adjuvant effect due to the PAMPs from the attenuated bacteria and also presents a strong capacity to stimulate CTL and Th1 cell-mediated immune response, crucial for parasite control [25,34]. Our laboratory has previously demonstrated the efficacy of this DNA delivery system as a vehicle for different T. cruzi immunogens: Cruzipain [35,36], Tc52 [15] and Tc24 [36].
To combine the advantage of the above-mentioned formulations, a heterologous prime boost scheme was performed based on DNA priming and recombinant protein boosting. In this way, we aimed to join the strong cellular response triggered by DNA immunization and the powerful humoral response generated by recombinant protein immunization [37,38]. Different heterologous prime boost protocols (DNA + protein) have been implemented with different T. cruzi immunogens obtaining promising results; among them: Cz [39], TcG2 and TcG4 [40], Tc52 [41], etc.
In the present manuscript, we demonstrated that immunization protocols that included the recombinant protein (rTc80im and Pboost groups) elicited high titers of Tc80-specific IgG. In contrast, antibodies titers were low in the group immunized only with DNA carried by Salmonella (STc80).
As expected, in Pboost group, rTc80 boosting increased humoral response compared with STc80 DNA vaccine, and displayed an IgG2a predominating isotype suggesting a Th1-oriented immune response. Although rTc80im group presented an important IgG2a/IgG1 ratio (~6), the Th1 response bias was much higher in a Pboost vaccinated mice since the IgG2a/IgG1 ratio was about 150. The same phenomenon has been described for other DNA-protein prime boost schemes against different intracellular pathogens such as T. cruzi [41], Leishmania donovani [42], Mycobacterium tuberculosis [43] and Brucella spp. [44]. This Th1-bias generated by DNA-priming could be antigen-specific due to the immune response elicited by Tc80 DNA, as well as a non-specific adjuvant effect of Salmonella delivery system.
Tc80-specific antibodies elicited by immunization were able to block parasite infection to non-phagocytic cells, mediate blood trypomastigote lysis by complement activation, and inhibit the enzymatic activity of Tc80. The latter effect might be directly related to the blocking of Tc80 activity as the use of specific enzyme inhibitors blocked the infection of non-phagocytic cells [6,8]. Interestingly, despite experimental infection in mice certainly induces Tc80-specific antibodies, we could not detect antibodies able to inhibit rTc80 enzymatic activity, even during the chronic phase of the infection (S1 Fig). All these results suggest that the elicited antibodies by vaccination would be effective in inhibiting parasite invasion to cells. Surprisingly, sera from STc80 group were able to mediate these functions despite their low specific antibody levels. In this regard, further exploration of STc80 sera with an amplification ELISA allowed us to evidence the presence of anti-Tc80 antibodies, although there was not significant difference compared to the control.
We did not found correlation between anti-Tc80 antibodies level and sera ability to lysis parasites by complement activation since the lysis rate was approximately 40%, in all vaccinated mice despite the differences in antibody titers between groups. Classical complement activation suggests that Tc80 is located at least temporarily in the parasite membrane. Although a vesicular location near the flagellar pocket has been described in trypomastigotes [6], its location on the surface cannot be ruled out since many T. cruzi protein has been shown to be temporary located on the membrane [45–47].
With regard to the cell-mediated immune responses triggered by the immunization protocols, we observed that all immunized groups developed a Tc80-specific cellular response, which was reflected in a high delayed hypersensitivity reaction, antigen-specific proliferation and production of Th1 cytokines (IL-2, IFN-γ and TNF-α) that are crucial for parasite control [48,49]. Additionally, we found that the administration of 4 doses of STc80 maximize the robustness of cell-mediated immunity. Surprisingly, while prime boost strategy was aimed to improve the immune response, halving the number of doses of attenuated Salmonella was noticeably detrimental in to the stimulation of cellular response. Similarly, halving the number of doses of recombinant protein affected the specific humoral response.
Polyfunctional CD4+ T cells are those that simultaneously perform at least 2 functions and generally present a protective correlate against different chronic infections such as HIV infection [50], hepatitis C [51], tuberculosis [52] and leishmaniasis [53]. We found that mice from STc80 group presented a significantly higher proportion of multifunctional T lymphocytes producing simultaneously IFN-γ and TNF-α.
Some authors described the loss of polyfunctionality of CD4+ T cells during chronic infections as an exhaustion phenomenon, similar to the very well characterized effect for the CD8+ subset [54]. This situation has also been described for T. cruzi human infections [55]. In this regard, Albareda et al. [56], described that patients with early T. cruzi infection preserve CD4 polyfunctionality (IFN-γ+ and TNF-α+), while individuals with longstanding infections have higher frequency of monofunctional cells (i.e. IFN-γ+), indicating that a lack of polyfunctionality is associated to Chagas disease progression. Thus, the fact that immunization with STc80 specifically induced an increase in the number of CD4+ IFN-γ+ TNF-α+ T cells becomes more relevant. In addition, we also demonstrated that these polyfunctional cells have a significantly higher cytokine production ability for each cytokine than monofunctional cells.
We also showed in vitro and in vivo that Tc80-based immunization protocols triggered CTL activity. We identified the SEAELRKKI as a MHC-I–restricted cytotoxic T cell epitope in the Tc80 antigen. Splenocytes from immunized mice were able to induce death of target cells loaded with this Tc80 H-2Kk epitope. These results highlight the importance of immunization with DNA to promote antigenic presentation by MHC-I and the subsequent activation of CD8+ T cells, which play an important role in eliminating T. cruzi infected cells [57]. Although we focus on the cytotoxic activity, other CD8+ cells functions such as the cytokine production and degranulation capacity remain to be studied in depth. It is probably that the increased secretion of cytokines by splenocytes from STc80 and Pboost groups with respect to rTc80im group has its origin in CD8+ T cells. Further studies will be focused on a broader peptide screening to identify other potential epitopes capable of being presented in the context of MHC-I.
To analyze whether the triggered immune response was able to confer protection in the acute and chronic phase of infection, different challenges were performed. In a lethal-acute infection model, vaccinated mice were challenged with T. cruzi RA strain. This strain is highly virulent, pantropic [58] and belongs to discrete typing unit (DTU) Tc VI [59]. We observed that all immunized groups presented lower parasitemias compared with the control group. However, only those mice that received at least 2 doses of Salmonella carrying Tc80 gene (STc80 and Pboost groups) showed a significantly higher survival with respect to the controls. The protection achieved in these groups highlights the crucial role of the stimulated cell-mediated immunity in controlling T. cruzi infection.
With the aim of resembling a human infection where most patients survive the acute phase [60], we implemented a non-lethal mice model of infection. This chronic model was carried out by infecting male C3H mice with T. cruzi K98 clone which is non-lethal, myotropic [58], and belongs to DTU TcI [59]. The rationale behind the choice of this strain was its low lethality allowing infected mice to survive the acute period and reach the chronic phase. Moreover, myotropism of K98 strain increases the probability for detecting muscle disorders. Additionally, in order to favor the development of chagasic pathology, we used male mice which are more susceptible to T. cruzi infection than female ones [61,62]. Once the challenge was performed, different parameters of tissue damage were evaluated at 100 dpi.
In experimental and natural T. cruzi infection, the damage of muscle fibers caused by parasitic persistence leads to increased blood levels of intracellular enzymes such as LDH, GOT and CK. Importantly, we observed that all immunized groups had lower serum levels of CK, CK-MB, GOT and LDH than the control group. In accordance with this, we showed that immunized mice presented less electrocardiographic alterations than those registered in the control group (SaroA). Prolongation of the QTc interval is one of the most frequently alteration described for T. cruzi murine infection model [63] and it is generally related to a slow ventricular repolarization [64]. We found that mice immunized with STc80 had a significantly less prolonged QTc interval (p<0.05) than control mice and most importantly, these values were similar to the QTc interval observed in non-infected mice.
Furthermore, the parasite burden in target tissues at 100 dpi was evaluated by qPCR. We observed lower parasitic loads in cardiac muscle compared with skeletal muscle. These results were consistent with the histopathological analysis where we found a few isolated inflammatory foci in heart muscle and a larger infiltrate in skeletal muscle of control infected mice. These differences between both target tissues may correspond to the K98 strain differential myotropism as well as differential immune response mounted on each tissue. Interestingly, we found that in the skeletal muscle of the immunized mice the parasite burden was significantly reduced compared with SaroA groups. These results are in agreement with the correlation between parasitism and inflammatory infiltrates as described in the literature [65–67]. Neither parasite burden nor inflammation was detected in heart tissue. In contrast, skeletal muscle, presented high parasitism and inflammation. Therefore, the ECG disorders and the elevated serum levels of CK-MB observed in SaroA group were not associated with an ongoing heart inflammatory process at 100 dpi. These physiological disorders may be the result of previous irreversible alterations despite the low inflammatory extent [68].
In conclusion, we have demonstrated that Tc80-based vaccines are able to confer protection against T. cruzi infection, and importantly this immunoprotection was extensive to different strains belonging to different DTUs. Ours results emphasize the importance of stimulating cell-mediated immunity by activation of polyfunctional CD4+ T cells and cytotoxic T lymphocytes in order to control T. cruzi infection. Even though our results highlight the importance of humoral responses as an effective weapon in the control of early stages of parasite infection, if cell-mediated immune response is not strongly stimulated by vaccination, mice are unable to survive against a lethal challenge, as we observed in rTc80im group.
Similar to other vaccine immunogens which were previously described for T. cruzi, formulations with Tc80 as DNA vaccine or in a prime boost strategy partially control the infection. Although the ideal vaccine will be the one that provides sterilizing immunity, avoiding progression to symptomatic forms of the chronic Chagas disease is indeed a very promising goal. Considering T. cruzi complexity, a multicomponent and/or chimeric vaccine approach has been proposed by us [36,69] and others [40,70]. In this context and based on the results presented here, Tc80 represents a novel target to be considered in the development of an effective vaccine against T. cruzi infection.
|
10.1371/journal.pcbi.1002946 | Restricted N-glycan Conformational Space in the PDB and Its Implication in Glycan Structure Modeling | Understanding glycan structure and dynamics is central to understanding protein-carbohydrate recognition and its role in protein-protein interactions. Given the difficulties in obtaining the glycan's crystal structure in glycoconjugates due to its flexibility and heterogeneity, computational modeling could play an important role in providing glycosylated protein structure models. To address if glycan structures available in the PDB can be used as templates or fragments for glycan modeling, we present a survey of the N-glycan structures of 35 different sequences in the PDB. Our statistical analysis shows that the N-glycan structures found on homologous glycoproteins are significantly conserved compared to the random background, suggesting that N-glycan chains can be confidently modeled with template glycan structures whose parent glycoproteins share sequence similarity. On the other hand, N-glycan structures found on non-homologous glycoproteins do not show significant global structural similarity. Nonetheless, the internal substructures of these N-glycans, particularly, the substructures that are closer to the protein, show significantly similar structures, suggesting that such substructures can be used as fragments in glycan modeling. Increased interactions with protein might be responsible for the restricted conformational space of N-glycan chains. Our results suggest that structure prediction/modeling of N-glycans of glycoconjugates using structure database could be effective and different modeling approaches would be needed depending on the availability of template structures.
| An N-glycan is a carbohydrate chain covalently linked to the side chain of asparagine. Due to the flexibility of carbohydrate chains, it is believed that the N-glycan chains would not have a well-defined structure. However, our survey of N-glycan structures in the PDB shows that the N-glycan structures found on the surfaces of homologous glycoproteins are significantly conserved. This suggests that the interaction between the carbohydrate and the protein structure around the glycan chain plays an important role in determining the N-glycan structure. While the global N-glycan structures found on the surfaces of non-homologous glycoproteins are not conserved, the conformations of the carbohydrate residues that are closer to the protein appear to be more conserved. Our analysis highlights the applicability of template-based approaches used in protein structure prediction to structure prediction and modeling of N-glycans of glycoproteins.
| Glycosylation represents one of the most important post-translational modifications [1], [2] and is ubiquitous in all domains of life. The glycosylation machinery is largely conserved in eukaryotes, and more than 50% of all eukaryotic proteins are expected to be glycosylated [3], [4]. An oligosaccharide moiety in a glycoprotein, referred to as a glycan, comes in a diversity of sequences and structures and plays critical roles in a vast array of biological processes [1]. The N-glycosylation pathway is the most common pathway in which an oligosaccharide is covalently attached to the side chain of asparagine [2]. In general, such an oligosaccharide appendage masks the protein surface, protecting the glycoprotein from degradation and nonspecific protein-protein interactions (reviewed in [5]–[7]). N-glycosylation also alters the biophysical properties in the vicinity of the glycosylation site and affects the folding rates and the thermal stability of the protein [8], [9]. Some N-linked oligosaccharides (N-glycans) are directly involved in specific molecular recognition events; e.g., lectins and antibodies can recognize specific N-glycans on viral envelope glycoproteins such as HIV gp120 [10]–[13].
The impact of glycosylation on the structure of the parent protein and vice versa has been of great interest in structural glycobiology [8], [14]–[17]. At this time, however, an understanding of which glycans are important components in protein function and how to modify these glycans to optimize the protein properties of interest remain an enigma. Therefore, knowledge of the structure and dynamics of N-glycans is central to understanding protein-carbohydrate recognition and its role in protein-protein interactions. An oligosaccharide chain is flexible in solution and has an ensemble of diverse conformations rather than a single well-defined structure [18]–[20]. The inherent flexibility of oligosaccharides often hinders crystallographic structure determination, and there are only a few crystal structures of oligosaccharides longer than 2–3 residues in the Cambridge Structure Database [21]. In contrast, there are many more crystal structures of glycoconjugates in the Protein Data Bank (PDB) [22], suggesting that the presence of the protein may reduce the conformational freedom of oligosaccharides or even favor a certain conformation over others [23]. For example, the N-glycan conformations in the crystal structures of the Fc domain [24]–[30] exhibit remarkable similarity (Figure S1 in Supporting Information), suggesting that the protein's structure around the glycan has an influence on the glycan's conformation.
The number of PDB entries containing carbohydrates has been steadily increasing, but obtaining the complete N-glycan structure remains challenging [23]. Mass spectrometric mapping of N-glycosylation sites is becoming common [4], providing information about glycosylation sites as well as the relative abundance of different glycoforms. In this context, computational modeling of N-glycan structures is an appealing approach to provide glycosylated protein structure models. In particular, a computational approach that can combine known glycoprotein structures and glycosylation information (i.e., glycosylation site, primary glycan sequence, and linkage information) would be very useful in a variety of applications in glycoscience. For successful template-based glycan structure modeling, it is essential to understand the conformational variability of an oligosaccharide chain when it is glycosylated. In addition, the influence of the protein residues around the glycosylation site can provide valuable insight into the design of new computational approaches that are optimized for glycoconjugates. Several structural database surveys have investigated the general features of N-glycosylation in terms of oligosaccharide and protein structures [14], [23], [31]–[35]. In these earlier studies, however, the oligosaccharide conformations were analyzed in terms of individual glycosidic torsion angles, making it difficult to recognize the actual structural variability of glycans en bloc. To the best of our knowledge, the conformational variability of N-glycans using the three-dimensional (3D) structures in the PDB has not been studied.
In this work, using the PDB crystal structures that contain N-glycans, we examined the conformational variability in various N-glycans. Using Glycan Reader [36], an automatic sugar recognition algorithm that we developed, all N-linked glycoprotein structures were obtained from the PDB and sorted by their N-glycan sequence. PDB entries with more than 3 Å resolution were excluded and N-glycan sequences with less than 20 PDB entries were also excluded, resulting in 35 N-glycan sequences (see the full list in Table S1 in Supporting Information). Using random background conformations of each N-glycan sequence, the statistical significance of glycan structural similarity was estimated. The N-glycan structures in the PDB show statistically significant similarity when the local structure around the protein is conserved. When the local protein structures are different, overall N-glycan structures are not conserved, but their internal substructures appear to be strongly conserved due to the proximity to the protein. Our results highlight the applicability of template-based approaches used in protein structure prediction to the structure prediction and modeling of N-glycans of glycoproteins. Although the N-glycan sequences examined in this work mostly represent oligomannose-type glycans due to the limited numbers of crystal structures of complex- and hybrid-type glycans, the conclusions might be applicable to other glycoconjugates' glycan sequences.
Because glycan sequences have branches and different linkages between monomers, alignment of glycan structures with different sequences is challenging and, to the best of our knowledge, there is no such an alignment tool for glycans. Therefore, in this study, pairwise structure similarity is measured using the root-mean-squared deviation (RMSD) among glycan structures having the identical glycan sequence. Assuming that homologous protein structures share similar surface features, the structural similarity of glycans found on homologous proteins would provide insight into the influence of the protein structure on the N-glycan structure. Therefore, N-glycan structure pairs with the identical glycan sequence are designated as “homologous” or “non-homologous” depending on the sequence similarity of their parent proteins (with a sequence similarity of 30% as a cutoff). Unless stated explicitly, highly homologous pairs (sequence similarity ≥90%) as well as redundant structure pairs were excluded from the analysis. There are a total of 289 homologous and 33,333 non-homologous glycan structure pairs in the final dataset (see Figure 1 and Methods for details). In this section, N-glycan structural similarity is examined and its statistical significance is estimated using random background conformations of each N-glycan sequence (see Methods for details). The structural similarity of the N-glycans is then discussed in terms of the protein's structure as well as the structural rigidity of the oligosaccharide regions that are closer to the glycosylation site on the protein.
The structural similarities of the N-glycans are measured by calculating the glycan RMSD after alignment of the oligosaccharide structures using the carbohydrate ring heavy atoms. N-glycan structural similarity including their orientations with respect to the protein is discussed separately below. Figure 2 shows the RMSD distributions of the N-glycan structure pairs in the PDB and random conformation pool. Note that the RMSD is only measured between glycan structures having an identical sequence. The average RMSD of all PDB structural pairs are 1.4±0.8 Å. The homologous and the non-homologous N-glycan structure pairs have RMSD values of 0.9±0.8 Å and 1.4±0.8 Å, respectively. Both the homologous and non-homologous N-glycans showed smaller RMSD values compared to those in the random glycan structure pool whose RMSD is 2.4±0.8 Å (Figure 2A).
Measuring the structural similarity using RMSD is straightforward, but it is not an objective measure when comparing structures of different lengths and sequences due to its length dependence. When the average RMSD values of the N-glycans are plotted against N-glycan length, i.e., the number of carbohydrate monomers (Figure 2B), a length dependence is observed for the random background and non-homologous glycan pairs, but homologous glycan pairs do not show such a length dependence. The smaller RMSD values of the homologous N-glycan structure pairs compared to the RMSD values of the non-homologous pairs indicate that the homologous N-glycan structures are more conserved than the non-homologous N-glycan structures.
Because our dataset contains different lengths of N-glycan sequences with different branching patterns (Table S1), we converted the RMSD values to their statistical significance (p-values) using the random background glycan structures (see Methods for details). By deriving the statistical significance using the random background having the identical N-glycan sequence, the length dependence is effectively removed. The generalized extreme value distribution (Eq. 1 in Methods) was used to estimate the statistical significance [37], and 35 sets of parameters were determined by fitting the generalized extreme value distribution to the original RMSD distribution of the random conformational pool of each glycan sequence (see the determined parameters in Table S2 and the fitting results in Figure S2). The calculated p-values (Eq. 2 in Methods) represent the probability of having randomly chosen two N-glycan structures whose RMSD is smaller than the random background. A list of p-values and the corresponding RMSD values averaged over different sequences are given in Table 1.
Figures 3A and 3B show the cumulative fraction of homologous and non-homologous glycans structure pairs as a function of their p-value. It is clear that about 67% of the homologous N-glycan structure pairs have a statistically significant level (p<0.05) of structural similarity, whereas about 36% of non-homologous N-glycan structure pairs have a statistically significant level of structural similarity. A correlation is also found between the sequence similarity of the glycoprotein and the structural similarity of the N-glycan (Figure S3). Specifically, about 81% and 91% of N-glycan structure pairs have statistically significant structure similarity when the parent proteins have sequence similarity greater than 50% and 60%, respectively. A similar analysis has been carried out independently using the global distance test (GDT) score [38] instead of RMSD, and the conclusion remains the same (Figure S4). Assuming that the proteins with similar sequences have similar surface features around the glycosylation site, such a high level of N-glycan structure similarity strongly indicates that the protein structure around the N-linked oligosaccharide plays an important role in determining the N-glycan structures.
Apparently, not all homologous glycans have significant structural similarity. Figure 4A shows an example of two homologous proteins, the Fc domain of IgG (PDB:2WAH) in green and the Fc domain of IgE (PDB:3H9Y) in orange, which share a sequence similarity of about 50% and have significantly different glycan structures (RMSD of 2.9 Å and p-value of 0.6). The structures of these two homologous proteins around the glycosylation site are similar and well aligned. Notably, the structural difference of the N-glycans arises mainly from the terminal residues at the 1–6 branches (or 1–6 arm). The PDB:2WAH IgG-Fc domain is glycosylated with a different glycoform than typical IgG-Fc glycans whose 1–6 arm carbohydrates are tightly packed with the proteins [24]–[30]. This may explain such a different glycan conformation in PDB:2WAH.
There are some non-homologous N-glycan structure pairs that have a statistically significant level of structural similarity. Visual inspection of several examples of non-homologous glycoproteins having similar N-glycan conformations shows no apparent similar protein surface features around the N-glycans. Figure 4B shows an example of two non-homologous glycoproteins, beta-galactosidase (PDB:3OG2) in green and the extracellular domain of the nicotinic acetylcholine receptor 1 subunit (PDB:2QC1) in orange, having a significant level of structural similarity of the N-glycan (RMSD of 0.9 Å and p-value of 0.009). Nonetheless, the structure alignment of these two N-glycans results in a poor alignment of the parent proteins.
The relative orientation of an oligosaccharide chain with respect to the parent protein can be affected by the Asn side chain conformation and the protein conformation in the vicinity of the glycosylation site. To examine N-glycan structural variability with respect to the parent protein, the heavy atoms of the glycosylated Asn residue were used for alignment of each pair, and then the Euclidean distance of the glycan portion was measured without further alignment. Figures 3C and 3D show the cumulative fraction of structure similarity of the homologous and non-homologous glycans aligned with glycosylated Asn residues. Clearly, structural similarity is greatly reduced when the Asn residues are used for the alignment. Given the fact that glycosylation has a bias towards turns and extended regions [32], it is not surprising that even homologous N-glycans show reduced structural similarity when the Asn residues are used for the alignment.
The observations so far indicate that a comparative modeling approach for N-glycan structures would successfully predict the N-glycan structure itself, especially when the homologous N-glycan templates are present in the PDB, but finding the global orientation of the glycan with respect to the protein would remain challenging. Such difficulties can be significantly alleviated when a partial glycan structure is available. In fact, there are large numbers of partial N-glycan structures available in the PDB, probably due to the removal of glycans prior to structural studies, due to crystallization conditions, or due to missing electron density resulting from flexible glycan structures. For example, as of December 2011, there were 2,517 PDB entries and 10,769 N-linked glycan chains in the RCSB database; 84% (9,027 chains) had partial glycan structures with less than two carbohydrate units and 15% (1,394 chains) of such partial structures showed their parent protein sequence similarity less than 50%. Assuming that one can find such partial glycan structures, Figures 3E and 3F show the cumulative structural similarity of the N-glycans when the first two carbohydrate units in the glycan chains are aligned. Both the structural similarities of the homologous and non-homologous N-glycan structures (especially the former) significantly increased, suggesting that the conformations of glycosylated Asn residues and the first few carbohydrates of the N-glycan are important in determining the N-glycan orientations.
What makes homologous N-glycan structures conserved compared to non-homologous N-glycans or random background? Possibly, the protein structures around the glycan may provide a steric barrier, thus restricting the conformational freedom of N-glycans nearby. In addition, specific protein-carbohydrate interactions may play an important role in favoring a certain conformation of the oligosaccharides. If local protein structure around the N-glycan is directly correlated with the N-glycan structure similarity, such information provides valuable criteria in N-glycan structure modeling.
Figure 5 shows the correlation between the local protein structure around the glycan chain and the N-glycan structure similarity. As expected, most homologous glycoproteins have similar local protein structures around the glycan chain. However, some homologous N-glycan structure pairs adopt significantly different conformations while their local protein structures are similar (p-RMSD>0.05 and p-local<0.01). Visual inspection of such structures shows that the structural differences are mainly due to the terminal residues, especially ones in the 1–6 branches, similar to the case in Figure 4A. The increased flexibility of the 1–6 linkage is not surprising because the 1–6 glycosidic linkage contains three rotatable torsional angles (compared to two for other glycosidic linkages), and the flexibility of the 1–6 linkage has been well documented by other experimental, computational, and structural database surveys [34], [39]–[42].
To examine the flexibility of different regions of N-glycan structures, we have used the GDT chart [38]. Figure 6 shows two example N-glycan sequences and the corresponding GDT charts, where each bar represents an alignment of an N-glycan pair and the bar is colored according to how well a certain region of the sequence can be aligned each other. Clearly, the increased flexibility of terminal residues is apparent and, in particular, the residues in the 1–6 branches are even more flexible.
Non-homologous N-glycan structures in the PDB do not show a correlation with local protein structure around the glycan. There could be several factors responsible for this observation, and the accuracy of local protein structure alignment might be one important factor. To compare the similarity of local protein structure, TM-align [43] was used because the algorithm is general and performed well compared to other local structure algorithms available in our internal testing [44]. However, the TM-align algorithm was developed for comparison of global protein structure, and it is possible that the algorithm is insensitive to the structural similarities of the small number of residues around the glycan chain. Thus, further in-depth investigations with robust local structure algorithms are warranted.
The lack of correlation between the local protein structure and non-homologous glycan structures suggests that the gapless threading approach to N-glycan modeling would be inapplicable when no homologous templates are present. It was reported that the majority of glycosylation sites are found to be in convex or flat regions of the protein surface [32]. When the N-linked oligosaccharides are situated in such regions, the terminal residues of a long oligosaccharide may not be able to interact with the protein surface residues, and experience a smaller influence of the local protein environment. Thus, local protein structure around glycan chains might have a stronger impact on the first few residues of the glycan chain rather than on the global structure.
Internal substructure conservation can be visualized with the two examples in Figure 6, showing that the flexibility of the carbohydrate residues increases as the residues move away from the protein. In addition, a large increase in flexibility is observed after the 1–6 linkage, which is known to be flexible. If the N-glycan substructure is more conserved, a threading or fragment assembly approach could be useful to model the N-glycan structures. To quantify the conservation of internal substructures, we compared the structural similarity of the N-glycans as a function of glycan chain length from the protein. Figure 7A shows the average RMSD of N-glycan internal substructures containing only the residues within the given residue distance from the Asn residue of the parent protein. The conservation of the internal substructure is apparent up to 3 or 4 residues away from the Asn residue. Note that N-glycan sequences can have branches, and thus, there could be more residues in a substructure within a certain residue distance. For example, in the two examples in Figure 6, there are in fact 5 sugar residues at a residue distance of 4 from Asn.
To avoid the inherent length dependence of RMSD (i.e., a smaller substructure has a smaller RMSD), RMSD values for the substructures are converted to p-values using the random background. Figure 7B and 7C show the cumulative fraction of the substructure similarity for homologous and non-homologous N-glycans, respectively. About 80% and 60% of the substructure up to a residue distance of 3 (black curve) show significant structural similarity for homologous and non-homologous N-glycans, respectively. The substructures are less conserved when residues up to a distance of 4 are included in the substructure (blue curve). As discussed above, due to its flexibility, the 1–6 linkage might contribute to the diversity of the N-glycan substructures more than other glycosidic linkages. Clearly, when structural similarity of substructures up to a residue distance of 4 is compared without residues linked by the 1–6 linkage (red curve), significant structural conservation is observed even for non-homologous N-glycans. This observation implies that the glycan residues closer to the protein surface have more restricted conformational space and conserved structures.
Elucidation of the factors influencing the conformational variability in N-glycans is essential to understand the dynamics of N-glycans and provides valuable insight into modeling and computational studies of the N-linked oligosaccharides. In this work, we have shown that the conformations of homologous N-glycans are restricted compared to the random background. About 67% of the homologous N-glycan pairs and 37% of the non-homologous N-glycan pairs show statistically significant level of structural similarity. Although excluded from the main analysis, more than 90% of highly homologous N-glycan structure pairs (protein sequence similarity ≥90%) show very significant structural similarity (Figure S5).
Why do homologous N-glycans have conserved conformations compared to the free oligosaccharides? First, protein-carbohydrate interactions may restrict the conformational freedom of the N-glycans. In addition, the shape of the local protein structure may also act as a non-specific steric barrier and restrict the N-glycans to adopt certain conformations. Lastly, crystallographic bias in the dataset could also play a role in conformational similarity of homologous N-glycan structures. Our dataset is composed of crystal structures of well-resolved N-glycan structures; hence, flexible N-glycan structures may not be included in our dataset.
Despite the biological importance of N-glycans, understanding the structure and dynamics of N-glycans is currently lacking due to the difficulties in crystallization of glycoproteins and other experimental techniques. The high level of structural similarity among the N-glycan structures found on the surface of homologous proteins strongly indicates that the comparative modeling and threading approach used in protein structure prediction [45]–[47] might perform well in glycan structure modeling if appropriate templates are present. Despite the structural similarity of N-glycans on the homologous glycoproteins, the absolute orientation of N-glycan with respect to the glycosylated Asn residue may differ because the glycosylation site are often found on the loop regions of the protein.
N-glycan modeling without good template structures appears to be challenging because of less conserved N-glycan structures found for non-homologous proteins. However, a higher level of internal substructure similarity exists even for non-homologous N-glycan pairs up to a residue distance of 4 without the 1–6 linkage. In fact, these carbohydrate structures that lie close to the protein are key determinants of the overall N-glycan orientation. Thus, a fragment assembly approach might perform well even without homologous N-glycans template structures because of this internal substructure conservation.
Extracting structural information of glycans from the PDB is nontrivial due to a lack of standardized nomenclature and the way the data is presented in the PDB. To recognize the PDB entries that contain carbohydrate molecules, we used Glycan Reader for automatic sugar identification [36]. Briefly, in Glycan Reader, the topologies of the molecules in the HETATM section of a PDB file are first generated using the CONECT section of the PDB file, and the candidate carbohydrate molecules (a six-membered ring for a pyranose and a five-membered ring for a furanose that are composed of carbon atoms and only one oxygen atom) are identified. For each carbohydrate-like molecule, the chemical groups attached to each position of the ring and their orientations are compared with a pre-defined table to identify the correct chemical name for the carbohydrates. Glycan chains are constructed by examining the glycosidic linkages between the carbohydrate molecules that have chemical bonds between them. As of December 2011, there were 2,517 PDB entries and 10,769 N-linked glycan chains in the RCSB database. The glycan fragment structure database, including the substructures of the original N-glycan chains, was generated, which resulted in a total of 48,568 N-glycan fragment chains.
From the N-glycan fragment database, we have collected glycan structures composed of more than 3 carbohydrate units. A glycan structure was excluded when its resolution was higher than 3 Å or when it had less than 20 structures in total, resulting in the 35 N-glycan fragment sequences listed in Table S1. An N-glycan structure pair is called “non-homologous” when the sequence similarity of the parent proteins is less than 30%. Because a glycoprotein can have multiple glycosylation sites in a single domain, if the distance between the backbone Cα atoms of the two glycosylated Asn residues is more than 10 Å after alignment of the glycoprotein chains using TM-align [43], the N-linked glycan structure pairs are considered “non-homologous” glycans. The rest of the N-glycan structure pairs are called “homologous” glycans. Figure 1 summarizes the protocol for building the N-glycan structure dataset.
To quantify the conformational variability of the PDB N-glycan structures, it is essential to know the upper bound of the conformational variability in a given oligosaccharide. In protein structural biology, the upper bound of conformational variability is estimated by using the non-homologous protein structure pool and sequence-independent structure alignment methods [48]–[50]. However, because such sequence-independent structure alignment methods are not available for oligosaccharides, it is difficult to estimate the upper bound of the conformational variability in oligosaccharides only using the crystal structures in the PDB.
Instead of using the crystal structures directly, a conformational pool that contains diverse conformations of a specific N-glycan sequence was generated as follows. For each of the 35 N-glycan sequences, a total of 1,000,000 glycan conformations were generated in an iterative fashion. The initial structures were generated by using the IC BUILD command in the CHARMM biomolecular simulation program [51] according to the glycan sequence. For each iteration, a glycosidic linkage was randomly selected and a new torsion angle value was also randomly chosen based on the accessible glycosidic torsion angles of the corresponding glycosidic linkage type. If the newly generated conformation had bad contacts with neighboring atoms, the conformation was rejected and the protocol was repeated until no bad contacts were found. If a conformation had no bad contacts, the conformation was recorded and the protocol repeated until 1,000,000 conformations were generated. A bad contact was defined by the CHARMM van der Waals energy higher than 10 kcal/mol. Accessible glycosidic torsion angle values were used rather than the values observed in the PDB because the number of observations is limited for certain types of glycosidic linkages. For example, Figure S6 in Supplementary Material shows the resulting glycosidic torsion angle distributions of the N-glycan core sequence using the accessible glycosidic torsion angle values, and Figure S7 shows the torsion angle values observed in the PDB, respectively.
To construct an accessible glycosidic torsion angle map, a total of 13 adiabatic (φ, φ, ω) potential maps were constructed for each distinct glycosidic linkage type found in the 35 N-glycan sequences. For each glycosidic linkage type, a disaccharide connected by the corresponding glycosidic linkage type was generated by CHARMM [51], and the CHARMM carbohydrate force field [52]–[54] was used to evaluate the energy. The adiabatic map was generated by evaluating the energy over a grid of glycosidic torsion angles with a grid spacing of 5°, resulting in a total of 373,248 grid points for (1→6) linkages (φ, φ, ω) and 5,184 grid points for the rest of the glycosidic linkages (φ, φ). At each grid point, the conformations were minimized with the dielectric-screened Coulombic electrostatic and Lennard-Jones potential energy while the glycosidic torsion angles were restrained and a harmonic restraint potential was applied to the carbohydrate rings to prevent the distortion of the ring geometry. The generated adiabatic potential energy map was converted to a torsion angle probability map using the Boltzmann distribution. Finally, the resulting distribution was compared with the glycosidic torsion angles observed in the PDB using the Glycan Fragment DB [55], available at www.glycanstructure.org. The glycosidic torsion angle probability maps and the observations in the PDB matched well in general. However, the torsion angle probability map was clearly more restricted (data not shown). To remedy the restricted conformational space, glycosidic torsion angle pairs having probability above 0.0001 were considered “accessible”; this covers on average about 65% of the observed glycosidic torsion angles in the PDB.
The N-glycan structural similarity was measured by calculating pairwise RMSD in the following three different ways: First, the heavy atoms in the carbohydrate ring (C1, C2, C3, C4, C5, and O5) were used for the alignment of two N-glycan structures and in the RMSD calculation. Second, to examine the variability of N-glycan orientations with respect to the protein, the heavy atoms of glycosylated Asn residues were used to define the alignment, and then the Euclidean distance of the N-glycan structures was calculated using the carbohydrate ring heavy atoms. Third, many crystal structures only have a few residues at the glycosylation site due to difficulties associated with glycan crystal structure determination, and these partial glycan structures can be used to model the rest of a full glycan structure. To examine the efficacy of such an approach in obtaining a better N-glycan orientation with respect to the protein, the carbohydrate ring heavy atoms of the first two residues were used for the alignment of N-glycan structures, and then the Euclidean distance of the N-glycan structures was calculated using the ring heavy atoms excluding the first two residues.
The statistical significance of the structural similarity between two glycan structures was estimated by comparing the structural similarity of 124,750 random glycan structure pairs for each N-glycan sequence. The structural similarity of random glycan structure pairs was calculated by the identical procedure described above. Using the statistical model, p-values of the corresponding structural similarity measure can be calculated. This allows us to compare structural similarity across different glycan sequences and lengths. Each RMSD distribution for each glycan sequence was modeled by the generalized extreme value distribution,(1)where . The variable represents the RMSD of a structure pair; , , and are the location, scale, and shape parameters, respectively. These parameters were obtained through the maximum likelihood estimates by the EVD package in R (http://www.r-project.org). 35 sets of determined parameters are given in Table S2 and the fitting results are shown in Figure S2. The resulting correlational coefficients () are generally good except for a few sequences. The correlation coefficients improved when more “liberal” protocols were used (e.g., when the glycosidic linkage was not restricted and larger energy cutoff values were used to define bad contacts; data not shown). However, such protocols may produce unrealistic random glycan conformers and are not used in this work. The p-value of a glycan structure pair from the PDB having RMSD values smaller than the random glycan conformation background was calculated by(2)
The local protein structures are defined for protein residues having any heavy atoms within 6 Å from any glycan heavy atoms. The local protein structures were derived from the PDB structure files in our dataset, and the TM-align algorithm [43] was used to compare the structural similarity of a given local protein structure pair. Any local protein structures having less than 5 residues were excluded. The TM-scores calculated by TM-align were normalized by the length of the smaller structure. To estimate statistical significance, we have derived a random local protein structure pool using the N-linked glycoproteins in the PDB. Briefly, from the PDB, a non-redundant N-linked glycoprotein structure list having at least one carbohydrate residue and protein sequence similarity less than or equal to 30% were generated. A random local protein structure pool was derived from the protein residues having any heavy atoms within 6 Å from any of the carbohydrate heavy atoms. The TM-align algorithm was used to calculate the distribution of TM-scores from the random local protein structure pairs. The calculated TM-score distribution was fit using the generalized extreme distribution (Eq. (1)), and the p-values of having TM-scores larger than the random background were estimated using Eq. (2).
Although there are several local structure alignment tools available [56]–[58], it was difficult to directly utilize them in this study because many of them are highly customized to specific domains, such as a protein-protein interface or protein-ligand interface. Thus, we used TM-align [43] to compare local structure similarity. Although TM-align is not designed to compare local structure similarity, it performed well in our internal testing and correctly found most homologous glycoproteins having similar local protein Cα structures; also see ref for protein local structure comparisons [44].
The residue distance is defined as the minimum number of glycosidic linkages between carbohydrate monomers, including the glycosidic linkage to Asn. For each of 35 N-glycan sequences, three types of internal substructures were generated; a) residue distance up to 3, b) residue distance up to 4, and c) residue distance up to 4, excluding residues linked by the 1–6 linkage. Then, the RMSD of substructures were measured after alignment using the carbohydrate ring atoms in the substructure. To estimate the statistical significance of the internal substructures, the random glycan internal structure pool was generated for each of three different types of substructures. The resulting random background distributions were fit using Eq. (1) and p-values were calculated using Eq. (2).
|
10.1371/journal.pbio.1000566 | A dp53-Dependent Mechanism Involved in Coordinating Tissue Growth in Drosophila | Coordination of growth between and within organs contributes to the generation of well-proportioned organs and functionally integrated adults. The mechanisms that help to coordinate the growth between different organs start to be unravelled. However, whether an organ is able to respond in a coordinated manner to local variations in growth caused by developmental or environmental stress and the nature of the underlying molecular mechanisms that contribute to generating well-proportioned adult organs under these circumstances remain largely unknown. By reducing the growth rates of defined territories in the developing wing primordium of Drosophila, we present evidence that the tissue responds as a whole and the adjacent cell populations decrease their growth and proliferation rates. This non-autonomous response occurs independently of where growth is affected, and it is functional all throughout development and contributes to generate well-proportioned adult structures. Strikingly, we underscore a central role of Drosophila p53 (dp53) and the apoptotic machinery in these processes. While activation of dp53 in the growth-depleted territory mediates the non-autonomous regulation of growth and proliferation rates, effector caspases have a unique role, downstream of dp53, in reducing proliferation rates in adjacent cell populations. These new findings indicate the existence of a stress response mechanism involved in the coordination of tissue growth between adjacent cell populations and that tissue size and cell cycle proliferation can be uncoupled and are independently and non-autonomously regulated by dp53.
| The coordination of growth within and between organs contributes to the generation of functionally integrated structures and well-proportioned animals and plants. Though these issues have fascinated biologists for centuries, the responsible molecular mechanisms remain largely uncharacterized. In this work, we have used the Drosophila wing primordium to show that adjacent cell populations grow and proliferate in a coordinated manner. By reducing growth rates in specific territories within the developing wing, we showed that the tissue responds as a whole and that in adjacent cell populations the growth and cell cycle rates are concomitantly reduced, thus maintaining tissue proportions and normal wing shape. Interestingly, we show that the Drosophila tumour suppressor protein dp53 and apoptotic machinery play a key role in coordinating this tissue-wide response. Both growth and proliferation rates are regulated in a coordinated and non-autonomous manner by the activity of dp53, whilst the apoptotic pathway has a specific and non-autonomous role in regulating cell proliferation rates. Our studies describe a novel mechanism for regulating tissue growth in developing organs that may ultimately be relevant for other processes involving coordination of growth, such as tissue renewal, regeneration, and cancer.
| In multicellular organisms, coordination of growth between and within organs contributes to the generation of well-proportioned organs and functionally integrated adults. Although the mechanisms that help to coordinate the growth between different organs start to be unraveled [1], the underlying molecular mechanisms that contribute to generating well-proportioned adult organs remain largely unknown.
In Drosophila, primordia of the adult head, thorax, and terminalia are established in the embryo as imaginal discs, which grow and proliferate within the feeding larva, fuse during metamorphosis, and give rise to the adult animal [2]. Even though the size, shape, and pattern of each adult structure are genetically determined in an autonomous manner by each imaginal disc, several humoral mechanisms contribute to coordinating the growth between imaginal discs to generate a well-proportioned adult fly. Insulin-like growth factors (IGF) and Target of Rapamycin (TOR) kinase signaling couples the nutritional status of the animal with the growth of each imaginal disc, and steroid and neuropeptide hormones coordinate the termination of growth with developmental timing (reviewed in [3]). Damage or growth retardation of imaginal tissue induces, through the activity of steroid and neuropeptide hormones, a larval developmental delay to ensure that termination of growth is coordinated among growing tissues and all organs attain a characteristic final size [4],[5]. Similarly, the activity of IGFs modulates the growth of each imaginal disc to give rise to well-proportioned adult flies, with variable sizes depending on nutritional status (reviewed in [6]).
Here we used the wing imaginal disc to first analyze whether adjacent cell populations within an organ grow in a coordinated manner to give rise to a well-proportioned structure and afterwards to determine the molecular mediators involved in this coordination. The wing disc is a mono-layered epithelium that grows about a thousand-fold in mass and cell number during larval development. After metamorphosis, it gives rise to the adult wing, a flat structure with a species-specific shape, size, and pattern. By reducing the growth rates of defined territories within the developing wing primordium and analyzing the non-autonomous response throughout development of the adjacent cell populations, we demonstrate that adjacent cell populations respond as a whole by decreasing their growth and proliferation rates. This non-autonomous response occurs independently of where growth is affected, and it is functional throughout development. We underscore a central and non-autonomous role of Drosophila p53 (dp53) and the apoptotic machinery in these processes. While the decrease in growth and proliferation rates are regulated in a coordinated and non-autonomous manner by the activity of dp53, effector caspases have a non-autonomous role in reducing proliferation rates. These new findings indicate the existence of a stress response mechanism involved in buffering local variations in growth in order to maintain the relative contribution of each cell population to the final organ and that tissue size and cell cycle proliferation can be uncoupled and are regulated by two different mechanisms downstream of dp53.
In order to address whether adjacent cell populations grow and acquire a final size in a coordinated manner, we quantified the size and shape of adult wings when tissue growth was reduced in defined territories of the developing wing. Growth, defined as accumulation of cell mass, can be modulated by changing the biosynthetic capacity of cells (reviewed in [7]). In Drosophila larvae, starvation of dietary nutrients leads to smaller adult flies. Starvation modulates tissue size by reducing insulin/TOR signaling, which leads to the inhibition of ribosome synthesis, a decrease in nucleolar size, and a reduction in protein synthesis capacity [7],[8]. Thus, a set of genes with well-known growth inhibitory functions in ribosome function, protein biosynthesis, or insulin signaling were expressed with the Gal4/UAS system in specific territories of the developing wing primordium (Figure 1A). A cold-sensitive version of the Ricin toxin A chain (Ricincs), a protein toxin that belongs to the ribosome inactivating proteins that binds and reduces the translational activity of 28S rRNA and thus protein synthesis [9], was used to impair growth, as its activity can be easily modulated in time by changes in temperature [10]. 4E-BP is an important repressor of translation levels and can be inactivated by the protein kinase TOR [11]. 4E-BP binds eIF4E and impairs the recruitment of the 40S ribosomal subunit to the cap structure present at the 5′-end of all eukaryotic cellular mRNAs. To reduce elF4E activity and impair growth, we used a constitutively active form of 4E-BP (4EBPAA, [12]) that cannot be phosphorylated by TOR. Finally, the activity of the insulin/TOR pathway was reduced by expression of the tumor suppressor gene PTEN, a conserved negative regulator of the pathway [13],[14]. A collection of Gal4 drivers expressed in distinct domains of the developing wing primordium was used to induce the expression of Ricincs, PTEN, and 4EBPAA (Figures 1B, C and S1). Larvae containing the Gal4 driver and the UAS-Ricincs transgene were initially grown at 18°C (restrictive temperature) until early second instar and then switched to 29°C (permissive temperature) until eclosion of the resulting adult flies. Expression of PTEN and 4EBPAA was achieved by growing the larvae at 25°C. Larvae expressing an UAS-GFP transgene and the corresponding Gal4 driver were subjected to the same temperature schemes and used as controls. Expression of Ricincs, 4EBPAA, or PTEN caused a clear reduction in total wing area in all the genotypes analyzed (Figures 1B, C, S1 and Table S1). Interestingly, these wings conserved normal shape, proportions, and vein patterning. The maintenance of shape and wing proportions suggests that the size of the neighboring cell populations not expressing the transgene is reduced in a non-autonomous manner. In order to quantify these non-autonomous effects, we focused our attention on those Gal4 drivers expressed either in the anterior (A, ci-Gal4, ptc-Gal4, and dpp-Gal4) or posterior (P, en-Gal4, and hh-Gal4) compartments, cell populations that do not mix and give rise to defined structures of the adult wing ([15], e.g. Figure 1C). Both compartments, the one expressing the growth-reducing transgene and the adjacent one, decreased in size (Figures 1E and S1). A size reduction of 11%–14% was observed in the A compartment when Ricincs, 4EBPAA, or PTEN were expressed in P cells with the en-gal4 driver (p<10−3, Table S2). A similar reduction was detected in the P compartment (14%–32%) when these transgenes were expressed in A cells with the ci-Gal4 driver (p<10−2, Table S2). Similar results were obtained with hh-Gal4, ptc-Gal4, and dpp-Gal4 drivers (Figure S1). The observed non-autonomous effects in tissue size are most likely a local response to a signal coming from the growth-depleted territory rather than a more general response of the whole larvae induced by the insult to the imaginal cells, since expression of PTEN with a Gal4 driver expressed in the developing wing but not in other tissues (spaltPE-gal4, [16], Figure 1G) gave rise to smaller wings (82.3%±2.1%; p<10−9) without a large impact on the overall size of the animal, as visualized by quantification of adult weights (103%±8%; p = 0.02) and pupal lengths (101%±4%; p = 0.08) (Figure 1G–I and Table S1, see also [4]).
We next took advantage of the temperature sensitivity of the Ricincs transgene to conduct time-lapse experiments. Larvae expressing Ricincs in the A compartment with the ci-gal4 driver were initially grown at 18°C and switched to 29°C at different developmental stages until adult eclosion. The size of the resulting adult wings was measured. The reduction in size of the nearby P compartment ranged from 5% to 32% depending on when larvae were switched to the restrictive temperature (Figures 1D, S1 and Table S4). Longer exposure to Ricincs expression gave rise to the strongest phenotypes, thereby suggesting that the non-autonomous effects are operative throughout development (see below). We also noted that the non-autonomous reduction in size of the P compartment was proportional to the reduction in size of the A compartment (Figure 1D).
The non-autonomous reduction in tissue size could be a consequence of reduced cell growth, reduced cell number, or both. To address this issue, we quantified cell densities in the adult wing by counting the number of hairs (each cell differentiates a hair) in a defined area in the two compartments. Cell densities were slightly increased to various degrees in the transgene-expressing compartment (Figures 1F, S1). The increase in cell densities ranged from 1.7% to 19%, depending on the Gal4 driver and the UAS-transgene used (Table S3). In the non-expressing transgene compartment, a significant increase in cell densities was observed mainly in those wing discs with the highest reduction in tissue size (ci-gal4; UAS-Ricincs, 111.6%±2.2%; p<10−13). Thus, changes in cell size contributed in a smaller extent to the non-autonomous reduction in tissue size and mainly in those situations where this reduction was above a certain threshold.
The results presented so far indicate that growth depletion in defined territories of the wing induces a non-autonomous reduction in tissue size in nearby territories. In order to gain insight into whether this non-autonomous response is an active mechanism that takes place throughout development, we monitored the size of the A and P compartments throughout development after targeted expression of Ricincs in P cells. Larvae expressing Ricincs or GFP in the P compartment (with the en-Gal4 driver) were grown at 18°C and switched to 30°C from early second instar to late third instar stages. The size ratio between the transgene-expressing (P) and non-expressing (A) compartment was first quantified at a range of time points after induction of Ricincs expression. The P/A ratio was roughly maintained throughout the induction period in GFP-expressing wing discs (blue dots in Figure 2A). In Ricincs-expressing discs, this ratio was smaller in the first 48 h after transgene expression but reached a similar value in mature wing discs (red dots in Figure 2A, see also [17]). These results suggest that during development the Ricincs-expressing P compartment was relatively smaller than the nearby compartment, but both compartments reached a similar size ratio to that observed in control wing discs at the end of larval development. We next quantified and compared the absolute size of both compartments in GFP- and Ricincs-expressing wing discs. Interestingly, not only was the Ricincs-expressing P compartment already smaller 24 h after transgene expression, but also the adjacent A compartment showed a decrease in size when compared to GFP control wing discs (Figure 2B). As observed in adult wings, the A and P compartments of Ricincs-expressing mature wing discs were smaller than those of control GFP-expressing wing discs (Figure 2B), even though the Ricincs-expressing animals extended their larval period for about 24 h before entering metamorphosis (Figure S2).
We next induced neutral clones of cells at the beginning of the Ricincs induction period (early second instar) and examined the size of these clones 72 and 96 h later in third instar wing discs. The size of the clones in the A compartment was measured in en-gal4; UAS-Ricincs wing discs and was compared to the size of control clones induced in the A compartment of en-gal4; UAS-GFP wing discs subjected to the same temperature schemes. In Ricincs-expressing wing discs, the size of the clones visualized 72 or 96 h after clone induction was, respectively, one-half or two-thirds smaller than the size of clones quantified in GFP-expressing discs (Figure 2C). All together, the results presented so far indicate that growth rates are non-autonomously reduced when growth is depleted in defined territories of the developing wing.
In order to gain insight into whether cell proliferation rates are also regulated in a non-autonomous manner, we analyzed cell cycle progression in developing wing primordia exposed to transgene expression. Larvae expressing Ricincs in the P compartment (with the en-Gal4 driver) were grown at 30°C from early second (48 h AEL) to mid-late third instar stages (96 h AEL) and the expression of markers for each cell cycle stage was then analyzed. Larvae expressing GFP in the P compartment were subjected to the same temperature schemes and used as controls. S phase progression was first monitored. Imaginal discs were exposed to BrdU for 45 min, and its incorporation was subsequently analyzed. BrdU incorporation was strongly reduced in the adjacent A compartment, while it was not significantly affected in the Ricincs-expressing domain (Figure 2D, E). Similar non-autonomous effects in BrdU incorporation were observed by expression of PTEN or 4E-BPAA in the A or P compartment (with the en-Gal4 and ci-Gal4 drivers, respectively; Figure 2N–P). Consistent with the observation that the non-autonomous effects in tissue size were not exclusive to the A and P compartments (Figure 1), a non-autonomous reduction in BrdU incorporation levels was also observed when driving expression of Ricincs in other domains in the wing primordium (Figures 2M and S2). Similarly, the non-autonomous reduction in BrdU incorporation were also observed at earlier stages of wing development (Figure 2Q), thereby indicating that the non-autonomous effects in proliferation rates were operative throughout development (Figure 1D). The observed non-autonomous effects in BrdU incorporation are most likely a local response to a signal coming from the growth-depleted territory rather than a more general response of the whole larvae, since targeted expression of Ricincs with the ap-Gal4 driver, which is expressed in the dorsal compartment of the developing wing, caused a non-autonomous reduction in BrdU incorporation in wing cells while other imaginal tissues not expressing the transgene (e.g. eye) or expressing it late in development (e.g. leg) incorporated BrdU at normal levels (Figure 2R, S).
Next, using in situ hybridization, we monitored the expression levels of cyclin E (cycE) and string (stg, the Drosophila cdc25 homolog), two genes that act in wing disc cells as rate-limiting factors of G1/S and G2/M transitions, respectively [18]–[20]. Consistent with the non-autonomous reduction in BrdU incorporation levels, cycE mRNA levels were reduced in the A compartment of wing discs expressing Ricincs in P cells (with the en-Gal4 driver, Figure 2F, G). Interestingly, stg mRNA levels were also reduced in these cells (Figure 2H, I), suggesting that the G2/M transition was also compromised or delayed in a non-autonomous manner. Consistent with this view, mitotic activity, monitored with an antibody against a phosphorylated form of histone H3 at serine 10 (PH3) that labels mitotic figures, was reduced in these cells (Figure 2J, K). The number of PH3-positive cells observed in the A compartment of Ricincs-expressing wing discs decreased by 50% compared to GFP-expressing discs (p = 0.02, Figure 2L), whereas a similar number of PH3-positive cells was observed in the P compartment of both genotypes (p = 0.6, Figure 2L). Although we noted a slight increase in stg and cycE mRNA levels in cells expressing Ricincs (Figure 2G, I), no detectable changes in mitotic activity or BrdU incorporation were observed (Figure 2E, J, L).
The non-autonomous reduction in mitotic activity, BrdU incorporation, and cycE and stg mRNA levels observed suggest that a general reduction in proliferation rates, without any obvious arrest in any particular cell cycle stage, is non-autonomously induced when growth is impaired in a defined territory of the developing wing. In order to confirm this hypothesis, we used a fluorescence-associated cell sorter (FACS) to collect data about the cellular DNA content of dissociated cells from 96 h AEL wing discs and analyzed the cell cycle profile of these cells. A forward scatter (FSC) analysis was also carried out to compare cell sizes. Cell cycle profiles and cell sizes of A cells dissociated from wing discs expressing Ricincs and GFP or GFP alone (with the en-Gal4 driver) in the P compartment were very similar (Figure S2). Comparable results were obtained with the ci-Gal4 driver (Figure S2). All together these data indicate that upon growth depletion in defined territories of the wing primordium, the adjacent cell populations reduced their growth and proliferation rates, giving rise to smaller structures with a smaller number of cells. The slight non-autonomous reduction in cell size contributes to a small extent to the non-autonomous reduction in adult tissue size (Figure 1E, F) and is most probably occurring during post-larval stages, since no major change in cell size was observed in imaginal tissues (Figure S2).
We noticed that the levels of BrdU incorporation, stg and cycE expression, and mitotic activity were largely unaffected in the transgene-expressing compartment when compared to GFP control wing discs (Figure 2). This might reflect a process of compensatory proliferation due to the large number of cells being lost by cell death (see below) or by other means.
In mammalian cells, inhibition of the insulin/TOR pathway or inhibition of protein biosynthesis increases the activity of the tumor suppressor gene p53 (reviewed in [21]) and induces the activation of the apoptotic machinery that breaks down cells in a highly controlled fashion by the action of caspases, a specialized class of cysteine proteases. Non-apoptotic functions of activated caspases have been previously described, and Drosophila and vertebrate caspases have been reported to regulate cell proliferation in various ways (reviewed in [22], see also [23]). Thus, we first monitored the contribution of the apoptotic machinery to the non-autonomous regulation of tissue growth and cell proliferation rates observed in developing wing discs and resulting adult wings.
We performed a TUNEL assay to label DNA strand breaks induced by apoptotic cell death. A clear increase in TUNEL-positive cells was observed in those territories expressing Ricincs (Figure 3A, B), PTEN, or 4E-BPAA (Figure 3D, E) as well as in adjacent cells. We next used an antibody against the activated form of human Caspase 3, a marker of Caspase-9-like Dronc activity in Drosophila tissues [24]. We observed increased levels of Dronc activity in those territories expressing the transgene as well as in adjacent cells (Figure 3C and unpublished data). The increase in TUNEL-positive cells and Dronc activity raises the question of whether apoptosis participates in the non-autonomous response observed in developing wing primordia and adult wings. Caspase activities are regulated by inhibitor-of-apoptosis proteins (IAPs), and in Drosophila, DIAP1 binds and inhibits Dronc and the effector caspases DrICE and Dcp-1 (Figure 3Q, reviewed in [25]). The Drosophila pro-apoptotic genes hid, grim, and reaper bind and repress DIAP1, thus alleviating repression of initiator and effector caspases. In order to analyze the contribution of caspases to the non-autonomous effects in proliferation rates, we tested the requirement of various elements of the genetic cascade that drives apoptosis in Drosophila (Figure 3Q). When apoptosis was reduced in the whole tissue by halving the dose of hid, grim, and reaper (in Df(H99)/+ wing discs) or by depleting dronc expression (in droncl29 wing discs), a clear reduction in TUNEL-positive cells was observed upon Ricincs expression (compare Figure 3B with 3H, I). Interestingly, the non-autonomous reduction in BrdU incorporation levels caused by Ricincs expression was largely rescued in these discs (Figure 3M, N and compare with Figure 2E). We then expressed DIAP1 and p35 in the same domain as Ricincs to analyze whether caspase activation is required in the growth-depleted territory or in the neighboring cell populations. DIAP1 represses both DrIcE and Dronc, while the baculovirus protein p35 specifically represses effector caspases DrIce and Dcp-1 and maintains fully active Dronc (reviewed in [25], Figure 3Q). Expression of either DIAP1 or p35 in the same domain as Ricincs caused a clear autonomous reduction in the number of TUNEL-positive cells (Figure 3J and unpublished data) as well as a clear non-autonomous rescue of BrdU incorporation levels (Figure 3O and unpublished data). A similar non-autonomous rescue in the expression levels of cycE and string mRNA and in the number of mitotic figures was observed upon p35 expression (Figure 3K, L, P). Similar results were obtained when PTEN or 4E-BPAA were co-expressed with p35 in different domains of the wing disc (Figure 3F, G). We then used FACS and FSC analysis to characterize the cell cycle profile and the size of A cells upon expression of Ricincs and GFP, or Ricincs and p35 in the P compartment (with the en-Gal4 driver, Figure S3). Similar cell cycle profiles and cell sizes were obtained in both genotypes. Comparable results were obtained with the ci-Gal4 driver (Figure S3). The capacity of p35 expression to rescue the non-autonomous reduction in BrdU, string, and cycE levels and in mitotic activity caused by Ricincs expression, even though Dronc is fully active under these conditions [25], suggests that effector caspases, such as DrIce and Dcp-1, are required in the growth-depleted territory to induce a non-autonomous reduction in cell proliferation rates in the nearby cell populations.
We next analyzed the resulting adult wings when the activity or expression of various elements of the apoptotic machinery was depleted in Ricincs-expressing larvae. The autonomous reduction in tissue size was not rescued and the cell densities were either unaffected or increased when apoptosis was reduced (Figure 3R, S), thereby suggesting that effector caspases do not play a major role in the autonomous reduction in tissue size caused by Ricincs expression. Surprisingly, the non-autonomous reduction in tissue size was not rescued either (Figure 3R). Nevertheless, and consistent with the role of effector caspases in regulating cell proliferation rates in nearby territories, cell densities were significantly increased in these territories in all the genetic backgrounds tested (Figure 3S). The non-autonomous reduction in cell size is most probably occurring during post-larval stages, since no major change in cell size was observed in imaginal tissues (Figure S2). These results imply that growth and proliferation rates are independently regulated and that the decrease in cell proliferation rates does not play a major role in the observed reduction in tissue size. While effector caspases are required in the growth-depleted territory to induce a non-autonomous reduction in cell proliferation rates in the nearby cell populations, the non-autonomous reduction in tissue size relies on a caspase-independent mechanism.
The transcription factor and tumor suppressor p53, a short-lived, non-abundant protein in healthy cells, plays a major role in regulating the response of mammalian cells to stress, in part through the transcriptional activation of genes involved in apoptosis and cell cycle regulation [26]. Impaired TOR signaling, ribosomal biogenesis, and protein translation increase p53 activity [21],[27]. Although the regulation of dp53 in Drosophila has not been fully elucidated, the biological function of p53 is well-conserved between flies and mammals [28]. dp53 mediates a variety of stress responses by inducing the expression of the pro-apoptotic gene reaper [29],[30]. Interestingly, expression of reaper was induced in Ricincs- and in PTEN-expressing cells and this expression depended on the activity of dp53 (Figure 4A-D). We did not find any evidence of increased dp53 mRNA levels in the transgene-expressing cells (unpublished data), thus suggesting that the activation of dp53 is post-transcriptional. The increase in the number of TUNEL-positive cells caused by expression of Ricincs was largely rescued by reducing the activity of dp53 in the whole wing disc (in dp53ns mutant larvae) or in the transgene-expressing domain (co-expressing dp53DN or dp53dsRNA; Figure 4E–H). Most interestingly, the effects of Ricincs on the levels of BrdU incorporation and on the number of mitotic PH3-positive cells were also largely rescued by dp53 depletion (Figure 4I–M). These results indicate that dp53 and the effector caspases are required in the growth-depleted territory to non-autonomously reduce the proliferation rates in neighboring cell populations.
We next analyzed the resulting adult wings when the activity or expression of dp53 was depleted in the domains expressing Ricincs. Consistent with the observation that the apoptotic machinery and effector caspases do not play a major role in the autonomous reduction in tissue size caused by expression of these transgenes, the autonomous reduction in tissue size was still observed and the cell densities were either unaffected or increased when dp53 activity was reduced (Figure 5A). However, and in contrast to the apoptotic machinery and the effector caspases, dp53 exerts a fundamental role in the non-autonomous regulation of tissue size. The non-autonomous reduction in tissue and cell size caused by Ricincs expression was either weaker or not observed when dp53 activity was depleted in the transgene-expressing domain (Figure 5A), and the resulting adult wings were not well-proportioned and exhibited a clear asymmetric shape with respect to the AP boundary (Figure 5B). In other words, in a situation of reduced dp53 activity in the Ricincs-expressing compartment, the neighboring compartment exhibited a size nearly identical to the one observed in GFP control wings. These results indicate that, upon growth depletion, dp53 exerts a fundamental and non-autonomous role in reducing the size of the adjacent cell populations. In order to address whether growth rates are also regulated in a non-autonomous manner by the activity of dp53, we induced neutral clones of cells at the beginning of the Ricincs induction period (early second instar) and examined, 72 h and 96 h later, the size of clones located in the A compartment of wing discs expressing GFP, or Ricincs and GFP, or Ricincs and dp53dsRNA in the P compartment (with the en-gal4 driver, Figure 4N). Interestingly, the non-autonomous reduction in clone size caused by Ricincs expression was largely rescued by co-expression of dp53dsRNA (Figure 4N).
The Drosophila wing primordium increases about a thousand-fold in cell mass and cell number during larval development. Even though molecules of the Wnt, TGF-β, and Hedgehog families are well known to organize tissue growth and patterning of this primordium and to play a fundamental role in the generation of an adult wing with a species-specific size, shape, and patterning [31], the molecular mechanisms that contribute to buffering local variations in tissue growth caused by different types of stress and help to generate well-proportioned adult wings under these circumstances remain largely unknown. Here we underscore a new role of dp53 and the apoptotic machinery in these processes. Depletion of the insulin pathway or the protein biosynthetic machinery in defined territories of the developing wing primordium induces an autonomous reduction in growth rates as well as a non-autonomous decrease of growth and cell proliferation rates in the adjacent cell populations. This non-autonomous response occurs independently of where growth is affected and is functional throughout development. We present evidence that targeted depletion of the insulin pathway or the protein biosynthetic machinery induces the activation of dp53 and consequently the apoptotic machinery. While growth and proliferation rates are regulated in a coordinated and non-autonomous manner by the activity of dp53, effector caspases have a unique role, downstream of dp53, in reducing proliferation rates in adjacent cell populations (Figure 5C). Thus, tissue growth and proliferation rates can be uncoupled and are regulated by two different mechanisms downstream of dp53. These new findings underscore a new role of dp53 and effector caspases in buffering local variations in tissue growth and in maintaining the relative proportions of distinct cell populations within a tissue to give rise to a functional adult structure.
Independent lines of evidence support the view that adjacent cell populations are able to buffer local variations in tissue growth caused by different means, and not only when the activities of the insulin pathway or the protein biosynthetic machinery are compromised. The halteres and wings of Drosophila are homologous thoracic appendages, which share common positional information provided by signaling pathways. The activity in the haltere discs of the Ultrabithorax (Ubx) Hox gene establishes the differences between these structures, their different size being an obvious one (reviewed in [32]). In Contrabithorax mutant wings in which one compartment is reduced in size due to the transformation to haltere by the ectopic expression of Ubx, the adjacent compartment adjusts its final size and results in a smaller wing territory [33]. Similar non-autonomous effects in tissue size were observed when inducing clones of cells with reduced EGF-Receptor activity [34], upon depletion of the dMyc proto-oncogene or over-expression of the hippo tumor-suppressor gene (Figure S5). Nevertheless, we cannot rule out the possibility that in other situations in which growth is being challenged, the tissue might not respond as a whole. In these situations, we speculate that dp53 is inactive or not activated. We would like also to point out that the dp53 dependent mechanism described in this work might be functional to buffer local and slight variations in growth rates. Above a certain threshold in the reduction of tissue size, this mechanism might not be sufficient to generate well-proportioned organs.
In vertebrates and invertebrates, p53 and caspase-dependent cell death also play a fundamental role in regeneration (reviewed in [35]) as well as in response to stress or tissue damage, whereby the damaged tissue undergoes extra cell proliferation to compensate for cell loss [23],[36]–[38]. Upon tissue damage in Drosophila tissues, a feedback loop mediated by dp53 and the initiator caspase Dronc is required in undifferentiated dying cells to induce cell proliferation in surrounding cells [30]. A specific role of effector caspases has been described in differentiated neurons to induce, upon tissue damage, cell proliferation in surrounding cells [39]. These non-autonomous effects are mediated by the ectopic activation of signaling molecules of the Wnt, TGF-β, and Hedgehog families. In contrast, we did not find any clear change in the expression of Wnt, TGF-β, or Hedgehog in the growth-depleted territory or in the activity of their signaling pathways in adjacent cell populations (Figure S4 and unpublished data). Also, while dp53 and caspases have a role in inducing cell proliferation within the damaged tissue [30], our data indicate that the same molecules reduce growth and proliferation rates in adjacent unaffected cell populations. These observations suggest that different types of signaling molecules induced or activated by dp53 and caspases exert different effects in the damaged tissue and in nearby ones. It is interesting to note in this context that effector caspases have been recently shown to promote wound healing and tissue regeneration in mice by mediating the cleavage of iPLA2 (Calcium-independent Phospholipase A2) to trigger the production of the growth signal Prostaglandin E2 [23]. Whether the non-autonomous reduction in cell proliferation rates is due to the caspase-dependent cleavage and activation of specific signaling molecules or whether dying cells release a variety of signaling molecules as a consequence of cell demolition [40] that could have a general role in the reduction of cell proliferation rates remains to be solved. Similarly, the signaling molecules that mediate the non-autonomous role of dp53 in regulating tissue growth remain to be identified.
Finally, we would like to highlight the central role of p53 in tissue homeostasis and stress response. While the role of p53 in regulating the response of mammalian cells to stress through the transcriptional activation of genes involved in apoptosis and cell cycle arrest is cell-autonomous [41], we speculate that the non-autonomous role of dp53 defined in this study also makes a major contribution to stress response and tissue homeostasis. Upon tissue damage, neighboring populations of healthy cells might reduce their growth and proliferation rates until damaged cells have been recovered. The combined autonomous and non-autonomous activities of p53 might be fundamental in tissue homeostasis.
The Drosophila strains include: UAS-Ricincs ([10], a cold sensitive version of the Ricin toxin A Chain), UAS-4EBPAA (a 4E-BP constitutively active form with two threonine-to-alanine substitutions (T37A, T46A) [12]), UAS-PTEN [13]; salPE-Gal4 [16]; UAS–dp53RNAi (ID 10692, VDRC); and dp53ns (a knock out of dp53 created by ends-in homologous recombination, Flybase). Other stocks are described in Flybase or Text S1.
Mouse anti-BrdU (Developmental Studies Hybridoma Bank); rabbit anti-PH3 (Upstate Biotechnology); rabbit anti-GFP (Abcam); rabbit anti-cleaved-Caspase 3 (Asp175, Cell Signaling Technology). Other antibodies are described in Text S1. Secondary antibodies were obtained from Molecular Probes. TUNEL analysis, BrdU staining, and in situ hybridization were performed as described in [20]. A Digoxigenin-RNA Labeling kit (Roche) was used to synthesise probes of string, cycE, and reaper.
Wing size and size of the A and P compartments were measured using the Image J Software (NIH, USA). Cell density was measured as the number of hairs (each wing cell differentiates a hair) per defined area. Two conserved regions between veins L4 and L5 (P compartment) and veins L2 and L3 (A compartment) in both dorsal and ventral surfaces of the wings were used to measure cell densities. The final area and cell density values were normalized as a percent of the control Gal4-driver; UAS-GFP values. At least 10 adult wings per genotype were scored. Only adult males were scored. The average values and the corresponding standard deviations were calculated and t test analysis was carried out.
en-Gal4/SM6a-TM6b/UAS-Ricincs or en-Gal4, UAS-GFP females were crossed with w1118/Y males and allowed to lay eggs for 5 h at 18°C. After 6 days at 18°C, the resulting larvae were transferred to 29°C. Wing discs were dissected after different time points of Ricincs induction and tissue size was measured from confocal images by means of Image J software (NIH, USA). At least 10 wing discs were used per time point. Average values and the corresponding standard deviations were calculated and t test analysis was carried out. Note that, consistent with the temperature sensitivity of Ricincs, en-gal4;UAS-Ricincs adult wings grown at the restrictive temperature (18°C) showed a size nearly identical to the one of en-gal4;UAS-GFP adult wings (Figure S1),
hs-FLP, ubi-GFP, FRT19 or hs-FLP, ubi-GFP, FRT19; +/+; UAS-p53RNAi females were crossed with control FRT19/Y and experimental FRT19/Y; en-Gal4/SM6a-TM6b/UAS-Ricincs males and allowed to lay eggs for 5 h at 18°C. After 6 days at 18°C, the resulting larvae were heat-shocked at 37°C for 30 min and then transferred to 29°C. Wing discs were dissected 72 h and 96 h after clone induction. The size of clones was quantified from confocal images with Image J software (NIH, USA). At least 15 clones were quantified per time point. Average values and the corresponding standard deviations were calculated and t test analysis was carried out.
Wing discs of larvae carrying different UAS transgenes and the en-Gal4 or ci-gal4 drivers were dissected and cells were dissociated according to the protocol described in [42]. Hoescht and GFP fluorescence were determined by flow cytometry using a MoFlo flow cytometer (DakoCytomation, Fort Collins, CO). Excitation of the sample was carried out using a Coherent Enterprise II argon-ion laser. Excitation with the blue line of the laser (488 nm) permits the acquisition of forward-scatter (FS), side-scatter (SS), and green (530 nm) fluorescence from GFP. UV emission (40 mW) was used to excite Hoescht blue fluorescence (450 nm). Doublets were discriminated using an integral/peak dotplot of Hoechst fluorescence. Optical alignment was based on optimized signal from 10 µm fluorescent beads (Flowcheck, Coulter Corporation, Miami, FL). DNA analysis (Ploidy analysis) on single fluorescence histograms was done using Multicycle software (Phoenix Flow Systems, San Diego, CA).
Forty hatched first-instar larvae were sorted from egg laying plates (egg laying period of 5 h), transferred to fresh tubes, and kept at 18°C. Then, animals of the different genotypes were transfer to 29°C from early second instar and the number of pupae was counted daily. Results were expressed as percentage of the individuals from each genotype that attained the pupal stage.
|
10.1371/journal.ppat.1001153 | Functional Interchangeability of Late Domains, Late Domain Cofactors and Ubiquitin in Viral Budding | The membrane scission event that separates nascent enveloped virions from host cell membranes often requires the ESCRT pathway, which can be engaged through the action of peptide motifs, termed late (L-) domains, in viral proteins. Viral PTAP and YPDL-like L-domains bind directly to the ESCRT-I and ALIX components of the ESCRT pathway, while PPxY motifs bind Nedd4-like, HECT-domain containing, ubiquitin ligases (e.g. WWP1). It has been unclear precisely how ubiquitin ligase recruitment ultimately leads to particle release. Here, using a lysine-free viral Gag protein derived from the prototypic foamy virus (PFV), where attachment of ubiquitin to Gag can be controlled, we show that several different HECT domains can replace the WWP1 HECT domain in chimeric ubiquitin ligases and drive budding. Moreover, artificial recruitment of isolated HECT domains to Gag is sufficient to stimulate budding. Conversely, the HECT domain becomes dispensable if the other domains of WWP1 are directly fused to an ESCRT-1 protein. In each case where budding is driven by a HECT domain, its catalytic activity is essential, but Gag ubiquitination is dispensable, suggesting that ubiquitin ligation to trans-acting proteins drives budding. Paradoxically, however, we also demonstrate that direct fusion of a ubiquitin moiety to the C-terminus of PFV Gag can also promote budding, suggesting that ubiquitination of Gag can substitute for ubiquitination of trans-acting proteins. Depletion of Tsg101 and ALIX inhibits budding that is dependent on ubiquitin that is fused to Gag, or ligated to trans-acting proteins through the action of a PPxY motif. These studies underscore the flexibility in the ways that the ESCRT pathway can be engaged, and suggest a model in which the identity of the protein to which ubiquitin is attached is not critical for subsequent recruitment of ubiquitin-binding components of the ESCRT pathway and viral budding to proceed.
| The release of an enveloped virus particle from an infected cell requires the separation of the viral and cell membranes. Many enveloped viruses accomplish this by parasitizing a set of cellular proteins, termed the ESCRT pathway, that normally separates cellular membranes from each other. In some cases, viral structural proteins encode peptides motifs that bind directly to, and thereby recruit, the ESCRT machinery. Alternatively, viruses can recruit enzymes, termed ubiquitin ligases, that bind to other proteins, and catalyze the addition of ubiquitin to them. It has, heretofore, been somewhat unclear precisely how the recruitment of ubiquitin ligases leads to the engagement of the ESCRT machinery. We show that the simple recruitment of a fragment of a ubiquitin ligase that is responsible for the addition of ubiquitin to other proteins is sufficient to drive virus particle release, even when it is not possible to attach ubiquitin to viral proteins. Paradoxically, we also found that simple attachment of ubiquitin to the same viral protein can also drive particle release. These results show that there is flexibility in the ways in which the ESCRT machinery can be recruited and how ubiquitin can be co-opted to enable this.
| The membrane scission event that separates the lipid membrane of nascent enveloped virions from host cell membranes is, in many cases, an orchestrated event requiring the participation of the class E vacuolar protein sorting (VPS), or endosomal sorting complex required for transport (ESCRT) pathway. Ordinarily, the ESCRT pathway induces topologically equivalent cellular membrane scission events including the biogenesis of multivesicular bodies (MVBs) [1], [2] and the membrane abscission event at the conclusion of cell division [3], [4]. Components of the pathway can be recruited, either directly or indirectly, through the action of short peptide motifs called late (L-) domains in viral structural proteins [5], [6]. Three classes of viral L-domains and cognate cofactors have been defined thus far: PT/SAP motifs bind Tsg101 [7], [8], [9], [10], LxxLF or YPXL motifs bind ALIX [11], [12], [13], and PPxY domains bind Nedd4-like HECT ubiquitin ligases [14], [15], [16], [17], [18]. Disruption of late domain function results in the failure of membrane scission and the accumulation of assembled virions that remain tethered to the surface of the host cell by a continuous membrane.
The ESCRT machinery is composed of ∼25 proteins, many of which participate in the formation of several multiprotein complexes, known as ESCRT-0, -I, -II, -III [19], [20], [21]. ESCRT-III components are thought to drive the membrane scission event [22], [23], [24], [25] and appear to be generally required for L-domain-dependent viral budding [7], [11], [12], [13], [26]. In contrast, other components of the ESCRT-pathway appear to be required in an L-domain specific way. For example, PTAP-dependent budding is especially sensitive to ESCRT-I perturbation, while YPXL-dependent budding is especially sensitive to ALIX depletion. Since ALIX interacts directly with ESCRT-III via its Bro1 domain [11], [12], [13], [27], [28] and ESCRT-I indirectly interacts with ESCRT-III via ALIX and/or ESCRT-II, [11], [12], [13] these observations suggest that YPXL and PTAP motifs access the same core scission machinery via alternative routes.
In contrast, it has remained somewhat unclear how PPxY motifs access the scission machinery. Overexpression of certain HECT ubiquitin ligases that bind directly to PPxY or other motifs can markedly stimulate budding, and the catalytic activity of the HECT domain is essential for this activity [17], [29], [30], [31]. Indeed, overexpression of catalytically inactive or truncation mutants of the HECT ligase WWP1 inhibits PPxY-dependent budding [17], [29]. Some components of the ESCRT pathway are also required for PPxY-induced budding [7], [31], [32]. However, the precise means by which HECT ligase recruitment subsequently results in the engagement of the ESCRT machinery is not completely defined. One model invokes direct ubiquitination of Gag as the key event. This notion derives from observations that several components of the ESCRT pathway are thought to recognize ubiquitinated cargo through various low affinity ubiquitin-binding domains [7], [33], [34], [35], [36], [37] and that monoubiquitination of cellular cargos can serve as a signal for endosomal trafficking and delivery to the lysosome [21], [38], [39]. Indeed, several observations are consistent with the notion that ubiquitination of retroviral Gag promotes virus particle release. For example, studies have noted an enrichment of free ubiquitin in retrovirus particles, and ubiquitinated Gag species have also been detected therein [14], [40], [41], [42], [43]. Additionally, late budding defects have been observed in cells treated with proteasome inhibitors, perhaps due to the depletion of free ubiquitin [14], [44], [45]. Mutation of multiple ubiquitin acceptor lysine residues in Gag has been shown to inhibit particle production by retroviruses [46], [47]. Finally, direct fusion of ubiquitin to the C-terminus of Gag proteins has been shown to alleviate inhibition of particle release imposed by proteasome inhibitors, or to obviate the requirement for an L-domain in particle release [44], [48].
Other observations suggest that PPxY and ubiquitin ligase-dependent budding may involve mechanisms other than direct Gag ubiquitination. In particular, overexpression of wild-type WWP1 stimulates PPxY-dependent particle production by a lysine-free Gag protein [29] in the absence of detectable Gag ubiquitination. This finding suggests the possibility that HECT ligases may promote budding by catalyzing the ubiquitination of specific trans-acting host factors, rather than Gag. Additionally, a HECT-truncated WWP1 protein, lacking the entire HECT domain, inhibits murine leukemia virus (MLV) budding more potently than the full length WWP1 protein with a disrupted active site [17], suggesting that HECT domains may possess activities other than ubiquitin conjugation that are important for their function in viral budding. Moreover, HECT domains localize to aberrant endosomal (so called class E) compartments induced by overexpression of catalytically inactive ATPase VPS4 [17], which is required for the disassembly of ESCRT complexes after each round of budding [49], [50]. Since many VPS factors are trapped on VPS4-induced compartments, HECT domains may be recruited to these compartments by interaction with VPS proteins, either directly or through unidentified bridging factors. It has also been reported that HECT ubiquitin ligases can bind to, and/or catalyze the ligation of ubiquitin to, certain class E VPS factors [31], [32], [51]. Thus, the ubiquitin ligases might act as recruitment factors rather than, or in addition to, conjugating ubiquitin to key target proteins.
In this study we investigated the role of PPxY motifs, HECT ubiquitin ligase domains and ubiquitin in viral budding, using a lysine-free viral protein from the prototypic foamy virus (PFV), in which the attachment of ubiquitin to Gag can be rather precisely controlled. We show that the catalytic activity of a variety of HECT domains, targeted to a PPxY motif in assembling particles via a common C2/WW domain fragment of WWP1, is essential for their ability to promote PPxY-dependent VLP release. In each case, however, Gag ubiquitination is dispensable for their activity. Rather, the ability of the chimeric ubiquitin ligases to promote budding correlated broadly, albeit imperfectly, with their ability to catalyze autoubiquitination, Moreover, we show that artificial recruitment of an isolated HECT domain can also stimulate budding, while a HECT domain becomes dispensable for PPxY motif dependent budding if the C2/WW domains of WWP1 are directly linked to the C-terminal domain of Tsg101, an ESCRT-I subunit. Finally, we demonstrate that direct fusion of a single ubiquitin moiety to the C-terminus of PFV Gag is also capable of promoting budding, in a manner that recapitulates the ESCRT protein requirement for budding induced by PPxY-dependent ubiquitin ligase recruitment in the absence of ubiquitin acceptors in Gag. These results support a model in which PPxY motif-induced HECT ubiquitin ligase recruitment leads to the deposition of ubiquitin at or near the site of viral budding. However, the identity of the protein to which ubiquitin is attached, be it Gag or a bystander protein, perhaps including the HECT ubiquitin ligase itself, does not appear to be critical in order for subsequent recruitment of ubiquitin-binding class E VPS proteins and viral budding to proceed.
To ascertain what properties of HECT domains are important for stimulation of virus particle release, we compared the properties of a panel of HECT domains. Nine members of the Nedd4-like HECT ubiquitin ligase family have been described in humans and these have the same domain organization as a single prototype member of this family in yeast, namely Rsp5 (reviewed in [52]). Specifically, an N-terminal C2 domain directs the protein to membranes, a central cluster of ‘WW” domains binds ligands, such as PPxY motifs, and a C-terminal HECT domain harbors the E3 ubiquitin ligase activity. Some of the intact ubiquitin ligases have been shown to vary in their ability to promote PPxY-dependent MLV virion release, due at least in part to differences in the affinities of their WW domains for the MLV L-domain [17], but whether the various the C-terminal HECT domains are equivalently able to induce particle release has not been investigated. We reasoned that variation in the ability of HECT domains to stimulate virus budding, correlated with a given property of the HECT domains, might suggest properties that are important for inducing virion release. Since WWP1 has been previously shown to be efficiently recruited by a number of PPxY-type L-domains, including that of MLV [17], we constructed a panel of chimeric ubiquitin ligases, consisting of membrane targeting and PPxY motif-binding domains (C2 and WW domains) of human WWP1, coupled to various catalytic HECT domains derived from human WWP2, Nedd4, Nedd4L, Itch, Smurf1, Bul2 or yeast Rsp5 HECT ligases (Fig. 1A).
To determine whether these chimeric ubiquitin ligases could support viral budding, we co-expressed each of them with a plasmid expressing a modified PFV Gag protein. Importantly, PFV Gag offers the advantage that it is naturally almost devoid of lysine resides. While PFV Gag normally requires a cognate Env protein for particle release, we have previously shown that appending a myristoylated, palmitoylated peptide from Lck at its N-terminus can overcomes this requirement by directing PFV Gag to the plasma membrane and thereby allowing the generation of extracellular particles in the absence of any other viral protein [29]. Throughout this study we used this N-terminally modified Gag protein, termed Lck-Gag, bearing a K396R mutation that renders the PFV Gag completely lysine-free. Examples of engineered variants of this Gag protein are illustrated in Fig. 1B, and include those that is otherwise unmodified and encode the natural PSAP late domain (Lck-Gag(PSAP)), a PSAP mutant that contains no known L-domain (Lck-Gag(L-)) or another variant that has a PPxY late domain derived from MLV Gag appended to its C-terminus (Lck-Gag-PY, Fig. 1B). In addition, we used an Lck-Gag-PY derivative containing three lysine residues adjacent to a PPxY late domain (Lck-Gag-PY-3K) to assess HECT ligase-induced Gag ubiquitination ([29], illustrated in Fig. 1B).
Overexpression of ubiquitin ligases encoding a variety of HECT domains (WWP1 itself, WWP1/Nedd4, WWP1/Nedd4L, WWP1/Itch, WWP1/Smurf1, or WWP1/Bul2) stimulated PPxY-dependent budding of lysine-free Lck-Gag-PY (Fig. 2A,B). Conversely, WWP1/WWP2 and WWP1/Rsp5 did not stimulate budding or had marginal activity. The strongest stimulation was observed using chimeric ligases containing the Nedd4L and Itch HECT domains. Importantly, overexpression of chimeric ligases in which the catalytic cysteine was mutated to serine, failed to stimulate PPxY-dependent particle release (Fig. 2A), indicating that the catalytic activity of each HECT domains was required, even when the viral structural proteins lack ubiquitin acceptors.
To assess the relative catalytic activities of the chimeric HECT ligases, and assess whether this correlated with their differential ability to promote budding, we compared their abilities to carry out autoubiquitination and to ubiquitinate a Gag substrate encoding three lysine residues in close proximity to a PPxY late domain (Lck-Gag-PY-3K, see Fig. 1B). To accomplish this, we immunoprecipitated either Gag or HECT ubiquitin ligases from 293T cell lysates, prepared 36 hours after co-transfection with plasmids expressing Lck-Gag-PY-3K, HA-tagged ubiquitin, and each of the YFP-fused chimeric HECT ligases. Cell lysates were prepared using denaturing, detergent-rich buffer (containing 0.5% SDS) to ensure dissolution of protein complexes, and ubiquitinated species were detected by immunoprecipitation with either αPFV Gag or αGFP antibodies followed by immunoblot analysis of the precipitates with an αHA antibody (Fig. 3).
Each of the chimeric HECT ubiquitin ligases was able to reasonably efficiently catalyze the addition of 1 to 3 ubiquitin moieties to the Lck-Gag-PY-3K substrate (Fig. 3A, upper panels). There was some variation in the ability of the HECT domains to catalyze the ligation of ubiquitin to Lck-Gag-PY-3K, with WWP1/Rsp5 and WWP1/Bul2 catalyzing the highest and WWP1/Nedd4 the lowest levels of ubiquitin ligation to Lck-Gag-PY-3K (Fig. 3A). However, there was no correlation between the extent to which each HECT domain stimulated Lck-Gag-PY-3K ubiquitination (Fig. 3A) and the degree to which it stimulated the release of VLPs assembled using Lck-Gag-PY or Lck-Gag-PY-3K (Fig. 2A and data not shown). For example, WWP1/Bul2 and WWP1/Nedd4, which induced the highest and lowest levels of Gag ubiquitination, respectively (Fig. 3A), stimulated budding to a similar extent (about 6-fold, Fig. 2A). Moreover, WWP1/Rsp5, which efficiently catalyzed Gag ubiquitination (Fig. 3A), enhanced particle release only marginally (Fig. 2A), much less efficiently than the WWP1/Nedd4L that induced comparatively modest levels of Gag ubiquitination (Fig. 3A).
We observed a better, albeit imperfect, correlation between the ability of the chimeric HECT ligases to catalyze autoubiquitination and to stimulate VLP production (Fig. 3B, Fig. 2B). Chimeric ligases that strongly promoted Lck-Gag-PY VLP release (e.g. WWP1/Itch and WWP1/Nedd4L) were more heavily autoubiquitinated, while those that failed or only marginally promoted VLP release (WWP1/WWP2 and WWP1/Rsp5, Fig. 2) exhibited the lowest levels of autoubiquitination (Fig. 3B). The correlation was imperfect, however, since WWP1/Nedd4, which moderately enhanced particle release (Fig. 2), was consistently highly auto-ubiquitinated (Fig. 3B). Notably, there was no correlation between the ability of the HECT ubiquitin ligases to catalyze autoubiquitination, and their ability to catalyze ubiquitin ligation to Lck-Gag-PY-3K (Fig. 3A, B). Overall, these data confirm our previous finding that direct Gag ubiquitination is dispensable for HECT ligase-dependent budding [29] and further indicates that intrinsic catalytic activity of the HECT ubiquitin ligases is critical for their ability to stimulate budding.
We next asked whether the need to recruit a HECT domain in the context of PPxY/WWP1 interaction was necessary for particle release, or whether the HECT domain could be bypassed by direct recruitment of putative downstream effectors. Additionally, we asked whether recruitment of a HECT domain in the absence of the other domains (C2 and WW) found in the Nedd4-like family of proteins was sufficient to stimulate particle budding. To accomplish this, we constructed hybrid L-domain cofactors in which the essential domains were split and linked to putatively complementing domains in another L-domain cofactor (Fig. 4A). Specifically, Tsg101 is a core component of ESCRT-I and contains two domains that are functionally important with respect to viral budding. The N-terminal ubiquitin E2 variant (UEV) domain interacts directly with P(T/S)AP peptide motifs and ubiquitin [7], while the C-terminal portion of the protein is a key structural component of ESCRT-I, interacting with other components, e.g. VPS28 and VPS37 [53], [54], [55] and is essential to support Tsg101 dependent budding. We constructed an artificial putative chimeric L-domain cofactor in which the C2/WW domains of WWP1 were linked to the C-terminal portion of Tsg101 (Tsg-C) that constitutes the core structural component of ESCRT-I (residues 157–390, Fig. 4A). Notably, overexpression of this chimeric protein, termed WWP1-Tsg-C, stimulated Lck-Gag-PY particle release in a dose-dependent manner but had no effect on particle production by the L-domain-deficient Lck-Gag(L-) protein (Fig. 4B, left and middle panels). This chimeric protein, therefore, appeared capable of recruiting a functional ESCRT-I complex to PPxY L-domains and thereby stimulating particle production. Conversely, WWP1-Tsg-C overexpression inhibited Lck-Gag(PSAP) budding in a dose-dependent manner (Fig. 4B, right panel). We surmise that since this chimeric protein lacks the domains required for interaction with PT/SAP motifs, it acts as an inhibitor of PSAP-dependent budding by sequestering endogenous components (e.g. VPS28 and VPS37) into retargeted ESCRT-I complexes that can be recruited to PPxY, but not PT/SAP, L-domains. Thus, these experiments demonstrate that the requirement for a HECT domain (and, by inference, the requirement for ubiquitin ligation) in PPxY/ubiquitin ligase dependent viral budding can be bypassed, if an alternative link to the ESCRT machinery is provided.
In a reciprocal experiment, we asked whether the PPxY motif and the C2/WW domains of WWP1 could be functionally replaced in the context of HECT domain/ubiquitin dependent budding. In other words, we determined whether recruitment of a HECT domain is sufficient to stimulate particle release, in the absence of the other protein domains (C2 and WW) to which it would ordinarily be linked. Specifically, we attempted to redirect P(T/S)AP-dependent particle production through a HECT domain-dependent pathway, by constructing chimeric proteins, termed Tsg-WWP1, Tsg-Itch and Tsg-Nedd4L, that contained the N-terminal UEV domain (residues 1–157) of Tsg101 linked to one of the three respective HECT domains (Fig. 4A). To test the function of these artificial putative L-domain cofactors, we also constructed an attenuated “leaky” mutant of the PT/SAP motif in the Lck-Gag(PSAP) protein, namely Lck-Gag(ASAP), by mutating the first proline residue of the PSAP motif to alanine. In the context of the HIV-1 PTAP motif, such a mutation reduces the affinity for, but does not eliminate binding to the Tsg101 UEV domain [7]. Correspondingly, the budding of Lck-Gag(ASAP), was attenuated as compared to Lck-Gag(PSAP), but the ASAP motif clearly retained some weak residual ability to stimulate budding (Fig. 4C, leftmost three lanes), suggesting that it retains some residual ability to recruit the Tsg101 UEV domain.
Overexpression of Tsg-WWP1, Tsg-Itch or Tsg-Nedd4L, respectively) resulted in clear stimulation of Lck-Gag(ASAP) budding (Fig. 4C). Tsg101-Itch was the most potent of the three Tsg101-HECT proteins tested by this approach, and its overexpression resulted in a particle yield that matched or even exceeded that observed in the presence of the intact PSAP motif (Fig. 4C). In contrast, expression of catalytically inactive versions of Tsg-WWP1, Tsg-Itch or Tsg-Nedd4L inhibited rather than enhanced Lck-Gag(ASAP) particle production (Fig. 4C). Because the Tsg101 UEV domain contains ubiquitin-binding activity that might complicate the interpretation of these results, we repeated these experiments using a mutant Tsg101 UEV domain (N45A) that is defective for ubiquitin binding, linked to a WWP1 HECT domain. The mutant Tsg(N45A)-WWP1 fusion stimulated budding at least as efficiently as did the unmanipulated Tsg-WWP1 protein (Fig. 4D).
Overall, the experiments in Fig. 4 demonstrate that the domains of the PTAP and PPxY binding cofactors can be functionally split into modular, interchangeable domains that are (i) necessary for binding to the L-domain and (ii) interface with downstream effectors that are critical for budding. Most notably, these findings suggest that simple recruitment of a HECT domain to sites of particle budding, irrespective of its mode of recruitment, and in the absence of ubiquitin acceptors on the viral protein, is sufficient to stimulate particle release and that other HECT ubiquitin ligase domains are dispensable for budding.
The aforementioned experiments demonstrated that the requirement for a catalytically active HECT domain could be obviated by direct recruitment of ESCRT-I to a viral protein (Lck-Gag-PY) whose budding would normally be dependent on such recruitment. We next asked whether the requirement for HECT domain recruitment could similarly be obviated, in the context of a nearly identical viral protein, by simply depositing ubiquitin at the site of particle assembly, in the absence of ubiquitin ligase recruitment. To mimic the deposition of ubiquitin at sites of virion assembly, in the absence of ubiquitin ligase recruitment, we expressed an Lck-Gag protein, lacking L-domains, with a single ubiquitin appended at its C-terminus (Lck-Gag-Ub, Fig. 5A). Ubiquitin is normally conjugated to proteins by an isopeptide bond between the C-terminal glycine residue of ubiquitin and the ε-amino group of a lysine residue within the substrate protein. Therefore, to avoid aberrant conjugation of our Gag-ubiquitin chimeras to other proteins we deleted two glycine residues from the C-terminus of ubiquitin (Fig. 5A).
Cells expressing ubiquitin-fused, but L-domain-deficient Gag (Lck-Gag-Ub) generated extracellular particles while those expressing the unfused, L-domain deficient counterpart Lck-Gag(L-) protein did not (Fig. 5B). Directly fused ubiquitin-dependent particle release was strongly inhibited, in a dose dependent manner, by expression of a catalytically inactive version of the ATPase VPS4 (Fig. 5C), indicating that the ESCRT pathway was required for Lck-Gag-Ub particle release. Thus, in the context of Lck-Gag, direct ubiquitin fusion appeared capable of substituting for a PSAP or PPxY containing L-domain. These results are similar to findings made by Joshi et al. who showed that direct fusion of ubiquitin to EIAV Gag can functionally substitute for the ALIX-binding YPDL L-domain encoded therein [48]. Similarly, we also found that ubiquitin-dependent budding was dependent on the ubiquitin hydrophobic patch residues (L8 and I44) and additionally, marginally dependent on residues (Q62 and E64) that have been implicated in ubiquitin-Tsg101 UEV domain interaction (Fig. 5D). However, lysine residues (K48 and K63) that are often important for the conjugation of further ubiquitin molecules could be mutated without affecting fused ubiquitin-dependent particle release (Fig. 5D).
Next we analyzed the effect of combining L-domains and ubiquitin on VLP release. To accomplish this, Lck-Gag proteins containing various combinations of the L-domains and C-terminally fused ubiquitin (Fig. 6A) were expressed. Quantitative analyses revealed that directly fused ubiquitin-dependent (Lck-Gag-Ub) particle release was at least as efficient as that driven by PSAP (Lck-Gag(PSAP)) or PPxY (Lck-Gag-PY) L-domains (Fig. 6B). Moreover, and in contrast to the previous report with EIAV Gag [48], we found that the combined presence of fused ubiquitin and a PSAP L-domain (in Lck-Gag(PSAP)-Ub) resulted in strongly synergistic effects on particle release (Fig. 6B). Specifically, Lck-Gag(PSAP)-Ub generated ∼20-fold and ∼6-fold more particles than Lck-Gag(PSAP) and Lck-Gag-Ub, respectively (Fig. 6B). No such synergy was observed when a PPxY L-domain and ubiquitin were combined in the same Gag protein. In fact, the Lck-Gag-Ub and the Lck-Gag-PY-Ub generated extracellular particles with approximately the same efficiency (Fig. 6B). Less dramatic, but nonetheless synergistic enhancement of particle release was evident when PPxY and PSAP motifs were both present (in the absence of ubiquitin fusion, Fig. 6B). In this case, the presence of the PPxY motif (in Lck-Gag(PSAP)-PY) enhanced particle release approximately ∼5-fold as compared to the situation where the PSAP motif was the only L-domain (in Lck-Gag(PSAP), Fig. 6B). Overall these results are consistent with the notion that ubiquitin behaves essentially like an L-domain, and further suggests that it functions synergistically with a PT/SAP motif, and redundantly with a PPxY motif.
We next attempted to mimic a situation that is somewhat typical of retroviruses, where only a fraction of Gag expressed in cells carries ubiquitin. This was done by co-expressing ubiquitin-fused and unfused Lck-Gag proteins in varying proportions. When this was done in the context of a Lck-Gag proteins lacking a PSAP motif (by co-expressing Lck-Gag(L-) and Lck-Gag-Ub), particle production was most efficient when a large fraction of the total Lck-Gag protein carried ubiquitin, and no stimulation of particle production was detectable when less than 25% of the Gag protein carried fused ubiquitin (Fig. 6C, left panel). When similar experiments were done in the presence of a PSAP late domain, by co-expressing Lck-Gag(PSAP) and Lck-Gag(PSAP)-Ub, stimulation of particle release was observed when smaller fractions of Gag, as little as a few percent, carried ubiquitin (Fig. 6C, right panel). Nonetheless, larger fractions of ubiquitin fused Gag had larger stimulating effects on particle release. Thus, these experiments suggest that the greater the number of ubiquitin molecules that are present at sites of particle assembly, the more efficient is particle release; however, relatively modest amounts of ubiquitin can significantly enhance particle budding in the presence of a PSAP motif.
Several class E vacuolar protein-sorting factors have been reported to possess ubiquitin binding activity (Table 1). Although the affinity of such domains for monoubiquitin is generally quite weak (Kd>100µM), several class E factors form multiprotein complexes with several ubiquitin-binding surfaces, which could provide sufficient avidity for their retention at sites of virion assembly. Under such a scenario, efficient recruitment of ESCRT complexes might require deposition of relatively large numbers of ubiquitin molecules in the vicinity of the assembling particle, a notion that is consistent with the finding that a large fraction of Gag must carry ubiquitin to compensate for the absence of a late domain (Fig. 6C).
To determine which of the mammalian ESCRT complexes and associated proteins might be most important for ubiquitin dependent budding, we performed a directed yeast two-hybrid screen in which ubiquitin binding to a range of human class E VPS factors and associated proteins was surveyed. These included components of ESCRT-0 (Hrs, HBP/STAM), ESCRT-I (Tsg101, VPS28, VPS37A,B,C, Mvb12), ESCRT-II (Eap20, Eap30, Eap45) ESCRT-III (CHMP1A, 1B, 2A, 2B, 3, 4A, 4B, 4C, 5, 6), as well as several ESCRT-associated proteins or proteins that are known to bind to components of the class E VPS pathway (ALIX, LIP5, VPS4, UBPY,CMS, CIN85). Most of these proteins, including known ubiquitin binding factors (Table 1), gave either weak or non-specific signals. Since we were testing ubiquitin binding by each protein individually and outside of its natural context and in the absence of ESCRT complex partners, it was perhaps to be expected that this assay would fail to detect ubiquitin interactions in at least some instances. Nonetheless, HBP/STAM, ALIX, and UBPY binding gave robust signals in WT ubiquitin binding assays, and binding was abolished when the ubiquitin hydrophobic patch was mutated (I44A), (Fig. 7A).
We next determined the effect of siRNA mediated disruption of known ubiquitin-binding complexes, as well the additional ESCRT-associated factors that were positive in our yeast 2-hybrid survey (ALIX and UBPY), on PPxY-dependent and fused ubiquitin-dependent Lck-Gag budding. The core components of the known ubiquitin binding ESCRT complexes (ESCRT-0, ESCRT-I and ESCRT-II) were targeted using pools of four siRNAs directed to Hrs, Tsg101 and Eap45, respectively. The potency of the siRNA pools was estimated by cotransfecting them with plasmids expressing YFP-tagged target proteins, followed by quantitative western blotting. By these criteria the Tsg101, Eap45, ALIX and UBPY siRNAs appeared effective (Fig. 7B). However, knockdown of Hrs was inefficient, so its effect on budding could not be reliably assessed. Because antibodies to Tsg101 and ALIX were available, the level of endogenous proteins could also be monitored in these siRNA experiments. Quantitative western blotting analyses (examples are shown in Fig. 7C) indicated that Tsg101 and ALIX proteins were reduced to 38±9% and 16±4% of endogenous levels, respectively. Notably, control experiments showed that Lck-Gag(PSAP) particle release was specifically inhibited (∼5-fold) by Tsg101 siRNA, but only marginally affected by EAP45, ALIX and UBPY depletion (Fig. 7C,D), while EIAV Gag particle release was specifically inhibited (∼3-fold) by ALIX depletion, but not by depletion of the other ESCRT-associated proteins (Fig. 7C,D).
Ubiquitin-dependent (Lck-Gag-Ub) budding was modestly inhibited (∼3-fold) by depletion of either Tsg101 or ALIX but was barely affected by UBPY or Eap45 siRNAs (Fig. 7C,D), suggesting that ubiquitin binding to ESCRT-I and ALIX contributes to its ability to mediate particle release. This finding mirrors a previous report using ubiquitin fused to EIAV Gag [48]. Additionally, however, we further found that Lck-Gag-PY exhibited a similar pattern sensitivity to class E factor-targeting siRNAs, in that it was modestly sensitive to Tsg101 and ALIX but not Eap45 or UBPY siRNAs (Fig. 7C,D). Similarly, the budding of an MLV Gag protein, that carries the same PPxY L-domain was also modestly sensitive to depletion of Tsg101 and ALIX Fig. 7C,D).
Because ESCRT-I and ALIX perturbation both affected ubiquitin and PPxY-dependent budding, we sought to determine whether their simultaneous depletion would exhibit a stronger inhibitory effect. Unfortunately, cotransfection of the two pools of siRNAs (or each pool together with normalizing control RNA duplexes,) rendered each somewhat less effective, perhaps due to dilution of the active siRNAs (Fig. 8A). Specifically, Tsg101 protein levels were reduced to 42±2% and 50±2% of endogenous levels, while ALIX protein levels were reduced to 30±3% and 27±2% of endogenous levels, when the Tsg101 or ALIX targeted siRNAs were cotransfected together or with normalizing control siRNAs, respectively (Fig. 8A). Thus, under these conditions, siRNAs targeting ALIX did not inhibit Lck-Gag-Ub or Lck-Gag-PY particle release (Fig. 8A,B). Nevertheless, simultaneous (albeit partial) depletion of Tsg101 and ALIX had a significantly stronger inhibitory effect on Lck-Gag-Ub, Lck-Gag-PY and MLV Gag budding (Fig. 8A, B) than did the more effective individual depletion of either Tsg101 or ALIX alone (Fig. 7C, D), suggesting that they both proteins contribute to optimal PPxY and ubiquitin-dependent budding.
The precise role of HECT ubiquitin ligases in promoting PPxY-dependent virion release has, heretofore, been somewhat unclear. Our previous studies suggest that their ubiquitin ligase activity is critical for their ability to stimulate budding [17], [29], but the functionally relevant substrate for ubiquitination has been difficult to define. Additionally, there is some evidence suggesting that HECT ubiquitin ligases may also function as adaptors for bridging factors that recruit ESCRT proteins to assembling virions [17], [32], [51].
We compared the activities of HECT domains from various Nedd4-like family HECT ubiquitin ligases by fusing them to the C2 and WW domains of WWP1. While this strategy does not illuminate which ubiquitin ligases are responsible for viral budding in the natural context, it does allow an assessment of HECT domain function in a uniform background. We found that HECT domains varied significantly in their ability to stimulate PPxY-dependent particle release in this context. This variability was evident when there were no ubiquitin acceptors in the Gag protein and correlated better with the ability of the HECT domains to drive autoubiquitination than with their ability to ubiquitinate a modified Gag substrate that contained lysines proximal to a PPxY motif. The correlation between autoubiquitination and budding was imperfect, however, and it is possible that variation among the HECT domains in their ability to catalyze different lengths and types of ubiquitin chains (e.g. K48 versus K63-linked chains), or their ability to ubiquitinate other bystander proteins, could influence their ability to stimulate viral budding. In this regard it was notable that there was no correlation between the ability of the HECT domains to undergo autoubiquitination versus their ability to catalyze ubiquitin ligation to Lck-Gag-PY-3K. It was nonetheless true that the ability of the HECT domains to stimulate budding was, in every case, absolutely dependent on their ability to catalyze the ligation of ubiquitin to a substrate. This suggests that the proposed role of HECT domains as adaptors that bind directly to downstream factors is of secondary importance in stimulating budding, or that this adaptor function requires catalytic activity. This latter scenario could, conceivably, be operative as a result of HECT autoubiquitination.
These studies underscore the remarkable flexibility in the ways that the ESCRT pathway can be engaged to achieve viral budding (Fig. 9) Using a single viral Gag protein as a model, particle budding could be achieved by: (i) conventional direct recruitment of the ESCRT pathway via PTAP binding to Tsg101, (ii) direct recruitment of the ESCRT pathway via PPxY binding to a hybrid cofactor consisting of the C2/WW domains of WWP-1 linked to the C-terminal domain of Tsg101, (iii) recruitment of a HECT ubiquitin ligase via a PPxY motif, (iv) recruitment of an isolated HECT domain to a PTAP motif using a hybrid L-domain cofactor consisting of the UEV domain of Tsg101 linked to a HECT domain or (v) direct fusion of ubiquitin to Gag. These results suggest that the cellular factors (in this case Tsg101, ubiquitin ligases and ubiquitin) that are either directly recruited or deposited at the site of viral particle budding behave as modular entities, with domains that are necessary and sufficient for their own recruitment, and distinct domains that are necessary and sufficient for the subsequent recruitment of downstream effectors of particle release (Fig. 9).
When HECT domains were used to promote budding, the requirement for catalytic activity was absolute, irrespective of how they were recruited to Gag and, importantly, in the absence of ubiquitin acceptors on the viral Gag protein. This finding suggests that ligation of ubiquitin to trans-acting factors, perhaps including the HECT domain itself (i.e. autoubiquitination), rather than Gag is important for viral budding. It is superficially paradoxical, therefore, that ubiquitin could promote budding of the very same Gag protein even when ubiquitin was not ligated to a trans-acting factor, but rather was directly fused to Gag. These findings suggest that the identity of the protein(s) to which ubiquitin is attached is not of critical importance, and ubiquitination substrates can, in principle, include Gag, the ubiquitin ligase itself, or other trans-acting proteins. The mere presence of ubiquitin at the site of particle assembly appears sufficient to engage the ESCRT pathway and stimulate budding.
The intrinsic manipulability of L-domains, the proteins that bind to them (specifically ESCRT-I and HECT ubiquitin ligases) and the apparent lack of importance of the identity of ubiquitination substrate suggests that each serve simply as recruitment factors to engage the downstream machinery that mediates membrane fission and particle release. Since ubiquitin binds to the very same factors (ESCRT-I and ALIX) that are bound by PT/SAP and YPXL type L-domains, and depends on them to stimulate budding, then ubiquitin itself can be conceptually viewed, in the context of viral budding, as a transferable L-domain that acts in a position-independent manner. In essence, this notion is a simple extension of the concept originally demonstrated by Parent et al, who showed that conventional L-domains function in a position independent, transferable manner [56].
A finding that is consistent with the aforementioned arguments, is that budding that was dependent either on a PPxY motif or a ubiquitin fused directly to Gag exhibited similar dependence on particular components of the ESCRT pathway. Notably, perturbation of individual segments of the pathway (ESCRT-I and ALIX) caused partial inhibition of ubiquitin-dependent Lck-Gag-PY, Lck-Gag-Ub and MLV Gag particle release. Previous work has shown that Mason-Pfizer monkey virus particle release, which is dependent on a PPxY motif, is quite strongly inhibited by depletion of Tsg101 [31] and that budding of a EIAV Gag-ubiquitin fusion protein is modestly inhibited by Tsg101 or ALIX depletion [48]. We found that simultaneous perturbation of ESCRT-I and ALIX resulted a stronger suppression of Lck-Gag-PY, Lck-Gag-Ub and MLV Gag particle release than did depletion of either protein alone, suggesting that both ESCRT-I and ALIX can contribute to optimal PPxY- and ubiquitin-dependent budding (Fig. 9). Indeed, the class E VPS pathway includes multiple ubiquitin-interacting factors, each of which could, in principle, provide parallel mechanisms for engaging the ESCRT machinery. While ESCRT-I and ALIX appeared to be most important for PPxY- and ubiquitin-dependent budding, these experiments do not exclude a contributory role for other ubiquitin binding complexes in the ESCRT pathway. A similar notion was recently demonstrated in yeast, where simultaneous disruption of ubiquitin binding by ESCRT-I, -II and Bro1 (the yeast homologue of ALIX) was necessary to block the sorting of ubiquitinated cargo to the lysosome [57]. Thus, ubiquitin has several potential entry points into the ESCRT pathway, and it appears that multiple interactions must be simultaneously inhibited in order to profoundly inhibit ubiquitin- or HECT ligase-dependent budding.
Since ubiquitin-binding class E VPS factors generally have a low affinity for individual ubiquitin molecules (Table 1), the efficiency with which they are recruited to, and retained at, sites of particle assembly is likely related to the number of ubiquitin molecules that are locally present. Indeed, in the context of direct ubiquitin fusion to Lck-Gag, particle release efficiency increased as the proportion of Gag molecules that carried a ubiquitin was increased, and directly fused ubiquitin could effectively bypass the need for a conventional L-domain only when a large fraction (>50%) of the Gag molecules were fused to ubiquitin. This approximates to ∼1000 to 2500 ubiquitin molecules per assembling virion.
Previous studies have shown that direct ubiquitin fusion to RSV or EIAV Gag can alleviate a late budding defect imposed by proteasome inhibitors or functionally replace a YPDL L-domain [44], [48]. However, this study is the first to demonstrate that ubiquitin can act synergistically with a PTAP motif, resulting in dramatically enhanced particle release when both are present. Moreover, the ability of fused ubiquitin to stimulate budding became evident at significantly lower Gag-ubiquitin abundance (5% to 25% of total Gag) when a PTAP motif was also present in Gag. Since ubiquitin could serve as an additional docking site for Tsg101, it might synergize with PTAP motifs by increasing the overall affinity of the assembling Gag lattice for individual ESCRT-I complexes. In fact, this property was predicted by previous binding studies involving Tsg101 UEV domain, PTAP containing peptides and ubiquitin [7]. Ubiquitin might also synergize with PTAP motifs by providing binding sites for distinct class E VPS factors (e.g. ALIX), thereby optimally utilizing all the available components of the ESCRT machinery. Consistent with these ideas, PTAP and PPxY L-domains behaved synergistically in driving particle release, as did PTAP and Gag-fused ubiquitin. However, a PPxY motif and Gag-fused ubiquitin behaved redundantly, consistent with the notion that that they ultimately function through the same mechanism.
pCAGGS-based expression plasmids encoding Lck-Gag(PSAP), Lck-Gag(L-), Lck-Gag-PY, and Lck-Gag-PY-3K plasmids have been described previously [29]. The Lck-Gag(ASAP) plasmid was derived from Lck-Gag(PSAP) by PCR-based site-directed mutagenesis. The Lck-Gag(PSAP)-PY plasmid was generated by replacement of a StuI/XhoI fragment from the Lck-Gag(PSAP) plasmid with the corresponding fragment from the Lck-Gag-PY plasmid. cDNAs expressing Lck-Gag-Ub (ubiquitinΔGG) fusion proteins were generated by overlap-extension PCR, using pCAGGs-Lck-Gag(PSAP), Lck-Gag(L-), and Lck-Gag(PSAP)-PY as templates for the N-terminal portions and pHA-ubiquitin as the template for the C-terminal portion. The K48R, K63R, F4A, L8A, I44A, and QE62,64AA point mutations were introduced into the Lck-Gag-Ub construct by PCR-based mutagenesis. Each cDNA was cloned into pCAGGs for expression in mammalian cells.
DNAs encoding the HECT domains from WWP1 (residues 543–922), WWP2 (491–870), Nedd4 (520–902), Nedd4L (593–975), Itch (483–862), Smurf1 (374–757), and Rsp5 (431–809) were amplified from plasmids encoding the full-length HECT ligases [17], [58]. The Bul2 HECT domain (encoding residues 1189–1572) was PCR amplified from 293T cell cDNA. The catalytically inactive WWP1 HECT domain (C890S) was amplified from a previously described full-length mutant WWP1 ligase [17]. Catalytic point mutants of the remaining HECT domains were made by PCR-based mutagenesis. Chimeric ubiquitin ligases, comprising the C2 and WW domains (residues 1–542) of WWP1 and each of the HECT domains described above were generated by overlap PCR. Likewise, plasmids expressing Tsg-WWP1, Tsg-Nedd4L and Tsg-Itch (residues 1–157 of Tsg101 fused to HECT domains of WWP1, Nedd4L, or Itch) as well as WWP1-Tsg-C (residues 1–542 of WWP1 fused to residues 157–390 of Tsg101) were constructed by overlap-extension PCR. All cDNAs encoding chimeric proteins were inserted into pCR3.1/YFP, to express proteins fused to the C-terminus of YFP, for in mammalian cells. The class E VPS factor yeast two-hybrid library and plasmids expressing Vps4 E228Q, Tsg101, Hrs, ALIX, UBPY, and Eap45 fluorescent fusion proteins in mammalian cells have been described previously [8], [12], [54]. Yeast two-hybrid plasmids encoding wild type and I44A mutant ubiquitin were constructed by PCR amplification of ubiquitinΔGG from the pHA-ubiquitin plasmid using 5′ and 3′ primers appended with EcoRI restriction sites and cloning into the pGBKT7 (Clontech) and pVP16 vectors [8].
For Gag particle release assays, 5×105 293T cells in six-well plates were transfected using polyethylenimine (Polysciences) with 1 µg of pCAGGs/Gag-derived plasmids, alone or with 1 µg of pCR3.1/YFP, pCR3.1/YFP-WWP1/HECT, pCR3.1/YFP-Tsg-HECT, or pCR3.1/YFP-C2-WW-Tsg-C plasmids, or the indicated amounts of pCR3.1/YFP-Vps4 E228Q plasmid. For EIAV and MLV VLP release assays, 293T cells were transfected in the same format with 500ng of, pCR3.1/EIAVGag or pCR3.1/MLVGag-HA plasmids. VLPs were pelleted by ultracentrifugation of 2 ml of 0.22-µm-filtered culture supernatants, collected 48 hours after transfection, over a 2ml 20% sucrose cushion for 90 min at 115,000×g. VLP and cell lysates were analyzed by Western blotting.
293T cells (5×105) in six-well plates were cotransfected with 1 µg of pCAGGs/Lck-Gag-PY-3K, 500 ng of pHA-ubiquitin, and 1 µg of the indicated chimeric pCR3.1-WWP1-HECT ligase. At 36h after transfection, cells were thoroughly lysed at room temperature in detergent-rich RIPA buffer (50mM Tris pH 7.4, 150mM NaCl, 1mM EDTA, 1.0% glycerol, 0.5% SDS, supplemented with protease inhibitor tablets (Roche) and 5mM N-ethylmaleimide to inhibit deubiquitination) and cleared of cellular debris by microcentrifugation. The lysates were then diluted 5-fold in the same buffer containing NP-40 rather than SDS, to adjust the concentration of SDS to 0.1% and NP-40 to 1.0%, and split into two fractions. From one fraction, Gag proteins were immunoprecipitated with αPFV serum and protein G-Sepharose beads. From the other fraction, YFP-HECT ligase proteins were immunoprecipitated with αGFP monoclonal antibody and protein G-Sepharose beads. Immunoprecipitates and unfractionated cell lysates were analyzed by Western blotting.
293T cells (3×105) in six-well plates were transfected with siGENOME siRNAs targeting Luciferase, Tsg101, Hrs, Alix, UBPY, or Eap45 (Dharmacon) using Lipofectamine 2000 (Invitrogen). After 24h, cells were transfected with the same siRNAs and the indicated Gag expression plasmids. VLP and cell lysates were prepared 48 h after the second transfection. To assess knockdown efficiency, 293T cells were transfected once with YFP-Tsg101, -Hrs, -ALIX, -UBPY, or -Eap45 expression plasmids and corresponding siRNAs. GFP expression in cell lysates harvested 48 h after transfection was assayed by quantitative Western blotting.
Virion and cell lysates and immunoprecipitates were separated on polyacrylamide gels, transferred to nitrocellulose membranes, and probed with various antibodies: anti-PFV human serum, anti-HIV-1 p24CA (183-H12-5C), anti-EIAV equine serum (VMRD, Inc.), anti-GFP (Roche), and anti-HA (HA.11, Covance) anti-Tsg101 (4A10, Abcam, Cambridge, MA) or anti-ALIX rabbit serum (a gift from Wes Sundquist). Subsequently, the blots were probed with species-specific peroxidase-conjugated secondary antibodies and chemiluminescent substrate reagents. Alternatively, for quantitative Western blotting, membranes were probed with species-specific antibodies conjugated to IRDye800CW, and fluorescent signals were detected and quantified using a LICOR Odyssey scanner.
Yeast cells (Y190) were transformed with the pGBKT7- and pVP16-derived plasmids described above. Transformants were selected and protein-protein interactions were assayed by β-galactosidase reporter activity as previously described [8].
|
10.1371/journal.pntd.0001294 | Increased CD4+/CD8+ Double-Positive T Cells in Chronic Chagasic Patients | CD4+/CD8+ double positive (DP) T cells have been described in healthy individuals as well as in patients with autoimmune and chronic infectious diseases. In chronic viral infections, this cell subset has effector memory phenotype and displays antigen specificity. No previous studies of double positive T cells in parasite infections have been carried out.
Seventeen chronic chagasic patients (7 asymptomatic and 10 symptomatic) and 24 non-infected donors, including 12 healthy and 12 with non-chagasic cardiomyopathy donors were analyzed. Peripheral blood was stained for CD3, CD4, CD8, HLA-DR and CD38, and lymphocytes for intracellular perforin. Antigen specificity was assessed using HLA*A2 tetramers loaded with T. cruzi K1 or influenza virus epitopes. Surface expression of CD107 and intracellular IFN-γ production were determined in K1-specific DP T cells from 11 chagasic donors. Heart tissue from a chronic chagasic patient was stained for both CD8 and CD4 by immunochemistry. Chagasic patients showed higher frequencies of DP T cells (2.1%±0.9) compared with healthy (1.1%±0.5) and non-chagasic cardiomyopathy (1.2%±0.4) donors. DP T cells from Chagasic patients also expressed more HLA-DR, CD38 and perforin and had higher frequencies of T. cruzi K1-specific cells. IFN-γ production in K1-specific cells was higher in asymptomatic patients after polyclonal stimulation, while these cells tended to degranulate more in symptomatic donors. Immunochemistry revealed that double positive T cells infiltrate the cardiac tissue of a chagasic donor.
Chagasic patients have higher percentages of circulating double positive T cells expressing activation markers, potential effector molecules and greater class I antigenic specificity against T. cruzi. Although K1 tetramer positive DP T cell produced little IFN-γ, they displayed degranulation activity that was increased in symptomatic patients. Moreover, K1-specific DP T cells can migrate to the heart tissue.
| Chagas disease, produced by the blood parasite Trypanosoma cruzi, is considered a public health problem in Central and South America. Non sterile immunity can be achieved after acute infection. Parasite persistence can induce tissue damage in nearly 20% to 30% of chronically infected individuals. Indeed, chagasic cardiomyopathy is one of the consequences of the chronic infection. Antigen persistence and dysfunctional cellular immune response have been implicated in T. cruzi pathogenesis. Here, a higher frequency of circulating CD4+/CD8+ double positive T cells in chronic chagasic patients is reported as compared with non infected donors, including those with a non-chagasic cardiomyopathy. This cell subset also expressed more activation markers and stored more intracellular perforin. We have previously reported that CD8+ T cells from T. cruzi infected donors recognized the HLA-A*0201 restricted K1-peptide derived from the KMP-11 protein. Here, double positive T cells displayed higher percentages of recognition for the K1 peptide than single CD8+ T cells. These cells produce little IFN-γ, but display degranulation activity that was increased in the symptomatic group. Finally, double positive T cells can be localized in the heart tissue from a chronic chagasic donor.
| Expression of either CD4 or CD8 on mature peripheral CD3+ T cells is considered to be a mutually exclusive event as a result of the thymic selection, reflecting the specific functions of each major T cell subpopulation. Contrary to this conventional dichotomy, circulating CD4+/CD8+ double positive (DP) T cells have been identified in human peripheral blood and represents between 1 and 3% of the total T lymphocytes population [1]. The role of these DP T cells in health and disease is still under investigation.
Some healthy individuals can display a significant proportion of DP T cells in peripheral blood. Furthermore, previous evidence has suggested that their frequency in blood can increase or they can be localized in specific tissues during several inflammatory diseases [2]–[8], including: a) chronic viral diseases, like EBV infectious mononucleosis [3] and HIV [4]; b) autoimmune pathologies characterized by chronic lymphocytes activation, such as autoimmune thyroiditis [5], myasthenia gravis [6] and systemic sclerosis/scleroderma [7]; c) allergy i.e. atopic dermatitis [8] and d) some neoplasias [3], [9].
Based on the intensity of CD4 and CD8 expression by flow cytometry, two subsets of DP T cells have been defined: CD4dim/CD8bright and CD4bright/CD8dim lymphocytes [3]. Previous studies of their phenotypic characteristics have shown that the majority of these cells have memory phenotype (CD45RO+). They are also more differentiated than mono-positive T cells, based on its low level of expression of CD27, and they frequently produced either intracellular granzyme B or perforin [10], [11]. Functional assays showed that during chronic viral infections, DP T cells secrete cytokines, such as IFN-γ, in response to cognate class I restricted antigens [10]. All these results suggest that, although small, DP T cells constitute a highly differentiated memory subpopulation acting in the adaptive immune response against infectious agents [1]–[4], [10], [11].
As this cell population is expanded in several viral and immunological mediated chronic diseases, it seems plausible that DP T cells can contribute to the immune response against chronic parasitic infections. The goal of this study was to determine the frequency, phenotype and effector potential of circulating DP T cells in patients chronically infected with T. cruzi, the etiological agent of Chagas disease. This hemoflagellate parasite, which is found in Central and South America, is transmitted by vectors of the Reduviidae family and produces both an acute and a chronic phase [12], [13]. Prevalence of Chagas diseases in Colombia is calculated to be 1'300,000, while 3'500,000 people are at risk of contracting the infection, representing a potential public health problem [13], [14]. After T. cruzi inoculation, during the acute phase, the parasite extracellular stages are found in peripheral blood and infected individuals develop constitutive symptoms. Activation of the immune system allows parasite control, although it is not completely eliminated. Parasites invade cells such as monocytes/macrophages, dendritic cells, fibroblast and myocardiocytes, shedding their flagella and becoming intracellular [15]. By mechanisms not clearly understood, parasites persist leading to the chronic phase of the disease. Most of the infected individuals will remain asymptomatic for several years, being classified as indeterminate. Those patients are usually detected by routine serological tests in blood banks. Nearly 20% to 30% of the chronic infected patients will develop tissue damage, being Chagasic cardiomyopathy or digestive disease the most common pathologies [12], [13]. What mediates the tissue damage in Chagas is not well understood, but some evidence suggests that parasite persistence and dysfunctional cellular immune response could contribute to this process [12], [15]. During the infection, antibodies elicited against T. cruzi antigens help to control blood circulating parasite [16], [17] and specific CD4+ and CD8+ T cells act against intracellular forms [18], [19]. Mice models demonstrated that T cells response seemed to be crucial for parasite control [20]. Interestingly, acute infection with T. cruzi in mice showed a large increase of CD4+/CD8+ DP T cells in their subcutaneous lymph nodes [21]. Given the characteristics of the DP T cells (memory phenotype and presence of cytotoxic granules), it is of special interest to determine their functional characteristics in chronic chagasic patients.
Research protocols and informed consents were approved by the Ethical Committees of the Universidad de los Andes (039-2009), Pontificia Universidad Javeriana (01-2010) and the Fundación Abood Shaio (134-2010), Bogotá, Colombia, following the national regulations and the Declaration of Helsinki.
Forty-one volunteers who signed the informed consent were enrolled in this study and divided into three groups. The first group included 17 chronic chagasic patients with positive immunofluorescence indirect assay (IFI) and ELISA tests. It was composed of 13 females and 4 males with ages ranging from 36 to 67 years old, recruited at the Fundación Abood Shaio (Bogotá, Colombia) or at Instituto Nacional de Salud; and classified according to the Kuschnir grading system [22]. Seven individuals were classified as indeterminate chagasic patients with no significant findings during clinical assessment (G0 or asymptomatic) and 10 as cardiac chronic chagasic patients (symptomatic) graded as follows: 3 with only abnormal ECG results (G1), 5 with abnormal ECG results and cardiac enlargement (G2) and 2 with abnormal ECG results, cardiac enlargement and clinical signs of heart failure (G3). The second group included 12 donors with cardiomyopathy of non-infectious etiology, including 8 females and 4 males with ages ranging from 28 to 80 years old, and whose anti-T. cruzi antibodies were negative. In this group, cardiomyopathy etiology was: ischemic cardiomyopathy (n = 6), systemic arterial hypertension (n = 4) and idiopathic dilated cardiomyopathy (n = 2). This group of patients was recruited at the Department of Cardiology, Hospital Universitario San Ignacio, Bogotá, Colombia. The third group included 12 uninfected donors who never lived in a Chagas endemic area, and with negative IFI and ELISA tests. They were 9 females and 3 males with ages ranging from 34 to 68 years old. The characteristics of the three groups, denoted as chagasic patients (CP), non-chagasic cardiomyopathy (NCC) and healthy controls (HC), including main clinical, electrocardiogram and echocardiography findings, are resumed in Table 1.
Blood samples were obtained from each donor using EDTA vacuntainer tubes. Total blood (100 µl/tube) was stained with anti-CD3 APC (clone SK7), anti-CD4 PerCP (SK3), anti-CD8 PE (RPA-T8), anti-HLA-DR PE-Cy7 (L243) and anti-CD38 FITC (HIT2). All monoclonal antibodies were purchased from BD Bioscience (San Jose, CA, USA). Blood was stained in darkness for 30 minutes at 4°C and incubated with cell lysis buffer (BD Bioscience) for 10 minutes at room temperature. Then, cells were washed twice in PBS 0.01 M pH 7.4 and gently re-suspended. Samples were acquired and analyzed in a FACS Canto II with FACSDiva software (BD Bioscience). At least 5×104 cells were acquired in the lymphocyte population gate according to their forward scatter (FSC) versus side scatter (SCC) features. Dead cells were excluded by light scatter (FSC-H versus FSH-A). All data shown for CD4+/CD8+ DP T cells are based on the CD3+ gate.
Analysis was done with 2×106 of peripheral blood mononuclear cells (PBMC) from each individual, obtained by ficoll-hypaque density gradient (Sigma, St. Louis, MO, USA). Viability was assessed by trypan blue exclusion (Sigma). Surface staining was done with antibodies against CD3 APC, CD4 PerCP and CD8 PE-Cy7 for 30 minutes at 4°C. After one wash, the cell membrane was permeated with Cytofix/Cytoperm solution for 20 minutes at 4°C, followed by washing with Perm/Wash 1× (BD Bioscience). Intracellular staining was done with anti-perforin FITC antibody (clone δG9) or mouse IgG2b isotype control (clone 27–35) for 30 min at 4°C. Lastly, cells were washed, re-suspended and analyzed as described above. At least 2×105 CD3+ cells were recorded by flow cytometry and perforin expression was determined in CD4+/CD8+ gate.
As HLA-A2 is one of the most common class I allele in Colombia [23], donors were typed for HLA-A2 and subtype for HLA-A*0201 by flow cytometry and SS-PCR as previously described [24], [25]. Tetramer analysis was only done in HLA-A*0201+ individuals, including: 13 chagasic donors (6 asymptomatic and 7 symptomatic) and 5 healthy controls. Cells were incubated with anti-CD3 APC, anti-CD4 PerCP, anti-CD8 PE-Cy7 and anti-HLA-A*0201 PE tetramers loaded either with T. cruzi or influenza virus derived epitopes. T. cruzi epitope or K1, is a 9 mer peptide (TLEEFSAKL) derived from N-terminal region (amino acids 4–11) of T. cruzi KMP-11 protein (Swiss-Prot accession number: Q9U6Z1) [24], [25]. Influenza virus peptide was a modified epitope derived from the viral matrix proteins (a.a. 58–66) denominated MP-Flu (GILGFVTTL) [25]. Tetramers were synthesized by the National Institute of Health (NIH) Tetramer Facility (Atlanta, USA). At least 2×105 cells were analyzed by flow cytometry in the CD3+ gate and tetramer expression was determined in the CD4+/CD8+ gates.
PBMC from 5 asymptomatic (G0) and 6 symptomatic patients (two of each: G1, G2 and G3) were isolated using density gradient as described above. A total of 2×106 PBMC were culture in the presence of anti-CD28 (1 µg/ml) and anti-CD49d (1 µg/ml) and one of the following conditions: medium alone, 10 µg/ml of K1 peptide or 3.7 µg/ml of Staphylococcal enterotoxin B (SEB). Cultures were incubated for 3 hours at 37°C with 5% of CO2, followed by an additional 6 hours of incubation in the presence of Brefeldin A 10 µg/ml (BD Pharmingen). For degranulation assays, anti-CD107a and CD107b FITC (clones H4A3 and H4B4, respectively) were added immediately following stimulation. For surface staining, antibodies for CD3-Pacific Blue (UCHT1), CD4-PeCy5.5 (SK3), CD8-APCH7 (SK1) and K1 PE tetramer were used. PBMC were washed once and then fixed/permeated using Cytofix/Cytoperm (BD Bioscience). Cells were washed twice with Perm/Wash 1× (BD Bioscience), stained with 10 µl anti-IFN-γ PE-Cy7 (clone B27), rewashed once and re-suspended in FACs Flow (BD Bioscience). Population was gated according viability and CD3 expression. Expression of CD107 and production of IFN-γ was based on the K1 tetramer positive CD4+/CD8+ cells. At least 1×103 tetramer positive cells were acquired and analyzed in a FACS Aria with FACSDiva software (BD Bioscience).
Cardiac tissue from a chronic chagasic patient biopsy was used for immunochemistry. This specimen was obtained from a 47 year old male who was diagnosed with cardiac failure class III (NYHA classification) and underwent a cardiac transplant. Formalin-fixed and paraffin-embedded cardiac biopsy tissue was cut, deparaffinized in xylene and rehydrated with alcohol. Antigen retrieval was done heating the tissue in the presence of EDTA buffer pH 9.0. Slides were stained with hematoxylin/eosin to evaluate cellular infiltration and presence of parasites. Immunohistology was carried out with anti-human CD4 (clone 1F6, Vector Laboratories, Burlingame, CA, USA) and CD8 (C8/144B, DakoCytomation, Carpinteria, CA, USA) antibodies. After blocking endogenous enzymes, antibodies were revealed with EnVision™ G|2 Doublestain System kit (DakoCytomation) using peroxidase/DAB+ Chromogen for CD8 and then alkaline phosphatase/Permanent Red Chromogen for CD4. Slides were evaluated in a light microscope (40–100× objective magnification) where CD8+ cells yielded a brown-color and CD4+ a red-color end product; DP T cells showed a red-brown color.
Descriptive statistics (mean, standard deviation and percentages) was used to describe the populations and to present the flow cytometry data. Non-parametric analysis was carried out (Statistix 8.0 software) for groups comparisons by Kruskal Wallis (K-W) test followed by Dunn post hoc tests. Comparison of two groups was done by Mann-Whitney U test (M-W). Significance was considered a p value<0.05.
Cell frequency between chronic chagasic patients (CP) and healthy controls (HC) matched by age and gender were compared. The average percentage of total circulating DP T cells was higher (p = 0.0017) in CP (X = 2.1%±0.9) compared with HC group (1.1±0.5) (Figure 1B). Representative flow cytometry dot plots and gating strategy are shown in Figure 1A. To determine if this difference was attributed to the chagasic patients immune response and not to the cardiac damage or its pathological consequences, DP T cells percentages from CP were compared with those from donors with non-chagasic cardiomyopathy (NCC) (Figure 1A and 1B). Likewise, DP T cells frequency from CP was higher (p = 0.034) compared with NCC group (X = 1.2±0.4). For the following analysis, DP T cells were divided according to their phenotype into CD4high/CD8low and CD4low/CD8high. The frequency of CD4high/CD8low T cell was higher (p = 0.0058) in CP (X = 1.61%±0.87) than in HC (0.88±0.56) and NCC (0.87±0.41) donors (Figure 1C). Regarding to CD4low/CD8high T cell, significant differences were only observed between CP and HC (p = 0.015) frequencies (Figure 1A and 1C).
To define the activation status of DP T cells in chronic Chagas patients, CD38 and HLA-DR surface markers were analyzed. Since, it was previously reported that CD4high/CD8low and CD4low/CD8high T cells differed in the expression of several surface markers, phenotypic analysis was done separately [1]–[4], [10]. The percentage of DP T cells expressing HLA-DR was three times higher in CP in both CD4high/CD8low (p = 0.0031) and CD4low/CD8high (p = 0.01) populations than in the two control groups. No differences were observed between HC and NCC control groups (Figure 2B). Representative flow cytometry density plots and percentages of HLA-DR positive DP T cells are shown in Figure 2A and Figure 2B, respectively. Analysis of CD38 in DP T cells showed that CD4high/CD8low subset had increased (p = 0.002) percentage of cell expressing this marker in CP (X = 38.9%±22.0) compared with NCC (9.6±5.3), and not different with HC (20.6±10.6). In contrast, the frequency of CD4low/CD8high expressing CD38 was higher in CP (X = 30.0%±10.5) (p = 0.0001) when compared with both HC (12.6±7.7) and NCC (11.0±10.5). When co-expression of both activation markers (CD38 and HLA-DR) was analyzed, CP doubled the percentage of activated cells compared to HC and NCC donors in both CD4high/CD8low (p = 0.009) and CD4low/CD8high T cells (p = 0.001) (Figure 2A and 2C). There were no differences when the analysis by clinical status was done. Due to the similarity of both subsets of DP T cells (Table S1), data is presented on the whole population in subsequent analyses.
To assess the cytotoxic potential of DP T cells, intracellular perforin was measured [10]–[11]. Increased (p = 0.002) percentages of cells containing perforin in CP (9.9±4.8) were found when compared with HC (3.3±3.2) and NCC (3.3±2.9) donors. Representative flow cytometry dot plots and percentage of perforin positive cells are shown in Figures 3A and 3B, respectively. No difference in perforin expression was found (p = 0.29) between asymptomatic (9.9±3.1) and symptomatic (9.9±5.7) chagasic patients.
Next, the antigen specificity of DP T cells was determined. To do so, the frequency of HLA-A*0201/peptide recognition was assessed using epitopes from T. cruzi and influenza virus, denominated K1 (Figure 4A) and MP-Flu (Figure 4B), respectively [25]. We found that both asymptomatic (X = 3.0%±1.9) and symptomatic (4.9±3.6) CP had higher percentages (p = 0.0006) of circulating DP T cells specific for the K1 cytotoxic epitope than HC donors (0.6±0.4). There was no difference according to the disease stage (p = 0.267) (Figure 4C and Table S2). The percentage of K1-specific DP T cells in the Chagasic donors was in average 27 times higher (p = 0.0005) than the mono-CD8+ K1-specific ones (X = 4.0±2.9 versus X = 0.17±0.08, respectively), Table S2. Representative flow cytometry dot plots and percentage of K1 specific DP T cells are shown in Figure 4A and Figure 4C, respectively. Regarding MP-Flu response, asymptomatic (3.9±1.5), symptomatic (2.9±2.0) and HC (3.6±2.0) donors had similar percentages of specific DP T cells (p = 0.39) for this viral epitope. Representative flow cytometry dot plots and percentage of MP-Flu specific DP T cells are shown in Figures 4B and 4D, respectively.
Production of IFN-γ and surface expression CD107 a/b, as marker of degranulation, were analyzed in K1 tetramer-positive DP T cells. The percentage of IFN-γ positive cells in medium alone was similar between asymptomatic (n = 5) (0.48%±0.45) and symptomatic patients (n = 6) (0.97%±0.97) (p = 0.52). Likewise, in the presence of K1 peptide, asymptomatic patients (1.87%±1.83) showed a no significant increase in the percentage of IFN-γ positive cells when compared with the symptomatic group (0.56%±0.43) (p = 0.26). However, after polyclonal stimulation (SEB), this percentage was significantly higher in the asymptomatic donors (14.6%±10.3 versus 5.55%±4.98; p = 0.049). Representative flow cytometry dot plots and percentage of IFN-γ positive cells are shown in Figures 5A and 5B, respectively. Furthermore, degranulation in K1 tetramer-positive DP T cells was significantly higher in than symptomatic donors in the presence of K1 peptide (21.50%±13.38 versus 66.25%±30.78; p = 0.0455) and SEB (24.74%±21.74 versus 74.20%±31.59; p = 0.03); this seemed to be also the trend for cell cultured in medium alone (16.58%±17.52 versus 67.0%±32.93; p = 0.052), in spite that no significant difference between the asymptomatic and symptomatic groups was found. Representative flow cytometry dot plots and percentage of CD107a/b positive cells are shown in Figures 5A and 5C, respectively.
Cardiac tissue from a chronic chagasic patient was analyzed in order to identify the presence of CD4+/CD8+ cells. The immunohistology analysis showed that myocardiocytes presented reparative nuclear changes such as bigger size, hyperchromatic and visible nucleoli. Moderate patchy myocardial infiltration mostly conformed by lymphocytes, some plasmocytes, macrophages and scarce eosinophils was also observed. The number of infiltrating lymphocytes was 10 or 12 per high power field, and they were mostly CD8+ cells (near 95%, Figure 6A and 6B) and some CD4+ cells (Figure 6B). Double CD4+/CD8+ cells were found in the inflammatory infiltrate of the cardiac tissue in a frequency lower than 2%, as shown in Figure 6C.
Circulating CD4+/CD8+ double positive (DP) T cells represent between 1–3% of the total T lymphocytes population [1]. However, previous evidence has suggested that this frequency can increase during several inflammatory diseases [2]–[8]. In this study we demonstrated that the percentage of peripheral DP T cells was higher in chronic chagasic patients compared with control donors, including individuals with non-chagasic cardiomyopathy. This finding differs from previous studies in HVC and HIV infected patients in which virus infected and control donors did not have differences in DP T cells frequencies on peripheral blood [10], [26]. To the best of our knowledge, this research is the first report of an augmented percentage of DP T cell in human patients with a chronic parasitic infection.
Previous characterization of DP T cells indicates that this lymphocyte subpopulation is constituted mainly by terminally differentiated memory cells. A recent study showed that chronic chagasic patients had higher frequencies of CD4+ effectors memory T cells (TEM) and CD8+ central memory T cells (TCM) when compared with uninfected individuals [27]. Interestingly, in healthy donors CD4high/CD8low T cells were described as being TEM phenotype, meanwhile CD4low/CD8high were mainly TCM [10]. It will be interesting to determine if some of those CD4+ TEM and CD8+ TCM described in chagasic patients could include some DP T cells [27].
Similarly to DP T cells described in other human chronic infectious diseases with antigen persistence [10], [26], [28], [29], we found that the expression of activation markers in these cells were increased in the chagasic donors. Consistently, some studies in T. cruzi infected donors have shown that their T cells (CD3+, CD3+/CD4+ and CD3+/CD8+) tended to be more activated than cells from uninfected donors [30]–[32].
Regarding the source of the peripheral DP T cells, experimental data supports that they might either escape from the thymus [33] or represents over-stimulated mono-positive CD4+ or CD8+ T cells [34], [35]. Interestingly, some studies have shown severe CD4+CD8+ thymocytes depletion coexisting with 16-fold increase of these cells in the periphery (subcutaneous lymph node) after T. cruzi acute infection in mice [21], [36]. These findings support the “thymus escape” theory. However, expression of the “second” marker in mono-positive T lymphocytes as a consequence of antigenic over-stimulation seems also plausible in our study, given the consistent activation exhibited by the DP T cells [34]. More studies are needed to determine the origin of DP T cells in chagasic patients.
Perforin-mediated cytotoxicity has proved to be very important in the protection and pathogenesis of some parasitic infection, including cerebral malaria [37] and toxoplasmosis [38]. In animal models, there is evidence that implicates perforin expression with T. cruzi elimination during the acute disease [32], [37]. However, perforin has also been involved in tissue damage during chronic Chagas disease [39]–[41]. After T. cruzi infection, perforin- knockout mice had an increased parasite burden and increased number of IFN-γ producing T cells infiltrating their hearts. However, these perforin-deficient animals showed more preserved cardiac tissue and less electric conduction abnormalities than normal littermates [39]. Even more, it has been suggested that human cardiac damage is directly related to an increase in the ratio of perforin-positive/total inflammatory cells in heart tissue [40].
Another remarkable result in this study is that DP T cells can recognized a T. cruzi derived class I epitope during chronic infection. The percentage of DP T cells specific for the K1 peptide was exceptionally high (1.7% to 11.5%) compared with previous reports of K1 specificity for mono-CD8+ T lymphocytes (0.09% to 0.34%) in a similar chronic chagasic population [25]. Likewise, in human viral infections (HCV and HIV) upon antigen exposure, DP T cells displayed a much higher frequency of cell-single cytokine production than CD8+ and CD4+ T cells [10], [26]. Analysis of class I antigen recognition suggests that DP T cells in chagasic patients were parasite driven. Probably, as these peripheral cells are T. cruzi antigen-specific and have memory phenotype, they could migrate to the tissue where parasites persist and contribute to T. cruzi induced pathology. In fact, in situ cytotoxic lymphocytes (perforin or granzyme A positive cells) have been described in human heart [41] and gastrointestinal tract [42] of patients with T. cruzi induced tissue damage.
When activation markers, perforin and K1-recognition on DP T cells were compared in chagasic patients according to their clinical status (asymptomatic versus symptomatic patients), no differences were found. Nevertheless, it was notable that some symptomatic donors had higher percentages of K1 T. cruzi specific cells than asymptomatic ones, while percentages of recognitions for influenza virus epitope were similar. We also found that K1 specific DP T cells from symptomatic group displayed increased degranulation activity even with medium alone. However, very low percentages of these K1 tetramer specific DP T cells produced IFN-γ after K1 peptide stimulation, as similarly described for CD8+ T cells in chronic chagasic patients [25]. This was not the case for the production of IFN-γ after polyclonal stimulation, which was significantly augmented especially in the asymptomatic patients. Interestingly, K1 specific DP T cells that produce IFN-γ did not display degranulating phenotype, indicating that DP T population is functionally heterogeneous and complex. In summary, K1 specific cells from asymptomatic donors had higher capacity of IFN-γ secretion than cells from symptomatic donors which have greater cytotoxic potential.
In our study, we found a higher expression of perforin on peripheral DP T cells in the chagasic patients accompanied with an increased degranulation activity in the symptomatic ones. Also, it was demonstrated that DP T cells can migrated to the cardiac tissue. If the blood phenotype of these cells is maintained by the infiltrating ones, it might be possible to associate DP T cells with cardiac damage. A similar mechanism was suggested for HCV infected humans where DP T cells were found to infiltrate the liver [10].
Progressive loss of cytokines secretion is a characteristic of CD8+ T cells exhaustion, a phenomenon related to antigen persistence during chronic infections [43]. Indeed, IFN-γ that is associated with protection against T. cruzi infection [44], is one of the last effector activities to be extinguished in this process [43]. So, our data suggest that K1 specific DP T cells, mainly from symptomatic donors, should be on the pathway of exhaustion, while they keep their cytotoxic potential. Lastly, as TNFα production has been associated with cardiac damage in Chagas disease [45], it will be of interest to test the production of TNFα on T. cruzi specific DP T cells.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.